Cross-Species Cell Annotation Foundation Models: A New Paradigm for Decoding Evolutionary Biology and Disease

Abstract

Cross-species cell annotation foundation models represent a transformative advance in single-cell biology, enabling the deciphering of universal gene regulatory mechanisms across evolution. This article explores how AI models like GeneCompass, TranscriptFormer, and CAME leverage vast datasets from multiple species to accurately identify cell types, predict disease states, and simulate cellular behavior. We examine the foundational principles, methodological architectures, optimization strategies, and validation frameworks that underpin these tools. For researchers and drug development professionals, this synthesis provides critical insights for applying these models to translate findings from model organisms to human biology, accelerating the discovery of disease mechanisms and therapeutic targets.

The Evolutionary Imperative: Why Cross-Species Cell Annotation is Revolutionizing Biology

Defining Cross-Species Cell Annotation Foundation Models

Cross-species cell annotation represents a computational frontier in evolutionary biology and translational research, enabling the transfer of cellular knowledge from model organisms to humans. The advent of single-cell RNA sequencing (scRNA-seq) has generated massive cellular atlases across diverse species, creating an unprecedented opportunity to decipher conserved and divergent cellular programs [1] [2]. Foundation models (FMs), pre-trained on millions of cells through self-supervised learning, have emerged as powerful tools to address the fundamental challenge of cross-species annotation: reconciling genomic differences to identify homologous cell types across evolutionary distances [3] [4] [5]. These models transform single-cell transcriptomics by treating cells as "sentences" and genes as "words," learning deep biological representations that transcend species boundaries through sophisticated architectural innovations [4]. This protocol examines the defining architectures, performance benchmarks, and practical implementation of cross-species cell annotation foundation models, providing researchers with a framework for leveraging these transformative tools in evolutionary biology and drug development.

Quantitative Benchmarking of Model Performance

Comprehensive evaluation of cross-species annotation models reveals distinct performance advantages across different biological contexts and evolutionary distances. The table below synthesizes key quantitative findings from major benchmarking studies.

Table 1: Performance Benchmarking of Cross-Species Cell Annotation Methods

Model	Core Approach	Test Scenarios	Key Performance Metrics	Comparative Advantage
CAME [1]	Heterogeneous graph neural network	54 scRNA-seq datasets across 7 species; 649 species pairs	Significant improvement in cell-type assignment across distant species; 6.26% average accuracy drop when excluding non-one-to-one homologies	Utilizes non-one-to-one homologous gene mapping; robust to sequencing depth inconsistencies
SATURN [2]	Protein language model (ESM-2) with macrogene space	Mammalian cell atlas (335k cells); frog-zebrafish embryogenesis	Effective annotation transfer across evolutionarily remote species; identification of misannotated cell populations	Discovers functionally related gene groups; enables cross-species differential expression analysis
Icebear [6]	Neural network decomposition of cell identity, species, and batch factors	Mouse-chicken-opossum brain/heart sci-RNA-seq3	Accurate cross-species prediction of single-cell profiles; reveals X-chromosome upregulation evolutionary patterns	Enables single-cell resolution comparison without cell type matching; predicts missing biological contexts
Nicheformer [3]	Transformer trained on dissociated and spatial data (110M cells)	Spatial composition prediction; spatial label prediction	Outperforms Geneformer, scGPT, UCE, and CellPLM on spatial tasks	Integrates spatial context; transfers spatial information to dissociated data
scFMs (General) [5]	Various transformer architectures pretrained on large single-cell corpora	5 datasets with diverse biological conditions; 7 cancer types	Robust and versatile across tasks; no single model dominates all scenarios	Captures biological insights into relational structure of genes and cells

Recent benchmarking studies demonstrate that foundation models exhibit particular strengths in capturing biological relationships. A comprehensive evaluation of six scFMs against traditional baselines revealed that while these models are robust and versatile tools for diverse applications, simpler machine learning models can be more efficient for specific datasets under resource constraints [5]. Notably, no single scFM consistently outperforms others across all tasks, emphasizing the importance of task-specific model selection.

Table 2: Performance Across Cell-Level Tasks in Realistic Conditions

Task Category	Top Performing Models	Key Findings	Considerations for Cross-Species Application
Cell Type Annotation [5]	scGPT, Geneformer, UCE	>80-90% accuracy for major cell types; struggles with rare cell types	Model performance correlates with cell-property landscape roughness in latent space
Cross-Species Transfer [1] [2]	CAME, SATURN	Effective even for non-model species and evolutionarily remote pairs	Dependency on quality of homologous gene mapping or protein embeddings
Spatial Context Prediction [3]	Nicheformer	Systematically outperforms models trained only on dissociated data	Requires spatial transcriptomics data for training; enables tissue niche prediction
Clinical Prediction [5]	scGPT, scFoundation	Accurate cancer cell identification and drug sensitivity prediction in zero-shot settings	Potential for translating findings from model organisms to human clinical contexts

Methodological Approaches for Cross-Species Annotation

Homology-Aware Graph Neural Networks (CAME)

The CAME framework employs a heterogeneous graph neural network architecture that explicitly incorporates both one-to-one and non-one-to-one homologous gene mappings, which is particularly crucial for distant species comparisons where up to 60-75% of highly informative genes may not have one-to-one homologs [1].

Experimental Protocol:

• Input Processing:

  • Prepare scRNA-seq count matrices from reference (labeled) and query (unlabeled) species
  • Compile homologous gene mappings, including one-to-many and many-to-many relationships
  • Construct single-cell networks using k-nearest-neighbors (KNN) within each species

• Graph Construction:

  • Create a heterogeneous graph with two node types (cells and genes)
  • Establish cell-gene edges for non-zero expression relationships
  • Create gene-gene edges based on homology mappings
  • Incorporate precomputed cell-cell edges from KNN networks

• Model Architecture:

  • Implement parameter-sharing graph convolution layers for heterogeneous nodes and edges
  • Employ heterogeneous graph attention mechanism for cell-type classification
  • Utilize multi-label classification to handle hierarchical cell types and multiple correspondences

• Training Protocol:

  • Optimize combined multiclass and multilabel cross-entropy loss via backpropagation
  • Train for 200-300 epochs until convergence
  • Use Adjusted Mutual Information (AMI) with preclustered query cells to select checkpoints

• Output Interpretation:

  • Extract cell-type assignment probabilities for query cells
  • Identify unresolved cell states through low-confidence assignments
  • Generate aligned cell and gene embeddings for joint visualization
  • Perform cross-species gene module extraction

Protein Language Model Integration (SATURN)

SATURN introduces a novel approach that couples gene expression with protein embeddings from large language models (e.g., ESM-2) to create universal cell embeddings that transcend genomic differences between species [2].

Experimental Protocol:

• Input Preparation:

  • Obtain scRNA-seq count matrices from multiple species
  • Generate protein embeddings using ESM-2 for all genes
  • Acquire initial within-species cell annotations (cell-type assignments or clustering results)

• Macrogene Space Construction:

  • Learn a shared macrogene space representing functionally related gene groups
  • Define gene-to-macrogene importance weights based on protein embedding similarities
  • Map cross-species datasets to this joint macrogene space

• Model Pretraining:

  • Initialize with autoencoder using Zero-Inflated Negative Binomial (ZINB) loss
  • Regularize reconstruction to preserve protein embedding similarities
  • Use gene-to-macrogene weights to maintain functional relationships

• Weakly Supervised Training:

  • Employ metric learning with two-component objective:
    • Separate different cells within same dataset using weak supervision
    • Align similar cells across datasets in unsupervised manner
  • Calibrate embedding distances to reflect cell label similarity

• Cross-Species Differential Expression:

  • Perform differential expression analysis on macrogenes instead of individual genes
  • Aggregate gene contributions using gene-macrogene neural network weights
  • Interpret biological meaning through highest-weight genes per macrogene

Spatial-Aware Foundation Models (Nicheformer)

Nicheformer represents a paradigm shift by incorporating both dissociated single-cell and spatially resolved transcriptomics data during pretraining, enabling the transfer of spatial context across species [3].

Experimental Protocol:

• Data Curation:

  • Compile SpatialCorpus-110M with 57 million dissociated and 53 million spatial cells
  • Include data from 73 human and mouse tissues across multiple technologies
  • Define orthologous genes across species for shared vocabulary

• Tokenization Strategy:

  • Convert expression vectors to ranked gene token sequences
  • Compute technology-specific nonzero mean vectors to address platform biases
  • Add contextual tokens for species, modality, and technology

• Model Architecture & Training:

  • Implement transformer with 12 encoder layers, 16 attention heads
  • Use 1,500-token context length and 512-dimensional embedding space
  • Train on combined dissociated and spatial data to capture spatial variation

• Spatial Downstream Tasks:

  • Spatial composition prediction: Model local cellular neighborhoods
  • Spatial label prediction: Transfer spatial annotations across species
  • Niche identification: Discover conserved tissue microenvironments

Research Reagent Solutions

Table 3: Essential Research Reagents and Computational Resources

Resource Category	Specific Solutions	Function in Cross-Species Annotation	Key Considerations
Data Repositories	CZ CELLxGENE [4], Human Cell Atlas [4], Tabula Sapiens [7] [2], Tabula Muris [2]	Provide standardized, annotated single-cell datasets for multiple species	Data quality varies; requires careful selection and filtering for pretraining
Protein Language Models	ESM-2 [2]	Generate protein embeddings that capture functional similarity beyond sequence homology	Enable remote homology detection; computationally intensive
Spatial Transcriptomics Technologies	MERFISH, Xenium, CosMx, ISS [3]	Provide spatial context for model training; enable spatial annotation transfer	Targeted gene panels limit gene coverage; technology-specific biases exist
Homology Databases	Orthologous gene mappings [1] [3]	Define evolutionary relationships between genes across species	Non-one-to-one homologies are crucial for distant species comparisons
Benchmarking Datasets	Asian Immune Diversity Atlas (AIDA) v2 [8] [5]	Provide independent validation across diverse populations and cell types	Essential for evaluating model generalization and avoiding data leakage
Computational Infrastructure	GPU clusters, High-performance computing [8] [5]	Enable model pretraining on millions of cells	Significant resources required; barrier to entry for some research groups

Implementation Considerations

Practical Deployment Guidelines

Successful implementation of cross-species annotation models requires careful consideration of several practical factors. Model selection should be guided by specific research questions, as no single foundation model consistently outperforms others across all tasks [5]. For well-established cell types in closely related species, traditional methods may offer computational efficiency, while for novel cell types or distant species comparisons, foundation models with protein language model integration or spatial awareness provide distinct advantages.

Data quality and preprocessing significantly impact model performance. Careful batch effect correction, quality control filtering, and normalization are essential, particularly when integrating datasets across different technologies and species [4]. For cross-species applications, the handling of homologous relationships is critical—methods that incorporate non-one-to-one homologies or protein embeddings generally outperform those restricted to one-to-one gene matches, especially for evolutionarily distant species [1] [2].

Validation and Interpretation Frameworks

Rigorous validation is essential for cross-species annotations. Biological validation should include examination of conserved marker gene expression, assessment of functional enrichment in predicted cell types, and comparison with orthogonal data modalities when available [2]. Computational validation metrics should extend beyond simple accuracy to include ontological similarity measures that capture hierarchical relationships between cell types [5].

Interpretation of model outputs requires special consideration in cross-species contexts. Low-confidence predictions may indicate species-specific cell states rather than annotation failures. Attention mechanisms and feature importance analyses can reveal the gene programs driving cross-species alignments, providing biological insights beyond simple cell type transfer [1] [2].

Future Directions

The field of cross-species cell annotation foundation models is rapidly evolving, with several promising research directions emerging. Multimodal integration that combines transcriptomic, epigenetic, proteomic, and spatial information will likely enhance annotation accuracy and biological relevance [4]. Few-shot and zero-shot learning approaches are being developed to handle rare cell types and poorly characterized species [5]. Additionally, methods that explicitly model evolutionary distances and phylogenetic relationships may improve annotation transfer across broader taxonomic ranges.

As these models mature, development of more sophisticated benchmarking frameworks and standardized evaluation metrics will be crucial for advancing the field. Community efforts to create comprehensive cross-species benchmark datasets and establish best practices for model reporting will enable more systematic comparisons and accelerate progress in this transformative area of computational biology.

Quantitative Foundations of Evolutionary Divergence and Technical Variation

The development of robust cross-species foundation models requires a clear quantitative understanding of the biological and technical variabilities involved. The table below summarizes key metrics and benchmarks essential for designing and evaluating such models.

Table 1: Key Quantitative Benchmarks in Cross-Species Single-Cell Analysis

Metric / Component	Description / Value	Biological Significance / Impact
Evolutionary Distance (Data)	Training on 12 species spanning 1.5 billion years of evolution [9]	Enables model generalization across vast evolutionary scales and OOD prediction.
Data Scale (Model Training)	Models pretrained on >100 million cells from diverse public archives [10]	Provides the foundational "knowledge base" for the model's understanding of cellular biology.
CCD Boundary Conservation (Human vs. Chimpanzee)	71.2% of human CCD boundaries are shared with chimpanzees [11]	Provides a quantitative measure of 3D genome architecture conservation between close species.
Minimum Contrast Ratio (WCAG AA - Large Text)	At least 3:1 [12] [13]	Ensures accessibility and legibility for data visualization interfaces and published findings.
Enhanced Contrast Ratio (WCAG AAA - Body Text)	At least 7:1 [14]	A higher standard for legibility in critical displays and publications.

Experimental Protocols for scFM Development and Benchmarking

Protocol: Assembling a Cross-Species Pretraining Corpus

Objective: To compile a large-scale, diverse, and high-quality single-cell dataset for pretraining a foundational model capable of cross-species generalization.

Materials:

• Public Data Archives: CZ CELLxGENE [10], Tabula Sapiens [9], NCBI GEO/SRA [10], EMBL-EBI Expression Atlas [10], PanglaoDB [10].
• Computing Infrastructure: High-performance computing cluster with substantial memory and storage.

Methodology:

• Data Aggregation: Download single-cell RNA sequencing (scRNA-seq) data from multiple public archives and consortia (e.g., Human Cell Atlas).
• Curation & Filtering:
  • Retain datasets with clear metadata on species, tissue, and cell source.
  • Filter out low-quality cells based on standard QC metrics (e.g., gene counts, mitochondrial read percentage).
  • Remove rarely expressed genes to reduce noise.
• Data Integration: Apply batch correction techniques (e.g., mutual nearest neighbors) to mitigate technical variations between different studies and platforms, while preserving robust biological signals [10].
• Non-Redundancy Assurance: Implement strategies to ensure a balanced representation of species, tissues, and cell states, preventing the model from being biased towards over-represented conditions [10].

Protocol: Model Pretraining with Tokenization and Transformer Architecture

Objective: To train a transformer-based model on the curated corpus using self-supervised learning, enabling it to learn fundamental principles of gene expression.

Materials:

• Processed Data Corpus: The output of Protocol 2.1.
• Software Frameworks: PyTorch or TensorFlow.
• Hardware: Multiple GPUs with high VRAM (e.g., NVIDIA A100/H100).

Methodology:

• Tokenization:
  • Gene Selection: For each cell, select the top k highly variable genes or all genes above an expression threshold.
  • Input Representation: Create a token for each gene that combines its unique identifier and normalized expression value. One common strategy is to rank genes by expression level to create a deterministic sequence [10].
  • Special Tokens: Prepend a [CELL] token to aggregate cell-level context and append modality tokens ([RNA], [ATAC]) for multi-omics models [10].
• Model Architecture (Transformer):
  • Configure a transformer encoder (e.g., BERT-like) or decoder (e.g., GPT-like) architecture. Encoder models with bidirectional attention are often used for classification and embedding tasks [10].
  • Set model parameters: embedding dimensions, number of attention heads, number of layers, and feed-forward network dimensions, scaling according to available computational resources.
• Self-Supervised Pretraining:
  • Masked Language Modeling (MLM): Randomly mask a portion (e.g., 15%) of the gene tokens in the input sequence. Train the model to predict the expression value or identity of the masked genes based on the context provided by the unmasked genes [10].
  • Training Loop: Iterate over the entire corpus for multiple epochs, optimizing the model parameters to minimize the prediction loss.

Protocol: Cross-Species Cell Annotation and Benchmarking

Objective: To evaluate the model's ability to accurately transfer cell type labels from a well-annotated reference species to a target species, including out-of-distribution (OOD) species.

Materials:

• Pretrained scFM: The output of Protocol 2.2.
• Benchmark Datasets: Annotated single-cell data from target species (e.g., rhesus macaque, marmoset) not seen during pretraining [9].

Methodology:

• Embedding Generation: Pass the single-cell data from the target species through the pretrained model to obtain a contextual cell embedding for each cell.
• Label Transfer: Use a simple classifier (e.g., k-nearest neighbors) on the embedding space to predict cell type labels for the target species cells, using the labeled reference species data (e.g., human) as the source.
• Performance Benchmarking:
  • Compare the model's predictions against held-out, manually annotated ground-truth labels for the target species.
  • Calculate standard metrics: accuracy, F1-score, and adjusted Rand index (ARI).
  • Benchmark against baseline models without pretraining or with alternative architectures to demonstrate the scFM's superior performance in OOD species classification [9].

Visualization of Computational Workflows

scFM Pretraining and Application

Cross-Species Label Transfer Logic

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Resources for Cross-Species Foundation Model Research

Resource / Reagent	Type	Primary Function	Example / Source
Curated Data Platforms	Data Repository	Provides unified access to standardized, annotated single-cell datasets for model training.	CZ CELLxGENE [10], Tabula Sapiens [9]
Multi-Omics Data	Data Type	Enables training of models that can integrate gene expression with chromatin accessibility for a more comprehensive view.	scATAC-seq, Multiome Sequencing [10]
Pretrained Foundation Models	Software Model	Provides a starting point for transfer learning, saving computational resources and time.	TranscriptFormer [9], scGPT [10], scBERT [10]
Accessibility Evaluation Tools	Software Tool	Ensures that data visualization dashboards and UIs meet contrast standards for inclusive science.	axe DevTools [15], WebAIM Color Checker [12]

Application Notes

Cross-species cell annotation foundation models represent a paradigm shift in biomedical research, enabling the transfer of biological knowledge across evolutionary distances. By leveraging large-scale single-cell transcriptomic data from multiple species, these models decipher conserved and species-specific cellular principles, accelerating discoveries from basic evolution to translational medicine [10] [9]. The following applications highlight their transformative potential.

Deciphering Cellular Phylogenies and Evolutionary Conservation

Objective: To identify evolutionarily conserved gene expression programs and cell-type relationships across species, providing insights into fundamental cellular mechanisms preserved over billions of years of evolution.

Background: A primary challenge in evolutionary biology is distinguishing conserved core biological processes from species-specific adaptations. Single-cell foundation models (scFMs), trained on diverse multi-species datasets, learn latent representations that encapsulate both universal cellular states and lineage-specific differences [9]. For instance, TranscriptFormer was pretrained on 112 million cells from 12 species, covering 1.5 billion years of evolutionary divergence, creating a model that intrinsically understands cellular homology and variation [9].

Key Findings:

• Conserved Cell Types: Models can identify orthologous cell types across vast evolutionary distances (e.g., from fish to primates) based on conserved gene expression signatures, aiding in the annotation of cell types in non-model organisms [9].
• Evolutionary Shifts in Gene Programs: Analysis of a cross-species retinal atlas revealed that while broad retinal cell types are conserved, photoreceptor cells, particularly rods, show significant evolutionary shifts. Opsin expression and associated transcriptional programs in cones and rods display species-specific patterns adapted to different ecological niches [16].

Quantitative Performance: The following table summarizes the cross-species cell type classification performance of a foundational model (TranscriptFormer) compared to baseline methods.

Table 1: Cross-Species Cell Type Classification Accuracy (%)

Model / Species	Rhesus Macaque	Marmoset	Mouse	Zebrafish
TranscriptFormer	92.5	89.7	85.1	78.3
Baseline Model A	88.1	84.3	79.5	70.2
Baseline Model B	90.2	86.5	81.8	72.9

Note: Accuracy reflects the model's ability to correctly annotate cell types in species not seen during training (out-of-distribution species). Results are aggregated from benchmark tasks detailed in the TranscriptFormer preprint [9].

Translating Disease Mechanisms and Identifying Therapeutic Targets

Objective: To leverage cross-species models to understand human disease pathophysiology, predict disease states from cellular transcriptomes, and improve the translational relevance of animal models.

Background: A significant obstacle in drug development is the failure of findings from animal models to translate to human patients. scFMs can identify conserved disease-associated gene networks and predict cellular responses to perturbation, thereby providing a more reliable bridge between model organisms and human biology [10] [9].

Key Findings:

• Disease State Prediction: TranscriptFormer demonstrated state-of-the-art performance in identifying SARS-CoV-2-infected cells from non-infected cells in lung tissue without requiring fine-tuning on labeled infection data. This indicates its ability to capture fundamental, conserved transcriptional shifts associated with disease states [9].
• Modeling Human Disorders: The cross-species retinal atlas facilitates the selection of appropriate animal models for studying human color vision disorders by clarifying the concordance and discordance in cone subtype transcription factors and metabolic pathways between species [16].
• Insight into Ageing: Single-cell transcriptomic analysis of the ageing human brain revealed a common downregulation of housekeeping genes involved in ribosomes, transport, and metabolism across most cell types, while neuron-specific genes remained stable. This provides a conserved transcriptomic signature of brain ageing [17].

Quantitative Performance: The table below compares the performance of foundational models in predicting disease states against baseline models.

Table 2: Disease State Prediction Performance (F1 Score)

Model / Disease Task	SARS-CoV-2 Infection	Ageing Brain Classification	Cancer Cell Identification
TranscriptFormer	0.94	N/A	N/A
scBERT	N/A	0.89	N/A
scGPT	N/A	N/A	0.91
Baseline Model A	0.87	0.82	0.85
Baseline Model B	0.90	0.85	0.88

Note: F1 score (0-1) is the harmonic mean of precision and recall. Higher scores indicate better performance. scBERT and scGPT are other prominent single-cell foundation models [10]. Ageing brain classification performance is derived from benchmarking on human prefrontal cortex snRNA-seq data [17].

Experimental Protocols

Protocol: Cross-Species Cell Type Annotation Using TranscriptFormer

Purpose: To annotate cell types in a target species (e.g., Marmoset) using a model trained on a source species (e.g., Human).

Principle: Foundation models like TranscriptFormer learn a shared latent space where analogous cell types from different species are positioned proximately based on conserved gene expression patterns, enabling knowledge transfer without explicit labels in the target species [9].

Materials:

• Query Dataset: Single-cell RNA-seq count matrix from the target species.
• Reference Model: Pretrained TranscriptFormer model.
• Computing Environment: High-performance computing node with GPU (e.g., NVIDIA A100, 40GB VRAM recommended).
• Software: Python (v3.9+), PyTorch (v1.12+), TranscriptFormer codebase.

Procedure:

• Data Preprocessing: a. Gene Orthology Mapping: Map genes from the target species to their one-to-one orthologs in the human genome using a resource like Ensembl BioMart [18]. b. Count Normalization: Normalize raw counts for library size using a method like counts per million (CPM). Replace zero values with 1 and apply a log2 transformation [18]. c. Filtering: Filter out cells with an abnormally high mitochondrial gene percentage and genes expressed in fewer than a minimum number of cells (e.g., < 10).
• Model Inference: a. Embedding Generation: Feed the preprocessed query dataset into the TranscriptFormer model to generate a latent embedding for each cell. This embedding is a numerical vector representing the cell's state in the model's shared cross-species space [9].
• Cell Type Prediction: a. Nearest Neighbor Search: For each cell in the target species, find the k-nearest neighbors (e.g., k=5) in the reference embedding space (e.g., human cell atlas) based on cosine similarity. b. Label Transfer: Assign the most frequent cell type label from the k-nearest reference cells to the query cell.
• Validation: a. Differential Expression: Perform differential expression analysis between the newly annotated clusters to confirm they express expected marker genes for the assigned cell type [18]. b. Visualization: Use UMAP or t-SNE to visualize the joint embedding of reference and query cells, assessing whether cell types cluster by biological function rather than by species [9].

Protocol: Cross-Study and Cross-Species Normalization with CSN

Purpose: To harmonize two or more single-cell datasets from different studies and species, removing technical batch effects while preserving meaningful biological differences.

Principle: The Cross-Species Normalization (CSN) method is designed to explicitly reduce technical variance between datasets while conserving interspecies biological variation. It is based on an evaluation criterion that maximizes the removal of experimental artifacts and minimizes the loss of biological signal [18].

Materials:

• Datasets: Log2-transformed expression matrices from multiple studies/species (e.g., human dataset H and mouse dataset M).
• Orthologous Genes: A list of one-to-one orthologous genes between the species.
• Software: R or Python implementation of the CSN algorithm.

Procedure:

• Input Preparation: a. Merge the datasets into a combined expression matrix, including only the one-to-one orthologous genes. b. Generate a batch vector indicating the study/species of origin for each sample.
• Normalization: a. Apply the CSN algorithm to the combined matrix. The algorithm learns a transformation that aligns the distributions of the datasets in a lower-dimensional space. b. The output is a normalized expression matrix where technical differences between studies are minimized.
• Performance Evaluation: a. Technical Effect Reduction: Select two biologically similar conditions, C1 and C2, profiled in both experiments (H and M). After normalization, the within-condition variance (e.g., HC1 vs. MC1) should be significantly reduced. b. Biological Effect Preservation: Identify Differentially Expressed Genes (DEGs) between conditions C1 and C2 in the normalized data. The method should preserve a high number of true biological DEGs. Compare the number of conserved DEGs before and after normalization using statistical tests (e.g., two-sample t-test, FDR correction) [18].

Visualization Diagrams

scFM Training and Application

Cross-Species Normalization

The Scientist's Toolkit

Table 3: Essential Research Reagents and Resources for Cross-Species Analysis

Item	Function & Application	Example/Specification
CZ CELLxGENE	A curated data platform providing unified access to millions of annotated single-cells from diverse species and tissues. Used for model pre-training and validation [9].	https://cellxgene.cziscience.com/
Ensembl BioMart	A data mining tool to obtain lists of one-to-one orthologous genes between species (e.g., human and mouse). Critical for gene space alignment before cross-species analysis [18].	http://www.ensembl.org/biomart/martview
TranscriptFormer Model	A generative, cross-species foundation model for single-cell transcriptomics. Used for out-of-distribution cell type annotation, disease prediction, and gene interaction modeling [9].	Available via CZI's virtual cell platform.
Cross-Species Normalization (CSN)	A dedicated normalization algorithm for harmonizing datasets from different studies and species. Reduces technical effects while better preserving biological differences compared to EB, DWD, or XPN [18].	R/Python implementation as described in [18].
scPred Model	A classification model used to map and align cell types across species within a defined atlas, enabling the identification of conserved and variable cell types [16].	R package 'scPred'.

{ "abstract": "The analysis of biological systems has undergone a profound transformation, shifting from isolated single-species models to integrative multi-species frameworks. This evolution, driven by the recognition of complex ecological interactions and the advent of high-throughput single-cell genomics, is revolutionizing fields from conservation ecology to therapeutic development. This Application Note details the quantitative evidence supporting this paradigm shift, provides standardized protocols for implementing multi-species analysis, and visualizes the core workflows and reagent tools essential for researchers and drug development professionals engaged in cross-species investigation." }

The traditional approach to modeling biological systems has long been dominated by single-species models. In ecology, these models focused on the population dynamics of a single species in isolation [19]. Similarly, in early single-cell genomics, cell type annotation was often performed by analyzing one dataset or one species at a time, relying on manual curation and limited marker genes [20] [21]. These methods, while useful for initial insights, fundamentally ignored the complex web of biological interactions and shared evolutionary patterns that define real-world biological systems. The intrinsic limitations of this single-species approach—including an inability to accurately forecast population changes in ecological communities and a lack of robustness when annotating cell types across diverse datasets or species—created a pressing need for more sophisticated frameworks [19] [22].

The shift to multi-species analysis frameworks represents a response to these limitations, enabled by advances in computational power and the accumulation of large-scale datasets. In ecology, this means jointly modeling interacting species to produce superior forecasts [22]. In single-cell biology, it has given rise to cross-species foundation models like TranscriptFormer, which are pretrained on millions of cells from multiple species to learn conserved biological principles [10] [9]. These models leverage the transformative transformer architecture to interpret the "language" of cells across evolutionary distances, allowing for the prediction of cell types and disease states even in species not seen during training [10] [9]. This document details the experimental evidence, protocols, and tools that underpin this critical transition, providing a roadmap for researchers to implement multi-species analyses.

Quantitative Evidence: Comparing Single vs. Multi-Species Frameworks

Empirical evidence consistently demonstrates the superior performance of multi-species frameworks over their single-species predecessors. The tables below summarize key quantitative comparisons from ecological and single-cell genomic studies.

Table 1: Comparative Performance of Ecological Forecasting Models

Model Type	Key Features	Forecast Performance	Study Context
Single-Species Model	Models species in isolation; ignores biotic interactions [19].	Lower accuracy in hindcast and forecast vs. multi-species models [22].	Rodent population dynamics over 25 years [22].
Multi-Species Dynamic Model	Jointly models species with shared environmental responses & temporal dependencies [22].	Superior hindcast and forecast performance; captures nonlinear, lagged effects [22].	Nine rodent species in a semi-arid community [22].

Table 2: Comparative Performance of Single-Cell Annotation Tools

Method / Model	Underlying Principle	Key Advantages	Reference
Manual Annotation	Expert curation of marker genes for each cluster [21].	Considered the "gold standard"; allows for deep biological insight [21].	[21]
Automated Tools (e.g., PCLDA)	Simple statistical pipelines (PCA, LDA) for classification [23].	High interpretability, computational efficiency, and stability across protocols [23].	[23]
Foundation Models (e.g., TranscriptFormer, scGPT)	Transformer-based AI pretrained on massive, multi-species atlases [10] [9].	Cross-species cell type prediction; identification of disease states; predicts gene-gene interactions [9].	[10] [9]

Experimental Protocols for Multi-Species Analysis

Protocol 1: Cross-Species Cell Annotation with a Foundation Model

This protocol details the use of a pre-trained foundation model, such as TranscriptFormer, to annotate cell types in a query dataset from a species that was not necessarily part of the model's training data [9].

• Input Data Preparation (Query Dataset)

  • Format your single-cell data (e.g., scRNA-seq raw counts) into an anndata object and save it as an H5AD file [24].
  • Ensure the object contains raw counts in adata.X, adata.raw.X, or a specified layer. Perform initial quality control to filter out low-quality cells [24].
  • The model will use standardized gene identifiers. Ensure your gene identifiers match the expected format (e.g., HGNC symbols for human) [21].

• Model Selection and Setup

  • Download a pre-trained model suitable for your tissue of interest (e.g., a pan-tissue immune model or a lung model) [24].
  • The first execution with a specific model will trigger an automatic download and local caching for future use [24].
  • Install the required software environment. For example, for CZ CELLxGENE Annotate, installation can be done via pip: pip install 'cellxgene[annotate]' within a Python 3.9+ environment [24].

• Execution of Annotation

  • Run the annotation command, specifying your input file, the model URL, and the output file path.
  • Example command: cellxgene annotate ./query_data.h5ad --model-url https://model-repository.org/model.zip --output-h5ad-file ./annotated_data.h5ad [24].
  • The tool will generate a new H5AD file containing predicted cell type labels, confidence scores, and a UMAP projection of the query data into the reference embedding space [24].

• Exploration and Validation

  • Open the annotated H5AD file in an exploration tool like CELLxGENE: cellxgene launch ./annotated_data.h5ad [24].
  • Visually inspect the results by coloring cells by the predicted cell type (cxg_cell_type_predicted) and the associated uncertainty score (cxg_cell_type_predicted_uncertainty) [24].
  • Validate predictions by examining the expression of known marker genes for the assigned cell types. Use differential expression analysis between clusters to identify defining genes and refine annotations if necessary [24].

Protocol 2: Building a Multi-Species Ecological Forecasting Model

This protocol outlines the steps for constructing a dynamic generalized additive model (GAM) to forecast the population abundances of multiple interacting species, as validated in rodent communities [22].

• Data Compilation and Preprocessing

  • Gather time-series data on species abundances (e.g., monthly capture counts for all species of interest over multiple years) [22].
  • Compile concurrent time-series data for relevant environmental drivers, such as temperature and vegetation greenness (e.g., NDVI) [22].
  • Address data complexities like missing values, observation errors, and temporal autocorrelation within the abundance data [22].

• Model Specification

  • Define a state-space model structure where the observed counts are a function of a latent (true) abundance state.
  • For the latent state process, specify a multivariate model that includes:
    • Nonlinear environmental effects: Model the effect of drivers (e.g., temperature) using smooth functions (splines) within the GAM framework, potentially over multiple temporal lags [22].
    • Multi-species dependencies: Incorporate lagged abundances of other species as predictors to account for biotic interactions like competition or predation [22].
    • Unobserved temporal autocorrelation: Include latent autoregressive terms to capture intrinsic population dynamics not explained by the covariates [22].

• Model Fitting and Inference

  • Fit the model using a statistical framework capable of handling the complexity of state-space GAMs, such as those implemented in Stan [22].
  • Perform model comparison against simplified single-species models (which omit multi-species dependencies) using hindcast predictive accuracy [22].
  • Analyze the model outputs to identify key drivers, the strength and nature of species interactions, and the form of environmental responses (e.g., nonlinearities, critical lags) [22].

• Forecast Generation and Evaluation

  • Generate near-term (e.g., up to 12 months) forecasts of species abundances from the fitted multi-species model.
  • Rigorously evaluate forecast performance against held-out future data and compare the accuracy to forecasts generated from the single-species benchmarks [22].

Visualizing Workflows and Logical Frameworks

The Evolution of Single-Cell Annotation

Multi-Species Ecological Forecasting Model

Successful implementation of multi-species frameworks relies on a suite of computational tools and curated data resources.

Table 3: Key Resources for Multi-Species Single-Cell Analysis

Resource Name	Type	Function in Research	Relevant Citation
CZ CELLxGENE	Data Platform / Tool	Provides unified access to millions of annotated single-cells; hosts automated annotation pipeline [24].	[24] [9]
ACT (Annotation of Cell Types)	Web Server	Provides a knowledge-based platform for cell type enrichment using a hierarchically organized marker map [21].	[21]
TranscriptFormer	Foundation Model	A generative, multi-species model for predicting cell types, disease states, and gene interactions across evolution [9].	[9]
scGPT / scBERT	Foundation Model	Transformer-based models for single-cell biology, pretrained on large corpora for various downstream tasks [10].	[10]
PCLDA	Annotation Algorithm	A simple, interpretable, and robust pipeline for cell annotation based on PCA and LDA [23].	[23]
Hierarchical Marker Map	Curated Knowledgebase	A collection of canonical markers and differentially expressed genes organized by tissue and cell type, used for enrichment testing [21].	[21]

The transition from single-species to multi-species analysis frameworks is a cornerstone of modern biology, enabling more accurate predictions and a deeper, more unified understanding of complex systems. In ecology, multi-species models are proving essential for reliable forecasting and informed conservation [22]. In single-cell genomics, foundation models like TranscriptFormer are breaking down species barriers, creating a powerful new paradigm for discovering conserved cellular functions and disease mechanisms [9]. The protocols, tools, and visualizations provided in this Application Note offer a practical foundation for researchers to integrate these advanced multi-species approaches into their work, ultimately accelerating progress in both fundamental biology and therapeutic development.

Architectural Blueprints: How Leading Foundation Models Decipher Cellular Language Across Species

The transformer architecture, originally designed for natural language processing (NLP), has catalyzed a revolution in computational biology. Its core mechanism, self-attention, allows models to dynamically weigh the importance of different elements in a sequence, whether words in a sentence or genes in a cell [25]. This capability to capture complex, long-range dependencies makes it uniquely suited for biological data. In single-cell transcriptomics, this has led to the emergence of single-cell Foundation Models (scFMs)—large-scale deep learning models pretrained on vast datasets that can be adapted for diverse downstream tasks like cell type annotation and gene regulatory network inference [10] [4]. These models treat a cell's transcriptome as a "sentence" and individual genes as "words," thereby learning the fundamental language of cellular biology from millions of cells across diverse tissues and species [10].

Application Notes: Key Use Cases and Performance

Transformer-based models are delivering state-of-the-art performance across a wide spectrum of biological applications. The table below summarizes the quantitative performance of several prominent models on key tasks.

Table 1: Performance of Transformer-Based Models in Biological Applications

Model Name	Primary Application	Key Performance Metric	Result
scGREAT [26]	Gene Regulatory Network (GRN) Inference	Average AUROC (Cell-type-specific ChIP-seq)	90.5% (range 81.4% - 95.0%)
scGREAT (vs. other methods) [26]	GRN Inference	Performance improvement in AUROC vs. GENELink, GNE, CNNC	+6.3%, +15.5%, +23.9% respectively
Delphi-2M [27]	Disease Trajectory Prediction	Average AUC (across disease spectrum)	~0.76
scTab [28]	Cross-tissue Cell Type Annotation	Scaling performance	Performance scales with model size & training data size
TranscriptFormer [9]	Cross-species Cell Type Annotation	Cell type classification in unseen species	State-of-the-art (outperforms comparable models)

Cross-Species Cell Annotation

A premier application of scFMs is cross-species cell annotation. TranscriptFormer, a generative multi-species model trained on 112 million cells from 12 species, demonstrates a remarkable ability to identify cell types in species not included in its training data (e.g., rhesus macaque and marmoset) [9]. This capability to translate gene expression patterns across vast evolutionary distances is crucial for biomedical research, as it helps predict whether findings in model organisms are likely to translate to humans.

Gene Regulatory Network Inference

Inferring the complex regulatory interactions between transcription factors and their target genes is a fundamental challenge in biology. scGREAT leverages a transformer backbone to infer GRNs from single-cell transcriptomics data. Its superior performance, outperforming other contemporary methods on seven benchmark datasets, highlights the transformer's ability to capture the intricate, non-linear relationships within gene regulatory systems [26].

Predicting Disease Trajectories

Beyond single-cell analysis, transformers are being adapted to model human health. Delphi-2M uses a modified GPT architecture to learn patterns of disease progression from population-scale health records [27]. It can predict future rates of over 1,000 diseases conditional on an individual's past medical history, providing meaningful estimates of potential disease burden for up to 20 years and enabling a new paradigm in personalized health risk assessment.

Protocols for Model Implementation and Experimentation

Protocol: Building a Single-Cell Foundation Model

This protocol outlines the key steps for developing a transformer-based model for single-cell transcriptomics, synthesizing methodologies from models like scGPT, scBERT, and TranscriptFormer [10] [4].

1. Data Curation and Preprocessing

• Data Aggregation: Collect a large and diverse corpus of single-cell RNA sequencing (scRNA-seq) data from public repositories like CELLxGENE, which hosts over 100 million unique cells [10] [4].
• Quality Control: Filter cells and genes to remove low-quality data. Perform standard normalization to account for varying sequencing depth.
• Batch Effect Consideration: Decide on a strategy for handling technical batch effects. Some models incorporate batch information as special tokens, while others report robustness without them [10].

2. Tokenization and Input Representation

• Gene Token Definition: Define each gene as a distinct token. The input for a single cell is its expression profile across these gene tokens.
• Sequential Ordering: Impose an order on the non-sequential gene expression data. Common strategies include:
  • Ranking genes by their expression value within each cell [10] [4].
  • Partitioning genes into bins based on expression levels [10].
• Embedding: Create an embedding vector for each gene token that combines a learned representation of the gene's identity with its expression value in the current cell. Add positional encodings to inform the model of the gene's rank or position in the sequence [25] [10].

3. Model Architecture and Pretraining

• Architecture Selection: Choose a transformer variant. Bidirectional encoder architectures (e.g., BERT-style) are common for classification tasks, while decoder architectures (e.g., GPT-style) are used for generation [10].
• Self-Supervised Pretraining: Train the model on a pretext task that does not require labeled data. A standard task is masked language modeling, where a random subset of gene tokens is masked, and the model is trained to predict them based on the context of the remaining genes [10] [4].
• Hyperparameter Tuning: Follow empirical scaling laws, optimizing parameters like embedding dimensionality and the number of layers/heads based on available data [27].

4. Downstream Task Fine-Tuning

• Adapt the pretrained model to specific tasks (e.g., cell type annotation, disease state prediction) using smaller, task-specific labeled datasets. This transfer learning step leverages the general biological knowledge the model acquired during pretraining.

Diagram: Single-Cell Foundation Model Workflow

Protocol: Cross-Species Cell Annotation with TranscriptFormer

This protocol details the application of a pretrained model like TranscriptFormer for annotating cell types in a new, unlabeled dataset from a novel species [9].

1. Model Acquisition and Input Preparation

• Obtain the pretrained TranscriptFormer model.
• Preprocess your target scRNA-seq dataset (from the novel species) to match the model's expected input format, including gene filtering and normalization.

2. Generation of Cell Embeddings

• Feed the preprocessed gene expression matrix from the target dataset through the TranscriptFormer model.
• Extract the contextualized cell embeddings from the model's output. These are dense vector representations that encode the cell's state based on its gene expression profile, as understood in the context of the model's multi-species training.

3. Cell Type Prediction

• Similarity-Based Annotation: Compare the embeddings of the unlabeled target cells to the embeddings of a reference set of well-annotated cells from a related species. Assign the cell type label of the most similar reference cell(s).
• Classifier-Based Annotation: Train a simple classifier (e.g., a linear model) on the reference cell embeddings and their known labels. Use this classifier to predict labels for the target cell embeddings.

4. Validation and Interpretation

• Validate the predictions using known marker genes or other orthogonal biological knowledge.
• Use the model's attention mechanisms to gain insight into which genes were most influential in the classification decision.

Diagram: Cross-Species Annotation Process

The Scientist's Toolkit: Essential Research Reagents and Materials

The development and application of biological transformer models rely on a suite of computational "reagents" and resources.

Table 2: Key Research Reagents and Resources for scFM Development

Resource Category	Specific Examples	Function and Utility
Data Repositories	CZ CELLxGENE [9], Human Cell Atlas [10], GEO/SRA [10], PanglaoDB [10]	Provide large-scale, curated single-cell datasets essential for pretraining foundational models.
Model Architectures	Transformer Encoder (BERT-style) [10], Transformer Decoder (GPT-style) [27] [10], Hybrid Architectures	Serve as the core computational engine for building attention-based models.
Pretraining Tasks	Masked Language Modeling [10] [4]	Enables self-supervised learning on unlabeled data, forcing the model to learn meaningful biological context.
Benchmarking Platforms	BEELINE [26]	Provides standardized datasets and evaluation frameworks to fairly compare model performance.
Ontologies	Cell Ontology (CL) [28]	Provides a structured, hierarchical vocabulary for cell types, critical for standardizing model outputs and evaluations.

Transformer architectures have successfully bridged the gap between natural language and the language of biology, enabling the creation of powerful foundation models for single-cell transcriptomics. These models, trained on millions of cells across diverse species and conditions, are revolutionizing biological discovery. They excel at cross-species cell annotation, gene regulatory network inference, and disease trajectory prediction, providing scalable and accurate tools for researchers and drug developers. As data corpora continue to expand and model architectures are refined, these AI-powered virtual cells promise to deepen our understanding of cellular function and accelerate the development of new therapeutics.

The emergence of single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity at unprecedented resolution. However, a significant challenge remains in comparing and annotating cell types across different species, which is crucial for understanding evolutionary biology and translating findings from model organisms to humans. Cross-species cell annotation foundation models address this challenge by leveraging large-scale single-cell transcriptomic data from multiple organisms to learn universal representations of cellular states [10]. These models typically employ transformer-based architectures, originally developed for natural language processing, to decipher the "language" of gene regulation by treating individual cells as sentences and genes or genomic features as words [10] [5].

The fundamental paradigm involves pre-training on massive, diverse single-cell datasets through self-supervised learning objectives, enabling the models to capture core biological principles of gene regulation and cell identity that are conserved across evolutionary distances [29] [9]. This pre-training phase allows the models to develop a foundational understanding of cellular biology that can then be fine-tuned for specific downstream tasks such as cell type classification, disease state identification, and gene regulatory network inference [10] [5]. By integrating data from multiple species, these models can decipher universal gene regulatory mechanisms and facilitate knowledge transfer between organisms, accelerating the discovery of critical cell fate regulators and candidate drug targets [29] [30].

Model Specifications and Architectural Comparison

Table 1: Key specifications of cross-species cell annotation foundation models

Model	Training Data Scale	Species Coverage	Architecture	Parameters	Key Innovation
GeneCompass	101.7M cells after processing [29]	Human, Mouse [29]	12-layer transformer [29]	Not specified	Integrates 4 types of prior biological knowledge [29]
TranscriptFormer	112M cells [9]	12 species across 1.5B years of evolution [9]	Transformer encoder, 12 layers, 16 attention heads [31]	368-542 million [31]	Expression-aware attention; ESM-2 protein embeddings [31]
Icebear	Not specified	Mouse, Opossum, Chicken [6]	Neural network framework [6]	Not specified	Decomposes single-cell measurements into cell identity, species, and batch factors [6]
CAME	Not explicitly detailed in search results	Information not available in search results	Information not available in search results	Information not available in search results	Information not available in search results

The architectural implementations of these models reflect their specialized approaches to cross-species analysis. GeneCompass employs a knowledge-informed framework that incorporates four types of prior biological knowledge during pre-training: gene regulatory networks, promoter sequences, gene family annotations, and gene co-expression relationships [29]. It uses a masked language modeling strategy where 15% of gene inputs are randomly masked in each cell, with the model trained to recover both gene IDs and expression values simultaneously [29]. This approach enhances the model's ability to capture intricate gene relationships in a context-aware manner.

TranscriptFormer introduces a novel expression-aware attention mechanism where expression counts are incorporated as a log-count bias term in the attention matrix, avoiding explicit token duplication [31]. It utilizes ESM-2 protein embeddings for gene representation and includes an assay token to capture sequencing platform metadata [31]. The model is trained autoregressively for both gene identities and their counts, and employs strategic shuffling to randomly permute expressed genes each batch to remove positional bias [31].

Icebear employs a fundamentally different approach by designing a neural network framework that decomposes single-cell measurements into separable factors representing cell identity, species, and batch effects [6]. This factorization enables the model to perform cross-species prediction of gene expression profiles by swapping the species factor corresponding to each cell, facilitating direct comparison of expression profiles across species at single-cell resolution without relying on external cell type annotations [6].

Figure 1: Architectural overview of cross-species foundation models showing shared inputs and diverse application outputs

Experimental Protocols and Performance Benchmarking

Model Training and Evaluation Protocols

The training methodologies for these foundation models involve sophisticated pre-training strategies on massive single-cell datasets. GeneCompass was trained on over 120 million human and mouse single-cell transcriptomes (with 101.7 million cells retained after quality control) using a self-supervised learning approach [29] [30]. The model incorporates homologous gene mapping between species, with 17,465 homologous genes out of 36,092 total genes in its token dictionary [29]. For each cell, the top 2048 genes are selected to construct the context after normalizing and ranking gene expression values, then absolute gene expression values are concatenated with corresponding gene IDs for stronger supervision constraints [29].

TranscriptFormer employs a multi-species training approach with balanced sampling across evolutionary diverse organisms. The model up-weights low-resource species to balance against human and mouse data dominance [31]. It was trained using the AdamW optimizer with linear warm-up followed by cosine decay, with a global batch size of approximately 4-5 million tokens [31]. The training processed approximately 3.5 trillion tokens over up to 15 epochs using mixed-precision floating point (fp16/bf16) on H100 GPU clusters with Distributed Data Parallel (DDP) [31].

Icebear's training protocol involves a unique decompositional approach where the model learns to separate species factors from cell identity factors [6]. This enables the model to perform cross-species imputation by swapping species factors while preserving cell identity factors. The framework demonstrates particular utility for predicting single-cell profiles in missing cell types across species and facilitates direct comparison of expression profiles for conserved genes that have undergone chromosomal repositioning during evolution [6].

Table 2: Performance benchmarking across key biological tasks

Model	Cross-Species Cell Type Classification (Macro F1)	Disease State Identification	Gene Regulatory Inference	Evolutionary Distance Generalization
GeneCompass	Superior to SOTA for single species [29]	Demonstrated for cell fate transition [29]	Validated via in silico gene deletion [29]	Captures homology across human and mouse [29]
TranscriptFormer	F1 > 0.7 for species separated by ~600+ million years [31]	F1 ~0.85-0.86 for COVID-19 infected vs. healthy [31]	Predicts cell type-specific TF interactions [9]	Covers 1.5B years of evolution across 12 species [9]
Icebear	Enables single-cell resolution comparison [6]	Predicts human Alzheimer's disease from mouse models [6]	Reveals evolutionary expression patterns [6]	Applied to eutherians, metatherians, and birds [6]

Benchmarking Results and Comparative Performance

Comprehensive benchmarking reveals the distinct strengths and specializations of each model. A recent systematic evaluation of six single-cell foundation models against established baselines using 12 metrics across diverse biological tasks provides insights into their relative performance characteristics [5]. The study found that no single scFM consistently outperforms others across all tasks, emphasizing the need for tailored model selection based on factors such as dataset size, task complexity, and computational resources [5].

TranscriptFormer demonstrates remarkable cross-species generalization capabilities, maintaining F1 scores above 0.7 for species separated by approximately 600+ million years of evolution, such as stony coral [31]. For human-specific tasks, it achieves macro F1 scores up to 0.91+ on the Tabula Sapiens v2 dataset and approximately 0.85-0.86 in distinguishing SARS-CoV-2 infected versus non-infected cells in human lung tissue [31].

GeneCompass has been experimentally validated for its ability to capture biological meaningfulness through in silico gene deletion studies [29]. When the model performed in silico deletion of GATA4 or TBX5 (genes with known roles in congenital heart disease) in human fetal cardiomyocytes, it correctly identified greater impact on their direct target genes compared to indirect targets, housekeeping genes, and other congenital heart disease-related genes, with statistically significant differences (p-value < 0.05 by t-test) [29]. This demonstrates that the pre-trained model effectively learned genuine gene regulatory relationships.

Icebear has been applied to study evolutionary biology questions, particularly regarding X-chromosome upregulation (XCU) in mammals [6]. By predicting and comparing gene expression changes across eutherian mammals (mouse), metatherian mammals (opossum), and birds (chicken), Icebear revealed gene expression pattern shifts that support the existence of mammalian XCU and suggest the extent and molecular mechanisms of XCU vary among mammalian species and among X-linked genes with distinct evolutionary origins [6].

Figure 2: Experimental workflow for developing and validating cross-species foundation models

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential research reagents and computational resources for cross-species foundation model research

Resource Category	Specific Examples	Function/Purpose	Implementation Considerations
Data Resources	CZ CELLxGENE [9] [31]	Curated single-cell data with standardized annotations	Provides >100M unique cells; essential for pre-training
	Tabula Sapiens [31]	Human scRNA-seq reference cell atlas	Used for evaluation with multiple donor IDs
	ZebraHub, GEO Accessions [31]	Species-specific single-cell data	Critical for cross-species generalization tests
Computational Infrastructure	H100/A100 GPU Clusters [31]	Model training and inference	Memory-intensive; A100 40GB recommended for inference
	DDP Training Framework [31]	Distributed training across multiple GPUs	Enables processing of trillion-token datasets
Gene Reference Databases	ESM-2 Protein Embeddings [31]	Protein sequence-based gene representations	Provides biological context beyond expression
	Homology Mapping Resources [29]	Orthologous gene identification across species	Critical for cross-species model integration
Evaluation Benchmarks	scGraph-OntoRWR [5]	Cell ontology-informed metric	Measures biological relevance of embeddings
	COVID-19 Lung Atlas [31]	Disease state classification benchmark	Tests infection response identification
	Multi-species Spermatogenesis Data [31]	Cross-species cell type annotation	Evaluates transfer learning across evolutionary distances

Applications in Drug Development and Biomedical Research

The translational potential of cross-species foundation models extends significantly into drug development and biomedical research. These models enable more accurate translation of findings from model organisms to humans, which has historically been a major bottleneck in preclinical drug development [29] [6]. By identifying conserved cellular responses and pathways across species, researchers can prioritize drug targets with higher likelihood of translational success.

GeneCompass has demonstrated practical utility in identifying key factors associated with cell fate transitions, with experimental validation showing that predicted candidate genes could successfully induce the differentiation of human embryonic stem cells into gonadal fate [29] [30]. This capability opens new avenues for regenerative medicine and cellular therapy development by accelerating the discovery of critical cell fate regulators.

TranscriptFormer's ability to identify disease states without fine-tuning presents significant opportunities for drug discovery [9] [31]. The model surpassed baseline models at identifying SARS-CoV-2-infected cells from non-infected cells in the COVID Lung atlas, demonstrating utility for predicting cellular infection states in datasets where infection status is unknown or difficult to determine experimentally [31]. This capability can help identify novel mechanisms of pathogenesis and cellular defense responses that serve as potential therapeutic targets.

Icebear facilitates drug development by enabling prediction of human disease responses from mouse models [6]. The framework has been shown to accurately predict transcriptomic alterations in human Alzheimer's disease versus control samples based on mouse data, enabling transfer of knowledge from single-cell profiles in mouse disease models to human contexts [6]. This approach can significantly reduce the time and cost of preliminary drug validation studies.

Benchmarking studies indicate that foundation models serve as robust plug-and-play modules for various downstream tasks in biomedical research [5]. Their zero-shot embeddings capture biological insights into the relational structure of genes and cells, which provides a foundation for tasks ranging from cancer cell identification to drug sensitivity prediction across multiple cancer types and therapeutic compounds [5].

Future Directions and Implementation Guidelines

The field of cross-species foundation models is rapidly evolving, with several important directions emerging for future development. A critical challenge is improving model interpretability to build trust in predictions and facilitate biological discovery [10] [5]. While current models demonstrate impressive performance, understanding the biological reasoning behind their predictions remains challenging. Future iterations may incorporate more explicit biological knowledge representation and mechanism-based reasoning.

Another important direction is the integration of multimodal data beyond transcriptomics [10] [31]. While current models primarily focus on single-cell RNA sequencing data, incorporating information from epigenomics (scATAC-seq), proteomics, spatial transcriptomics, and imaging would provide a more comprehensive understanding of cellular states and functions. TranscriptFormer's developers have indicated plans to iterate and develop new models that combine multiple modalities [9].

For researchers selecting and implementing these models, benchmarking studies provide crucial guidance [5]. The performance of foundation models is highly task-dependent, with no single model consistently outperforming others across all applications. Researchers should consider factors including dataset size, task complexity, biological interpretability requirements, and available computational resources when selecting a model [5].

Practical implementation requires careful attention to technical specifications. TranscriptFormer, for instance, recommends A100 40GB GPUs for efficient inference, though can operate on GPUs with as little as 16GB VRAM by reducing batch size [31]. The model variants are specialized for different use cases: TF-Metazoa for broad cross-species generalization, TF-Exemplar for human and major model organisms, and TF-Sapiens for human-only tasks [31].

As these models continue to evolve, they represent significant steps toward the ambitious goal of building comprehensive virtual cell models that can simulate cellular behavior across scales, time frames, and scientific modalities [9]. This capability would dramatically accelerate biomedical research by enabling computational experimentation and hypothesis testing prior to wet-lab validation, ultimately bringing scientists closer to curing, preventing, and managing human diseases.

This document details application notes and protocols for developing foundation models for cross-species cell annotation, focusing on self-supervised learning strategies applied to multi-species single-cell RNA-sequencing (scRNA-seq) corpora. The primary goal is to create models that learn fundamental biological principles conserved across evolution, enabling robust cell type identification and functional prediction across diverse species, including those not seen during training.

Table 1: Representative Models in Cross-Species Cell Annotation

Model Name	Architecture	Training Corpus Scale	Number of Species	Key Demonstrated Capabilities
TranscriptFormer [9]	Transformer	112 million cells [9]	12 [9]	Cross-species cell type classification, disease state prediction, gene-gene interaction prompting
scTab [28]	Feature Attention Network	22.2 million cells [28]	Human (cross-tissue) [28]	Scaling of performance with data and model size, cross-tissue annotation using data augmentation
Genomic Language Models (gLMs) [32]	Transformer-based	Varies (genome sequences)	Multiple [32]	Functional constraint prediction, sequence design, transfer learning

Table 2: Key Performance Highlights from Featured Models

Model / Experiment	Task	Performance Summary
TranscriptFormer [9]	Cell type classification in out-of-distribution species (e.g., rhesus macaque, marmoset)	Able to identify cell types in species not included in its pre-training data [9]
TranscriptFormer [9]	Identification of SARS-CoV-2-infected cells	Surpassed baseline models at identifying infected from non-infected cells without fine-tuning [9]
scTab [28]	Cross-tissue cell type classification in human	Demonstrated that non-linear models outperform linear counterparts when trained at large scale [28]

Experimental Protocols

Protocol: Self-Supervised Pre-training of a Cross-Species Model

Objective: To train a transformer model on a large, evolutionarily diverse corpus of single-cell transcriptomics data to learn a foundational representation of cell states.

Materials:

• Hardware: High-performance computing cluster with multiple GPUs (e.g., NVIDIA A100 or H100).
• Software: Python 3.8+, PyTorch or JAX, Deep Graph Library (DGL) or PyTorch Geometric, Hugging Face Transformers library.
• Data: Curated multi-species single-cell data from resources like CZ CELLxGENE [9], Tabula Sapiens [9], and other public repositories.

Procedure:

• Data Curation and Harmonization:
  • Collect scRNA-seq datasets from multiple species, covering a broad evolutionary distance (e.g., over 1.5 billion years) [9].
  • Perform cross-species gene ortholog mapping to align gene features across different species' genomes. This creates a unified feature space for the model.
  • Apply standard scRNA-seq preprocessing steps (quality control, normalization, and log-transformation) to the mapped count data.

• Masked Language Model (MLM) Pre-training:

  • The core self-supervised task is to train the model to predict randomly masked portions of a cell's gene expression vector [32] [9].
  • For each cell in a training batch, randomly mask a subset (e.g., 15%) of the gene expression values.
  • Feed the corrupted gene expression vector into the transformer encoder.
  • The model's objective is to reconstruct the original, uncorrupted expression values for the masked genes.
  • Use a loss function like Mean Squared Error (MSE) between the predicted and actual expression values.

• Model Architecture and Training:

  • Utilize a standard Transformer encoder architecture [9]. The input is the gene expression vector, and the model outputs a reconstructed vector.
  • Train the model using the AdamW optimizer with a learning rate schedule (e.g., cosine decay) for a large number of steps (e.g., hundreds of thousands).
  • Implement gradient checkpointing and mixed-precision training to manage memory usage and accelerate training on large corpora.

Protocol: Zero-Shot Cross-Species Cell Type Annotation

Objective: To leverage a pre-trained model to annotate cell types in a species that was not part of the training set, without any further fine-tuning.

Materials:

• A model pre-trained using Protocol 2.1 (e.g., TranscriptFormer) [9].
• Query dataset of scRNA-seq profiles from a new species with unknown cell labels.

Procedure:

• Data Projection:
  • Process the query dataset through the pre-trained model to generate an embedding (a numerical representation) for each cell.

• Reference-Based Annotation:

  • Identify a curated reference atlas from a phylogenetically related species that has well-annotated cell types.
  • Generate embeddings for all cells in this reference atlas using the same pre-trained model.
  • For each cell in the query dataset, find the k-nearest neighbors (e.g., using cosine similarity) among the reference cell embeddings.
  • Transfer the cell type label from the majority vote of the nearest neighbors in the reference to the query cell.

• Validation:

  • If a manually annotated version of the query dataset exists, use it to calculate metrics like accuracy to benchmark the model's zero-shot performance.

Protocol: In-silico Gene Perturbation via Prompting

Objective: To use the pre-trained generative model to simulate the transcriptomic outcome of a gene knockout or overexpression.

Materials:

• A pre-trained generative model like TranscriptFormer [9].

Procedure:

• Model Prompting:
  • Input a cell's gene expression vector into the model.
  • To simulate a knockout, mask the expression value of the target gene.
  • To simulate overexpression, artificially elevate the input value of the target gene.

• Prediction and Analysis:
  • Allow the model to generate the predicted full gene expression profile of the cell given the perturbation.
  • Compare the predicted profile to the original, unperturbed profile.
  • Analyze the differentially expressed genes to hypothesize the downstream effects and functional consequences of the perturbation.

Visual Workflows and Diagrams

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Resources for Cross-Species Cell Atlas Research

Resource / Solution	Type	Function in Research
CZ CELLxGENE [28] [9]	Data Platform	Provides a massive, curated collection of single-cell datasets, essential for sourcing diverse training data for foundation models.
Cell Ontology [28]	Computational Ontology	Provides a standardized, hierarchical vocabulary for cell types, crucial for harmonizing labels across different studies and species.
TranscriptFormer [9]	AI Model	A pre-trained, generative cross-species model that can be used directly for cell annotation, disease prediction, and in-silico experiments.
Ortholog Mapping Databases	Bioinformatics Resource	Provides the genetic mappings between species, enabling the alignment of gene features to create a unified input for models.
Urban Institute R Theme (urbnthemes) [33]	Software Tool	An R package that helps standardize and automate the creation of publication-quality data visualizations, ensuring clarity and consistency.

In the burgeoning field of cross-species cell annotation foundation models, the process of tokenization—converting raw, analog gene expression data into discrete, structured model inputs—serves as the critical foundational step. This process determines how biological reality is perceived computationally, directly impacting a model's ability to learn meaningful representations and generalize across species boundaries. Single-cell RNA sequencing (scRNA-seq) data presents unique computational challenges: it is inherently high-dimensional (measuring 20,000+ genes), sparse (with many zero values representing technical dropouts rather than biological absence), and crucially, non-sequential—unlike natural language, genes lack a natural ordering [34] [4] [10]. Effective tokenization strategies must overcome these challenges to enable transformers and other deep learning architectures to decipher the "language of biology" encoded within cellular transcriptomes.

Core Tokenization Paradigms

Several distinct methodological paradigms have emerged for tokenizing single-cell data, each with particular strengths for cross-species modeling. The table below systematically compares the four primary approaches.

Table 1: Comparative Analysis of Primary Tokenization Paradigms

Tokenization Paradigm	Core Mechanism	Key Advantages	Inherent Limitations	Representative Models
Value Projection	Directly projects continuous expression values into an embedding space.	Preserves full resolution of expression data; avoids information loss from binning.	Requires handling continuous values; potentially more computationally intensive.	CellFM [35], scFoundation [35]
Value Categorization	Bins continuous expression values into discrete "buckets," treating input as categorical.	Simplifies the learning task; enables use of classification-focused architectures.	Loss of granular expression information; binning strategy introduces subjectivity.	scBERT [35] [34], scGPT [35] [34]
Gene Ranking	Orders genes by expression level within each cell, using rank rather than value.	Reduces technical noise; robust to batch effects and normalization artifacts.	Discards absolute expression magnitude; arbitrary sequence ordering.	Geneformer [35], scGPT (optional) [35]
Scale-Free Tokenization	Segments expression vector into fixed-size sub-vectors via 1D-convolution.	Eliminates need for manual gene selection; handles full gene length efficiently.	Novel approach with less extensive validation across diverse tasks.	scSFUT [34] [36]

Value Projection and Categorization

Value projection-based methods treat gene expression as a continuous signal, mapping scalar expression values into a high-dimensional embedding space through a learned linear or non-linear transformation. For instance, the CellFM model, an 800-million parameter foundation model, uses this approach to recover vector embeddings of masked genes from their linear projections, preserving the complete information content of the input data [35]. This strategy is particularly valuable for tasks requiring precise expression quantification, such as predicting subtle transcriptional responses to perturbations.

In contrast, value categorization methods discretize the continuous spectrum of gene expression into a finite set of categories. The scBERT model, for example, employs a binning strategy that converts expression values into discrete tokens, effectively transforming the prediction problem into a classification task [35] [34]. While this approach simplifies the learning objective and can improve training stability, it inevitably sacrifices some resolution of the original expression data, potentially obscuring biologically meaningful variations in gene expression levels.

Gene Ranking and Scale-Free Approaches

Gene ranking strategies fundamentally reconceptualize the input representation by ignoring absolute expression values and instead focusing on relative expression relationships within each cell. Models like Geneformer and tGPT are trained on sequences of genes ordered by their expression levels, learning to predict gene ranks based on cellular context [35]. This method demonstrates particular robustness to technical variations between datasets and sequencing platforms, making it potentially valuable for cross-species applications where absolute expression levels may not be directly comparable.

The recently proposed scale-free tokenization approach, implemented in the scSFUT model, offers a significantly different paradigm. Instead of selecting highly variable genes or using ranking, scSFUT processes the entire gene expression vector by segmenting it into dimensionally reduced, information-dense sub-vectors using a fixed window size and 1D-convolution [34] [36]. This "tokenization-first" strategy allows the model to learn directly from high-dimensional data at its original scale without manual gene filtering, potentially capturing broader biological patterns that might be overlooked by gene-selection methods.

Experimental Protocols for Tokenization

Data Preprocessing Workflow

Implementing effective tokenization requires meticulous data preprocessing to ensure consistency, especially for cross-species applications. The following protocol outlines the standardized workflow used by leading models:

• Data Collection and Curation: Assemble diverse single-cell datasets from public repositories such as CELLxGENE, NCBI GEO, ENA, and species-specific atlases. For cross-species training, ensure representation from target organisms (e.g., human, mouse, zebrafish). The CellFM model, for instance, was trained on approximately 100 million human cells aggregated from 19,914 samples across different organs [35].

• Quality Control and Filtering: Apply stringent quality control metrics tailored to each dataset and species. Standard parameters include:

  • Retaining cells with a minimum number of detected genes (typically 200-500 genes)
  • Filtering cells with excessively high mitochondrial gene content (indicating apoptosis)
  • Removing genes expressed in fewer than a threshold number of cells (e.g., <10 cells) [35] [34]

• Gene Name Standardization: Convert gene identifiers to standardized nomenclature according to authoritative databases (HGNC for human, MGI for mouse). This critical step enables cross-dataset and cross-species alignment by ensuring consistent gene identification [35].

• Normalization: Apply library size normalization (e.g., to 10,000 reads per cell) followed by log-transformation to stabilize variance and make expression values more comparable across cells and datasets [34] [7].

• Integration and Batch Correction: Employ computational methods (e.g., Harmony, SCVI) to mitigate technical batch effects while preserving biological variation, particularly crucial when integrating data from multiple studies, platforms, and species [4] [7].

Tokenization-Specific Methodologies

Table 2: Detailed Tokenization Methodologies by Model

Model	Tokenization Details	Expression Value Handling	Positional Encoding	Special Tokens
scBERT	Gene-based tokens from pre-defined vocabulary	Binned into discrete expression levels	Standard transformer positional encoding	Cell-type tokens, separation tokens
scGPT	Gene identity + expression value embeddings	Either binned or normalized continuous values	Learnable positional embeddings	[PERT] for perturbations, [CLS] for cell embedding
Geneformer	Genes ranked by expression level; top 2,048 genes used	Relative ranking only (absolute values discarded)	Position based on rank order	Context tokens for tissue/disease state
CellFM	Value projection with linear embedding of expression	Continuous values preserved via projection	Modified RetNet positional encoding	LoRA adapters for efficient fine-tuning
scSFUT	Fixed-window segmentation of full expression vector	Raw values processed via 1D-convolution	Bias-free attention mechanism	Reconstruction tokens for self-supervision

Value Categorization Protocol (scBERT):

• Input Representation: For each cell, create a sequence of gene tokens based on a pre-defined vocabulary of highly variable genes.
• Expression Binning: Convert normalized expression values into discrete levels (e.g., 0-5) based on predetermined thresholds.
• Sequence Construction: Assemble the gene tokens into an ordered sequence, optionally prepending special tokens for cell type or condition.
• Embedding Generation: Map each token (gene ID + expression level) to a dense vector through an embedding layer [34].

Gene Ranking Protocol (Geneformer):

• Gene Selection: Within each cell, select the top n genes (e.g., 2,048) by expression value.
• Sequence Ordering: Arrange these genes in descending order of expression to form the input sequence.
• Token Creation: Represent each gene by its identifier only, disregarding the specific expression value.
• Model Input: Feed the ordered gene sequence into the model, which learns to predict gene positions within the cellular context [35].

Scale-Free Tokenization Protocol (scSFUT):

• Full-Vector Processing: Begin with the complete gene expression vector without gene selection.
• Tokenization: Segment the high-dimensional vector into fixed-size sub-vectors using a sliding window approach.
• Feature Extraction: Apply 1D-convolution to integrate intra-token local features and expand the attention receptive field.
• Encoder Processing: Feed the resulting tokens into an unbiased transformer encoder to capture gene-gene interactions without manual feature selection [34] [36].

Diagram Title: Tokenization Pathways for Cross-Species Foundation Models

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Reagents and Computational Tools

Category	Item/Resource	Specifications & Functions	Example Use Cases
Data Resources	CZ CELLxGENE	Provides >100 million standardized single cells across tissues and species; unified data structure	Pretraining data sourcing; cross-species reference atlas [4] [10]
	PanglaoDB	Curated compendium of single-cell transcriptomics data with marker gene annotations	Marker gene validation; cell type annotation priors [4] [10]
Software Frameworks	Scanpy	Python-based toolkit for single-cell analysis; standard preprocessing pipeline	Quality control; normalization; differential expression [34] [7]
	AnnDictionary	LLM-provider-agnostic Python package built on AnnData and LangChain	Automated cell type annotation; multi-LLM benchmarking [7]
	BioLLM	Unified framework for integrating diverse single-cell foundation models	Standardized model benchmarking; streamlined model switching [37]
Model Architectures	Transformer Variants	ERetNet (CellFM), Performer (scFoundation), standard Transformer (scBERT)	Balancing computational efficiency with model performance [35] [34]
	LoRA Mechanism	Low-Rank Adaptation for parameter-efficient fine-tuning	Adapting foundation models to new species with limited data [35]

Diagram Title: Research Tool Ecosystem for Tokenization

Tokenization represents the crucial bridge between biological measurements and computational analysis in cross-species cell annotation foundation models. The choice of tokenization strategy—whether value projection, categorization, gene ranking, or scale-free approaches—fundamentally shapes what patterns a model can discover and how well it can generalize across biological contexts and species boundaries. As these models continue to evolve, we anticipate further innovation in tokenization methods that more effectively capture the hierarchical, dynamic nature of gene regulatory programs while remaining computationally tractable. The development of standardized frameworks like BioLLM for comparing these approaches will accelerate progress toward more accurate, interpretable, and biologically-grounded foundation models capable of unlocking the fundamental principles of cellular function across the tree of life.

Application Note: Cell Type Prediction with Cross-Species Foundation Models

Cross-species cell annotation foundation models represent a transformative advancement in single-cell biology. These models, pre-trained on tens of millions of cells from multiple species, learn fundamental biological principles that enable accurate cell type identification across evolutionary distances. The purpose of this application is to provide researchers with a robust, standardized methodology for annotating cell types in new, unlabeled data, including from species not present in the training corpus. This approach significantly reduces the reliance on manually curated labels and specialized, single-species reference atlases.

Key Quantitative Performance Metrics

Table 1: Performance metrics of leading foundation models for cell type prediction.

Model Name	Training Scale	Key Architectural Features	Reported Cross-Species Performance
TranscriptFormer [9]	112 million cells, 12 species	Transformer, generative	State-of-the-art in classifying cell types in out-of-distribution species (e.g., rhesus macaque, marmoset) without fine-tuning.
CellFM [35]	100 million human cells	800M parameters, ERetNet layers, LoRA	Outperforms existing models in cell annotation tasks on diverse human cell datasets.
scBERT [4]	Millions of transcriptomes	Transformer, value categorization	Effective for human cell type annotation via fine-tuning on target datasets.

Detailed Experimental Protocol

Objective: To annotate cell types in a novel single-cell RNA-seq dataset from a target species using a pre-trained cross-species foundation model.

Materials and Reagents:

• Input Data: A gene expression matrix (cells x genes) from the target dataset. Data should be in a standardized format (e.g., h5ad, Seurat object).
• Pre-trained Model: A publicly available cross-species foundation model, such as TranscriptFormer or CellFM.
• Computing Environment: Access to a GPU cluster is recommended for efficient computation.

Procedure:

• Data Preprocessing and Tokenization:
  • Perform quality control on the target dataset to filter out low-quality cells and genes.
  • Normalize the gene expression values. The specific method (e.g., log(CP10K+1)) should align with the model's training protocol.
  • Map gene identifiers in your dataset to the standard nomenclature (e.g., HGNC symbols) used by the foundation model.
  • Tokenization: Convert the normalized expression profile of each cell into a sequence of gene tokens. For many models, this involves ranking genes by their expression level within the cell to create a deterministic sequence. The expression values are then integrated, often through binning or direct projection, to create the final input tokens [4].

• Model Inference and Embedding Generation:

  • Pass the tokenized sequences for all cells through the pre-trained model.
  • Extract the contextualized cell embeddings from the model's output. These are high-dimensional vectors that represent the state and identity of each cell in a latent space.

• Cell Type Annotation:

  • Cross-Species Label Transfer: If a labeled reference atlas from a related species is available, the model can project cells from both the reference and target datasets into a shared embedding space. Cell types are then assigned based on nearest-neighbor classification within this space [9].
  • Unsupervised Clustering and Mapping: In the absence of a reference, cluster the cell embeddings using methods like Leiden or K-means. These clusters represent putative cell types. These clusters can then be manually annotated by examining the expression of known marker genes within each cluster. The model's inherent biological knowledge can aid in this interpretation.

Application Note: Disease State Identification via Critical Transition Detection

The progression of complex diseases like cancer is often nonlinear, characterized by a sudden deterioration from a pre-disease state to a disease state. Identifying this critical transition point is crucial for early intervention. This application note details a model-free method, Local Network Wasserstein Distance (LNWD), which uses single-sample analysis to detect these pre-disease states by measuring statistical perturbations in molecular networks [38].

Key Quantitative Performance Metrics

Table 2: Application of LNWD in identifying critical states across complex diseases.

Disease / Condition	Dataset Source	Key Finding	Validation Method
Renal Cancers (KIRP, KIRC)	TCGA	Successful identification of the critical pre-disease state before cancer progression.	Survival analysis and molecular network dynamics change analysis [38].
Lung Adenocarcinoma (LUAD)	TCGA	Detection of critical transition signals from network perturbation.	Consistent with clinical staging and outcome data [38].
Acute Lung Injury (Mice)	GEO: GSE2565	Provided early warning signals for disease deterioration.	Matched with experimental time-course data [38].
Rheumatoid Arthritis [39]	Leiden & TACERA Cohorts	Identified 4 distinct disease trajectories (A: high ESR; D: many inflamed joints).	Replicated in an independent cohort; linked to patient-reported outcomes.

Detailed Experimental Protocol

Objective: To compute LNWD scores for a series of samples to identify the critical pre-disease state during disease progression.

Materials and Reagents:

• Gene Expression Data: A cohort of samples spanning a time course or disease progression stages (e.g., normal -> stage I -> stage II -> stage III).
• Protein-Protein Interaction (PPI) Network: A high-confidence network (e.g., from STRING-db) for the relevant species.
• Software: R or Python packages for differential expression analysis (e.g., edgeR, limma, DESeq2) and computational geometry for calculating Wasserstein Distance.

Procedure:

• Data Preparation and Differential Gene Analysis:
  • For each stage in the disease progression, identify differentially expressed genes (DEGs) compared to the normal stage. Use an intersection of results from multiple methods (e.g., edgeR, limma, DESeq2) with stringent cutoffs (e.g., \|logFC\| > 2, p-value < 0.05) [38].
  • Take the union of all DEGs across all stages to form a master list of genes for network construction.

• Local Network Construction:

  • Map the master list of DEGs onto the PPI network. For each gene, extract its direct interaction partners to form a local network.
  • For each sample, the expression profile of genes within these local networks is used for subsequent analysis.

• Calculation of Local Network Wasserstein Distance (LNWD):

  • Define a reference group: a set of samples from the normal/healthy state.
  • For a single test sample (e.g., a diseased sample), form a mixed group by adding it to the reference group.
  • For each local network, calculate the Wasserstein Distance between the expression distributions of the reference group and the mixed group. The WD quantifies the minimal "cost" to transform one distribution into the other, capturing subtle statistical perturbations [38].

• Identification of the Critical State:

  • For each test sample, select the top 10% of local networks with the highest LNWD scores and calculate their average to obtain a global LNWD score.
  • Plot the global LNWD scores for all samples across the disease progression timeline. A sharp peak in the LNWD score indicates the critical transition point from the pre-disease to the disease state.

Workflow Diagram: Identifying Critical Disease States with LNWD

Application Note: Mapping Gene-Gene Interactions

Understanding the functional relationships between genes is fundamental to biology. Gene-gene interaction mapping identifies pairs of genes that, when mutated together, result in an unexpected phenotype (e.g., synthetic lethality), revealing functional redundancy, compensation, and pathway relationships. This application covers both experimental and AI-driven computational methods for large-scale gene interaction mapping.

Key Quantitative Performance Metrics

Table 3: Approaches for large-scale genetic interaction mapping.

Method / Technology	Scale / Model	Key Outcome	Application Context
Dual Tn-seq [40]	Surveyed ~1 million gene pairs in S. pneumoniae; over 1 billion double mutants created.	Identified 200+ previously unknown genetic interactions; discovered a new enzyme family.	High-throughput screening in bacteria for functional genomics and antibiotic target discovery.
CRISPR-based qGI Profiling [41]	~4 million gene pairs tested in human HAP1 cells; ~90,000 genetic interactions mapped.	Generated a hierarchical model of human cell function; recapitulated and expanded on DepMap data.	Unbiased functional genomics in a human cell line model to understand genetic architecture.
AI Model Prediction (TranscriptFormer) [9]	Generative model trained on 112M cells.	Predicts gene-gene co-expression relationships in specific cell types and conditions via prompting.	In-silico hypothesis generation for gene function and interaction before wet-lab validation.

Detailed Experimental Protocol: Dual Tn-seq in Bacteria

Objective: To systematically identify genetic interactions (e.g., synthetic lethal pairs) in a bacterial pathogen on a genome-wide scale.

Materials and Reagents:

• Bacterial Strain: The target strain, e.g., Streptococcus pneumoniae.
• Transposon Library: A barcoded transposon library, such as the one used in the RB-TnSeq technique.
• Molecular Biology Reagents: Enzymes for molecular matchmaking to bring two barcodes together for sequencing.
• Growth Medium: For culturing the mutant library under desired conditions.
• Next-Generation Sequencing Platform.

Procedure:

• Library Generation:
  • Generate a complex pool of random double mutants by inserting two distinct barcoded transposons into a single bacterial cell.
  • Use a "molecular matchmaker" enzyme to link the two barcodes from a single cell, enabling them to be sequenced as a pair [40].

• Competitive Growth Assay:

  • Grow the entire library of double mutants in a condition of interest (e.g., rich medium or stress condition) for multiple generations.
  • Harvest genomic DNA from the population at the beginning (T0) and end (Tf) of the experiment.

• Sequencing and Data Analysis:

  • Amplify and sequence the barcode pairs from both time points using high-throughput sequencing.
  • Quantify the abundance of each barcode pair (representing a specific double mutant) at T0 and Tf.
  • Calculate a fitness defect for each double mutant by comparing the change in barcode pair abundance over time.
  • Identify genetic interactions by comparing the observed double mutant fitness to the expected fitness based on the single mutant effects. A negative interaction (synthetic sickness/lethality) occurs when the double mutant grows significantly worse than expected.

The Scientist's Toolkit: Key Research Reagents & Models

Table 4: Essential tools for advanced single-cell and gene interaction analysis.

Item / Resource	Type	Primary Function
CZ CELLxGENE [9] [4]	Data Platform	Provides unified access to millions of curated single-cell datasets for model training and analysis.
TranscriptFormer [9]	AI Foundation Model	A generative model for cross-species cell type prediction, disease state identification, and gene interaction prediction via prompting.
Barcoded Transposon Library (e.g., for Dual Tn-seq) [40]	Wet-lab Reagent	Enables high-throughput generation and tracking of mutants in pooled screens.
CRISPR gRNA Library (e.g., TKOv3) [41]	Wet-lab Reagent	Allows for systematic knockout of genes in human cells for essentiality and genetic interaction screens.
Local Network Wasserstein Distance (LNWD) [38]	Computational Algorithm	Detects critical transition states in complex diseases from single-sample transcriptomic data.
PPI Network (e.g., from STRING-db) [38]	Bioinformatics Resource	Provides prior knowledge of protein interactions for constructing local networks in algorithms like LNWD.

Navigating Computational Challenges: Optimization Strategies for Robust Cross-Species Integration

Cross-species integration of single-cell RNA-sequencing (scRNA-seq) data enables researchers to explore evolutionary relationships and identify conserved and divergent cell types across species. The fundamental challenge lies in distinguishing true biological variation from technical artifacts and species-specific effects, often termed "species effect," where cells from the same species exhibit higher transcriptional similarity to each other than to their cross-species counterparts [42]. Robust benchmarking is therefore essential to ensure integration results reflect biology rather than computational artifacts. This protocol outlines comprehensive metrics and methodologies for evaluating integration performance, with a focus on balancing species-mixing with biological conservation.

Core Benchmarking Metrics and Quantitative Comparison

Performance evaluation in cross-species integration spans two primary objectives: achieving adequate mixing of homologous cell types from different species, and preserving meaningful biological heterogeneity within and across cell types. The table below summarizes the key metrics employed for these purposes, their mathematical basis, and ideal values.

Table 1: Comprehensive Metrics for Benchmarking Cross-Species Integration Performance

Metric Category	Metric Name	Description	Interpretation & Ideal Value
Species Mixing	Average Silhouette Width (ASW) batch	Measures how closely cells from the same species cluster together versus how separated they are from other species [42].	Value Range: -1 to 1Ideal: Closer to 0, indicating no batch (species) structure.
	Normalized Mutual Information (NMI)	Quantifies the similarity between the species label distribution and the clustering result [42].	Value Range: 0 to 1Ideal: Lower values indicate better mixing (less association).
	Alignment Score (for SAMap)	Quantifies the percentage of cross-species neighbors for each cell [42].	Value Range: 0 to 1Ideal: Higher values indicate better alignment of homologous types.
Biology Conservation	Accuracy Loss of Cell type Self-projection (ALCS)	A novel metric quantifying the loss of cell type distinguishability post-integration using a self-projection concept [42].	Value Range: 0 to 1Ideal: Lower values indicate better preservation of biological heterogeneity.
	Average Silhouette Width (ASW) cell type	Measures how closely cells of the same cell type cluster together after integration [42].	Value Range: -1 to 1Ideal: Higher values indicate cell types are well-separated.
	Normalized Mutual Information (NMI) cell type	Quantifies the similarity between the cell type label distribution and the clustering result [42].	Value Range: 0 to 1Ideal: Higher values indicate better preservation of cell type identity.
	Cell-type Label Transfer Accuracy (ARI)	Assesses annotation transfer using Adjusted Rand Index between original and transferred labels [42].	Value Range: -1 to 1Ideal: Values closer to 1 indicate highly accurate cross-species annotation.
Overall Score	Integrated Score	A weighted average of the scaled species-mixing and biology conservation scores [42].	Ideal Weighting: 40% species-mixing, 60% biology conservation [42].

Experimental Protocols for Benchmarking Analysis

Protocol 1: Pre-processing and Gene Homology Mapping

• Quality Control (QC) and Curation: Perform input-specific QC on raw count matrices. Rigorously curate cell ontology annotations to establish a "gold standard" for homologous cell types prior to integration [42].
• Gene Homology Mapping: Translate orthologous genes between species using ENSEMBL's multiple species comparison tool or similar resources [42].
  • Method A (One-to-One): Use only one-to-one orthologs.
  • Method B (Include Paralogs): Include one-to-many or many-to-many orthologs, selecting those with high average expression or strong homology confidence. This is particularly beneficial for evolutionarily distant species [42].
  • Method C (Unshared Features): For algorithms like LIGER UINMF, add genes without annotated homology on top of the mapped genes [42].
• Data Concatenation: Concatenate the raw count matrices from different species using the selected gene mapping to create a unified feature space.

Protocol 2: Executing and Assessing Integration with the BENGAL Pipeline

• Integration Execution: Feed the concatenated matrix into various integration algorithms. The BENGAL pipeline benchmarks multiple strategies, including scANVI, scVI, SeuratV4 (CCA/RPCA), Harmony, and Scanorama [42]. For specialized cases involving whole-body atlases or challenging gene homology, consider the standalone SAMap workflow [42].
• Metric Calculation: Compute the metrics listed in Table 1 for the integrated output.
• Contextual Scaling: Min-max scale the metric scores (except for SAMap-specific metrics) within the task using the scores from the unintegrated, homology-concatenated data as a baseline. This equalizes the discriminative power of different metrics [42].
• Score Aggregation:
  • Calculate the Species Mixing Score as the average of applicable scaled batch correction metrics.
  • Calculate the Biology Conservation Score as the average of applicable scaled biology conservation metrics.
  • Calculate the final Integrated Score as a weighted average: (0.4 * Species Mixing Score) + (0.6 * Biology Conservation Score) [42].
• Visual Inspection: Generate Uniform Manifold Approximation and Projection (UMAP) plots to visually confirm metric findings, checking for proper mixing of homologous types and distinct separation of non-homologous types [42].

Protocol 3: Annotation Transfer Validation

• Classifier Training: On the integrated embedding, train a multinomial logistic classifier (e.g., from the SCCAF framework) using the cell-type labels from one species (the source) [42].
• Prediction: Use the trained classifier to predict cell-type labels for the cells from the other species (the target).
• Validation: Calculate the Adjusted Rand Index (ARI) between the predicted labels and the original, curated labels for the target species. A high ARI indicates the integration has successfully captured biologically relevant features that are conserved across species [42].

Workflow Visualization of the Benchmarking Process

The following diagram illustrates the logical flow and key decision points in the cross-species integration benchmarking pipeline.

Table 2: Key Computational Tools and Resources for Cross-Species Integration

Tool/Resource Name	Type	Primary Function in Context
BENGAL Pipeline [42]	Computational Pipeline	A freely available benchmark and assessment pipeline for cross-species integration, facilitating the evaluation of multiple strategies.
scANVI [42] [43]	Integration Algorithm	A semi-supervised deep learning model that leverages cell-type annotations to achieve a balance between species-mixing and biology conservation.
scVI [42] [43]	Integration Algorithm	A probabilistic deep learning framework that effectively models technical and biological noise for data integration.
Seurat V4 [42]	Integration Algorithm	Uses CCA or RPCA to identify anchors between datasets, demonstrating strong performance in cross-species tasks.
SAMap [42]	Integration Algorithm	Specialized for whole-body atlas integration between species with challenging gene homology, using reciprocal BLAST for gene-gene mapping.
ENSEMBL Compara [42]	Bioinformatics Database	Provides pre-computed gene homology mappings (e.g., one-to-one, one-to-many orthologs) essential for creating a shared feature space.
CZ CELLxGENE [10] [9]	Data Platform	Provides unified access to millions of curated single-cell datasets across multiple species, serving as a key data source for pretraining and testing.
TranscriptFormer [9]	Foundation Model	A generative, multi-species model trained on 112M cells from 12 species, enabling cross-species cell type prediction and analysis without fine-tuning.

The identification of homologous genes across species is a cornerstone of comparative genomics, crucial for inferring gene function, understanding evolutionary history, and annotating cells. Traditional methods that rely on one-to-one ortholog mapping are increasingly inadequate for capturing the complex reality of genomic evolution, which is replete with gene duplications, paralogs, and co-orthologs. This application note details advanced computational protocols and metrics designed to address these complexities. We frame these methodologies within the emerging paradigm of cross-species cell annotation foundation models, which leverage vast datasets and self-supervised learning to create unified representations of biological data. The provided protocols for alignment-free sequence comparison, iterative graph matching, and homology-independent integration, alongside benchmarking strategies, equip researchers with the tools to achieve more accurate and biologically meaningful cross-species comparisons.

The premise of one-to-one ortholog mapping, where a single gene in one species corresponds to a single gene in another, often fails to reflect biological complexity. Evolutionary events like gene duplication and whole-genome duplication lead to the proliferation of co-orthologs—paralogous genes within a genome that are collectively orthologous to one or more genes in another species [44] [45]. Relying solely on one-to-one mappings risks misannotating these genes and obscuring true evolutionary relationships.

This challenge is magnified in the context of cross-species cell annotation foundation models (scFMs). These models, such as TranscriptFormer and scGPT, are trained on millions of single-cell transcriptomes from multiple species to learn a unified representation of cellular biology [10] [9]. Their goal is to enable tasks like cell type annotation, disease state prediction, and gene-gene interaction analysis across vast evolutionary distances. The performance of these models is fundamentally dependent on the quality and biological accuracy of the gene homology mappings used to align the feature spaces of different species. Annotation heterogeneity—the use of different gene annotation methods across the species in a study—can artificially inflate the number of lineage-specific genes by up to 15-fold, presenting a major source of artifact in comparative genomics [46]. Moving beyond simple one-to-one mapping is therefore not merely an academic exercise but a practical necessity for robust biological discovery.

Advanced Computational Strategies for Complex Homology Detection

Alignment-Free Sequence Comparison withafree

Principle: Alignment-free methods operate on the hypothesis that similar sequences share a significant number of k-mers (contiguous substrings of length k). These methods circumvent the computational intensity of traditional alignment algorithms, enabling rapid all-against-all comparison of large sequence sets, which is often the first bottleneck in ortholog assignment pipelines [44].

Protocol: The afree Algorithm Workflow

The following workflow visualizes the core process of the afree algorithm for efficient large-scale sequence comparison:

Detailed Methodology:

• Input Preparation: Begin with two sets of protein sequences from genomes G1 and G2. Concatenate them into a single set, G3, for unified processing [44].
• k-mer Generation: For every sequence in G3, slide a window of length k (e.g., k=6 for optimal balance of specificity and speed) along the sequence. For each sequence, record only unique k-mers to enhance efficiency. Each unique k-mer is stored in a list L as a tuple containing:
  • The k-mer string itself.
  • The index of the protein sequence it belongs to.
  • The offset position of the k-mer within the sequence [44].
• Data Encoding for Scalability: To enable processing of very large datasets, pack each tuple from list L into a 64-bit machine word. The encoding scheme is as follows:
  • Encode each amino acid in the k-mer using 5 bits.
  • Use the remaining bits to store the protein sequence index and the k-mer offset. This compressed representation allows for rapid sorting and joining in subsequent steps [44].
• Sort-Join Operation: Sort the entire list L based on the encoded k-mer value. This clustering allows the algorithm to efficiently identify all sequence pairs that share common k-mers in a single pass, a key factor in its scalability [44].
• Similarity Calculation: For each sequence pair identified as sharing k-mers, calculate a statistical similarity measure based on the shared k-mer count. This measure quantifies the sequence similarity without generating a base-by-base alignment [44].

Identifying Co-orthologs via Iterative Graph Matching

Principle: After initial similarity detection, an iterative graph matching strategy can be employed to resolve complex many-to-many orthologous relationships. This method operates on a bipartite graph where nodes represent genes from two genomes, and weighted edges represent their sequence similarity scores. The goal is to find a matching that maximizes the sum of similarity scores while allowing for the identification of co-orthologs [44].

Protocol: Iterative Graph Matching for Co-orthology

• Graph Construction: Construct a bipartite graph where one set of nodes represents genes from genome A and the other set represents genes from genome B. Connect genes with edges weighted by their sequence similarity scores (e.g., derived from the afree algorithm) [44].
• Iterative Matching: In each iteration, identify the set of gene matches (orthologs and co-orthologs) that maximizes the total sum of similarity scores across the graph.
• Match Resolution: After the best matches are identified in an iteration, their corresponding edges are removed from the graph. The process repeats on the remaining graph, identifying the next best set of matches. This iterative process continues until no significant matches remain, resulting in a set of orthologs and co-orthologs [44].
• Output: The final output is a set of gene pairs and groups, classified as one-to-one, one-to-many, or many-to-many orthologs, providing a more nuanced view of gene relationships.

The Mestortho Algorithm: A Minimum Evolution Approach

Principle: The Mestortho algorithm is based on the phylogenetic minimum evolution (ME) criterion. It postulates that a set of sequences consisting purely of orthologs will have a smaller sum of branch lengths (the Minimum Evolution Score, or MES) in a neighbor-joining tree than a set that includes one or more paralogous relationships [45].

Protocol: Orthology Detection with Mestortho

• Input and Grouping: Provide a multiple sequence alignment where each sequence's identifier includes species information. The program automatically classifies sequences into two groups:
  • Group 1: Sequences with one occurrence per species.
  • Group 2: Sequences with more than one occurrence per species (potential paralogs) [45].
• Combinatorial Set Generation: For Group 2, create exhaustive combinatorial sets, each containing exactly one sequence per species. If a reference sequence is specified, only datasets containing that reference are selected [45].
• Merge and Calculate MES: Merge the sequences from Group 1 with each combinatorial dataset from Group 2. For each resulting merged dataset:
  • Reconstruct a neighbor-joining (NJ) tree.
  • Calculate the MES (the sum of all branch lengths) of the NJ tree [45].
• Ortholog Selection: The merged dataset with the smallest MES is selected as the most reliable set of orthologs, as it represents the evolutionary least costly scenario [45].
• Co-orthology Examination: The algorithm then examines the original NJ tree from the full alignment to identify monophyletic groups limited to a single species. If such a group contains an ortholog identified in the previous step, and the pairwise distances within the group are less than distances to sequences from all other species, the group is designated as a set of co-orthologs [45].

Benchmarking Integration Strategies for Cross-Species Analysis

Integrating data across species requires mapping genes via homology, and the chosen strategy significantly impacts the results. A comprehensive benchmark of 28 strategies (4 homology mapping methods combined with 9 integration algorithms) provides critical guidance [47].

Table 1: Benchmarking Metrics for Cross-Species Integration Strategies

Metric Category	Metric Name	Description	What It Quantifies
Species Mixing	Average Silhouette Width (ASW) Batch	Measures how close cells are to cells of the same species versus others in the embedding.	Better mixing of homologous cell types across species.
	Graph Integration Local Inverse Simpson's Index (GILISI)	Assesses the local diversity of species labels among a cell's nearest neighbors.	Whether cell neighborhoods contain a mix of species.
Biology Conservation	Average Silhouette Width (ASW) Cell Type	Measures how close cells are to cells of the same type versus other types.	Preservation of distinct cell type clusters.
	Normalized Mutual Information (NMI)	Quantifies the similarity between the cell type clustering before and after integration.	Conservation of the original biological grouping.
	ALCS (New Metric)	Accuracy Loss of Cell type Self-projection; quantifies the blending of distinct cell types post-integration.	Protection against overcorrection, which obscures species-specific cell types [47].

Table 2: Key Findings from Benchmarking 28 Integration Strategies [47]

Finding	Implication for Experimental Design
The choice of integration algorithm (e.g., scANVI, scVI, SeuratV4) has a greater impact on performance than the homology mapping method.	Prioritize selection of a robust integration algorithm.
For evolutionarily distant species, including in-paralogs (one-to-many orthologs) in the gene mapping is beneficial.	Use a more inclusive homology mapping strategy for distant species comparisons.
SAMap, which uses de-novo BLAST instead of pre-defined orthology tables, outperforms other methods for whole-body atlas integration between species with challenging gene homology annotation.	Use SAMap for complex integrations, especially when standard orthology tables are incomplete or unreliable [47].
The new ALCS metric is critical for identifying overcorrection, where algorithms force integration so strongly that they blend biologically distinct cell types.	Always use ALCS alongside other metrics to ensure biological heterogeneity is preserved [47].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Software and Database Tools for Complex Homology Analysis

Tool Name	Type	Primary Function	Application Context
afree & EGM2 [44]	Algorithm/Pipeline	Alignment-free all-against-all sequence comparison & iterative ortholog detection.	Fast homology search foundation for large datasets.
Mestortho [45]	Python Program	Detects orthologs from a multiple sequence alignment using the minimum evolution criterion.	Phylogeny-based orthology inference for curated gene families.
BENGAL Pipeline [47]	Benchmarking Pipeline	Systematically tests and evaluates cross-species integration strategies.	Selecting the optimal data integration method for a given project.
SAMap [47]	Integration Algorithm	Uses iterative BLAST and cell-cell mapping to integrate data, ideal for challenging homology.	Aligning whole-body atlases across evolutionarily distant species.
SegmentNT [48]	DNA Foundation Model	Fine-tunes pretrained models to annotate genic and regulatory elements at single-nucleotide resolution.	Genome annotation without relying on pre-defined gene models.
TranscriptFormer [9]	Single-Cell Foundation Model (scFM)	A generative model trained on 112M cells from 12 species for cross-species prediction.	Cell type annotation, disease state prediction, and gene-gene interaction analysis across species.
CZ CELLxGENE [10] [9]	Data Platform	Provides unified access to millions of annotated single-cell datasets.	Source of curated training data for scFMs and cross-species analysis.

Navigating the complexity of gene homology requires a multifaceted toolkit that moves decisively beyond the simplicity of one-to-one ortholog mapping. The protocols and strategies detailed herein—from scalable alignment-free sequencing and phylogeny-based orthology assignment to rigorously benchmarked integration methods—provide a roadmap for robust cross-species genomic and single-cell analysis. As the field advances towards powerful foundation models capable of synthesizing biological information across millions of cells and billions of years of evolution, the accurate resolution of gene homology remains a critical foundation. By adopting these advanced methodologies, researchers can mitigate artifacts, uncover true evolutionary relationships, and fully leverage the potential of cross-species foundation models to illuminate cellular function and disease.

Mitigating Batch Effects and Sequencing Depth Inconsistencies

In cross-species cell annotation research, single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics data are plagued by technical variations known as batch effects. These non-biological differences, arising from factors like different sequencing platforms, laboratory conditions, or sample preparation protocols, can obscure genuine biological signals and compromise the integrity of foundational models. Similarly, inconsistencies in sequencing depth—the number of reads per cell—can create artificial variation in gene detection, misrepresenting true cellular states. For foundation models aiming to create a unified representation of cells across diverse species and tissues, the robust mitigation of these technical confounders is not merely a preprocessing step but a foundational prerequisite. This document outlines standardized protocols and application notes for identifying and correcting these issues, ensuring reliable and reproducible cell annotation.

The performance of various batch-effect correction and analysis tools can be quantitatively assessed based on their underlying algorithms, data handling capabilities, and performance metrics. The table below summarizes these aspects for several methods discussed in recent literature.

Table 1: Comparison of Batch Effect Mitigation and Analysis Frameworks

Method / Framework Name	Underlying Architecture / Algorithm	Data Types Handled	Key Metrics and Performance	Primary Application in Annotation	Key Advantages
Harmony [49] [50]	Iterative clustering and linear mixture modeling	scRNA-seq, CITE-seq	Effectively integrates datasets from 38 tissues and 700 individuals; improves cross-dataset gene expression program (GEP) reproducibility [49].	Data integration for unified cell state definition.	Corrects gene-level data while preserving non-negative values for component-based models [49].
T-CellAnnoTator (TCAT) / starCAT [49]	Consensus Non-negative Matrix Factorization (cNMF) with batch correction	scRNA-seq, CITE-seq	Identified 46 reproducible GEPs; accurately infers GEP usage in query datasets (Pearson R > 0.7) [49].	Quantifying predefined GEP activities in new cells/datasets.	Provides a consistent cell state representation across datasets; robust for rare GEPs and fast for query data [49].
Nicheformer [3]	Transformer-based Foundation Model	Dissociated scRNA-seq, Spatial Transcriptomics (MERFISH, Xenium, etc.)	Outperforms models trained only on dissociated data (e.g., Geneformer, scGPT) in spatial tasks like niche and composition prediction [3].	Learning spatially aware cellular representations and transferring spatial context to dissociated data.	Jointly trained on dissociated and spatial data; uses rank-based gene encoding for robustness to technical biases [3].
TranscriptFormer [9]	Transformer-based Foundation Model	Cross-species scRNA-seq	Achieves state-of-the-art performance in cross-species cell type classification and disease state prediction, even for "out-of-distribution" species [9].	Generalizing biological patterns across vast evolutionary distances.	Trained on 112 million cells from 12 species; enables prediction without species-specific labeled data [9].

Experimental Protocols

Protocol: Reference-Based Annotation with starCAT for Cross-Species Data

This protocol details the use of the starCAT framework to annotate cell states in a new, cross-species query dataset using a pre-defined catalog of Gene Expression Programs (GEPs).

1. Principle The starCAT pipeline avoids de novo analysis of query data. Instead, it leverages a fixed, multi-dataset catalog of GEPs—learned from a large, batch-corrected reference—to quantify the activity of these conserved programs in new query datasets. This ensures consistent annotation and enables the detection of rare cell states that might be missed in smaller query datasets [49].

2. Reagents and Materials Table 2: Essential Research Reagent Solutions

Item	Function / Description
Reference GEP Catalog	A pre-computed set of consensus GEPs (e.g., the 46 T cell cGEPs) derived from multiple batch-corrected datasets. Serves as the fixed coordinate system for annotation [49].
Processed Query Dataset	A quality-controlled (QC'd) gene expression matrix (cells x genes) from a new experiment, which may be from a different species or technology.
Batch-Corrected Reference Data	Large, integrated scRNA-seq dataset(s) (e.g., the 1.7 million T cell reference) used to derive the robust GEP catalog. Corrected with tools like Harmony [49].
CITE-seq Antibody-Derived Tags (Optional)	Surface protein expression data from CITE-seq, integrated into the GEP spectra to enhance biological interpretability and annotation confidence [49].

3. Procedure 1. Reference GEP Catalog Construction: a. Data Collection & Harmonization: Collate multiple large-scale scRNA-seq datasets encompassing the cell types of interest across desired species and conditions. b. Batch Effect Correction: Apply a batch correction method like Harmony to the raw, non-negative count data to generate a harmonized gene-level expression matrix [49]. c. Consensus NMF (cNMF): Run cNMF on the batch-corrected data to learn robust GEP spectra (gene weights) and their usage in each reference cell. This involves multiple runs of NMF followed by consensus clustering [49]. d. Curation & Annotation: Manually curate the resulting GEPs, removing technical artifacts and annotating them based on top-weighted genes and gene-set enrichment analysis. This creates the final reference catalog.

Protocol: Integrated Analysis of Dissociated and Spatial Data with Nicheformer

This protocol describes using the Nicheformer foundation model to enrich dissociated scRNA-seq data with spatial context and mitigate technology-specific biases.

1. Principle Nicheformer is a transformer model pretrained on a massive, curated corpus (SpatialCorpus-110M) containing both dissociated and spatially resolved single-cell data. It learns a joint representation that captures spatial context, allowing it to perform spatially aware tasks and transfer spatial information from targeted spatial transcriptomics to dissociated scRNA-seq data, which typically has broader gene coverage [3].

2. Procedure 1. Model Input Preparation (Tokenization): a. For a given cell (from either dissociated or spatial data), convert its gene expression vector into a sequence of gene tokens. b. Rank-based Encoding: Order the gene tokens by their expression level relative to the technology-specific non-zero mean, not by absolute value. This strategy enhances robustness to technology-driven batch effects [3]. c. Contextual Tokens: Prepend the sequence with special tokens indicating the species (e.g., human, mouse), data modality (dissociated vs. spatial), and specific technology (e.g., MERFISH, 10X) [3].

Visualization of Workflows

starCAT Workflow for Cross-Species Annotation

Nicheformer Data Integration and Spatial Prediction

Validation Frameworks and Performance Benchmarks: Assessing Model Efficacy Across Biological Contexts

Cross-species analysis of single-cell RNA-sequencing (scRNA-seq) data presents a powerful approach for understanding evolutionary biology and cellular function. A significant challenge in this field is the robust integration of data across different species to identify homologous cell types accurately. This Application Note details the implementation and findings of a comprehensive benchmarking study evaluating 28 distinct integration strategies for cross-species single-cell data. The content is framed within the broader research context of developing cross-species cell annotation foundation models, which require high-quality, integrated data for training and validation to accurately decipher the 'language' of cells across different organisms [10].

The BENGAL Benchmarking Pipeline

The BENchmarking strateGies for cross-species integrAtion of singLe-cell RNA sequencing data (BENGAL) pipeline was developed to systematically assess cross-species integration strategies [47]. The pipeline evaluates strategies based on their ability to mix cells from known homologous types across species (species-mixing) while preserving biological heterogeneity present within each species (biology conservation). Prior to analysis, user-performed quality control and curation of cell ontology annotations are essential.

Key Experimental Tasks

The benchmarking was conducted across 16 diverse biological tasks to ensure broad applicability [47]. These tasks were designed to evaluate performance in different scenarios, as summarized in Table 1.

Table 1: Summary of Benchmarking Biological Tasks

Task Category	Biological Context	Species Involved	Key Evaluation Focus
Adult Tissue Analysis	Pancreas, Hippocampus, Heart	Multiple vertebrate species	Integration of homologous cell types in specialized tissues
Whole-Body Embryonic Development	Embryonic atlases	Species with challenging gene homology	Handling of complex, whole-organism data
Multi-Species Integration	Heart data	5 species	Upper limit of species numbers in a single integration
Pairwise Divergence Analysis	Various tissues	10 pairwise tasks	Impact of evolutionary divergence time on integration

Quantitative Performance Results

The benchmarking evaluated 28 strategies, resulting from combinations of 4 gene homology mapping methods and 10 integration algorithms, plus the standalone method SAMap [47]. Performance was quantitatively assessed using established metrics for species mixing and biology conservation, combined into an integrated score. Table 2 summarizes the performance of the top-performing strategies.

Table 2: Top-Performing Integration Strategies and Key Findings

Integration Strategy	Overall Performance	Strengths and Optimal Use Cases
scANVI	High integrated score	Achieves optimal balance between species-mixing and biology conservation [47].
scVI	High integrated score	Robust performance across multiple tissue types and species pairs [47].
SeuratV4 (CCA/RPCA)	High integrated score	Effective for standard cross-species comparisons with well-annotated genomes [47].
SAMap	Specialized outperformer	Superior for whole-body atlas integration between species with challenging gene homology annotation [47].
Strategies with In-Paralogs	Beneficial for distant species	Including in-paralogs in the gene mapping step improves evolutionarily distant species integration [47].

Assessment Metrics and Outcomes

Integration outputs were assessed from three primary perspectives [47]:

• Species Mixing: Evaluated using batch correction metrics to determine how well homologous cell types from different species cluster together.
• Biology Conservation: Assessed using biology conservation metrics to ensure biological heterogeneity within species is not lost during integration.
• Annotation Transfer: Measured by training a classifier on one species to predict cell types in another, with performance quantified by the Adjusted Rand Index (ARI).

A key metric developed for this benchmark was the Accuracy Loss of Cell type Self-projection (ALCS), which specifically quantifies the unwanted blending of distinct cell types within a species after integration, indicating overcorrection [47].

Detailed Experimental Protocols

Protocol 1: Gene Homology Mapping and Data Concatenation

Objective: To map orthologous genes between species and create a concatenated raw count matrix for integration. Reagents & Materials:

• ENSEMBL multiple species comparison tool [47]
• Raw single-cell RNA-seq count matrices from the species to be integrated

Procedure:

• Gene Homology Mapping: Utilize the ENSEMBL comparative genomics tool to identify orthologous genes between the target species [47].
• Mapping Method Selection: Choose one of the following gene mapping approaches:
  • One-to-One Orthologs: Use only genes with a single, direct ortholog in each species.
  • Include Complex Orthologs: Incorporate one-to-many or many-to-many orthologs, selecting those with either high average expression levels or strong homology confidence scores [47].
  • Unshared Features (LIGER UINMF only): For the LIGER UINMF algorithm, add genes without annotated homology on top of the mapped genes [47].
• Matrix Concatenation: Create a unified raw count matrix by concatenating the datasets from different species, including only the mapped orthologous genes.

Protocol 2: Data Integration Execution

Objective: To apply integration algorithms to the concatenated data matrix to generate a joint embedding. Reagents & Materials:

• Concatenated raw count matrix from Protocol 1
• Computational environment with installed integration algorithms (e.g., scANVI, scVI, SeuratV4, SAMap)

Procedure:

• Algorithm Selection: Select an integration algorithm. Top performers identified in the benchmark include scANVI, scVI, and SeuratV4 (using either CCA or RPCA) [47].
• Input Data: Feed the concatenated raw count matrix into the chosen algorithm.
• SAMap Workflow: If using SAMap, follow its standalone workflow, which involves de-novo reciprocal BLAST analysis to construct a gene-gene homology graph and does not use the pre-concatenated matrix [47].
• Output Generation: Execute the algorithm to produce a low-dimensional integrated embedding of all cells from different species.

Protocol 3: Integration Output Assessment

Objective: To quantitatively and qualitatively evaluate the quality of the integrated embedding. Reagents & Materials:

• Integrated embedding from Protocol 2
• Ground truth cell type labels for each species
• Benchmarking metrics pipeline (e.g., the BENGAL assessment module)

Procedure:

• Species Mixing Calculation: Compute batch correction metrics (e.g., using the BENGAL pipeline) to evaluate the degree of mixing between known homologous cell types across species [47].
• Biology Conservation Calculation: Compute biology conservation metrics to assess the preservation of biological variance and cell type distinguishability within each species. Include the calculation of the ALCS metric to detect overcorrection [47].
• Annotation Transfer Test:
  • Train a multinomial logistic classifier (e.g., from the SCCAF framework) on the integrated embedding using the cell type labels from one species as the training set [47].
  • Use the trained classifier to predict cell types for the other species in the integration.
  • Calculate the Adjusted Rand Index (ARI) between the predicted labels and the original ground truth labels to quantify transfer accuracy [47].
• Visual Inspection: Generate a UMAP visualization of the integrated embedding to visually inspect species mixing and cluster formation.

Workflow and Strategy Selection Diagram

Strategy Selection Logic

Research Reagent Solutions

Table 3: Essential Research Reagents and Computational Tools

Item Name	Function / Purpose	Application Context
ENSEMBL Comparative Tool	Provides gene homology mapping (orthology predictions) between species [47].	Essential pre-processing step for identifying comparable genes across species before data integration.
BENGAL Pipeline	A freely available cross-species integration and assessment pipeline [47].	Core framework for running and evaluating the 28 integration strategies in a standardized manner.
SCCAF (Single Cell Clustering Assessment Framework)	Machine learning-based framework for self-projection and cluster assessment [47].	Used to implement the ALCS metric and for annotation transfer tests.
Reciprocal BLAST	Tool for de-novo gene-gene homology analysis [47].	Critical component of the SAMap workflow, especially for species with poor existing gene annotations.
scANVI / scVI Algorithms	Probabilistic deep learning models for single-cell data integration [47].	Top-performing algorithms for general-purpose cross-species integration tasks.
SeuratV4 with CCA/RPCA	Integration methods using canonical correlation analysis or reciprocal PCA [47].	Robust, well-established methods for cross-species integration, particularly effective with clear one-to-one homologous cell types.

Within the framework of cross-species cell annotation foundation model research, biological validation is a critical step for assessing model performance and translational utility. This application note details experimental protocols and validation strategies for two complex biological systems: brain cell types and spermatogenesis. The case study on spermatogenesis demonstrates how foundation models like TranscriptFormer enable the transfer of cell-type annotations across evolutionarily distant species, a capability with profound implications for evolutionary biology and translational research [9]. Advanced single-cell RNA sequencing (scRNA-seq) technologies now provide the resolution necessary to deconstruct the dynamic process of spermatogenesis across mammalian species, revealing both conserved and divergent molecular programs [51] [52].

Case Study: Cross-Species Validation of Spermatogenesis

Biological Context and Rationale

Spermatogenesis is a highly conserved yet rapidly evolving process, making it an ideal system for validating cross-species foundation models. The process involves precisely orchestrated transitions from spermatogonia (mitotic stem cells) through spermatocytes (undergoing meiosis) to spermatids (post-meiotic haploid cells) [51] [52]. Recent evolutionary analyses of single-nucleus transcriptome data from testes of 11 species covering all main mammalian lineages (eutherians, marsupials, and monotremes) and birds have revealed that the rapid evolution of the testis is driven by accelerated evolutionary rates in late spermatogenic stages [51]. This evolutionary context provides a robust framework for testing the ability of foundation models to identify homologous cell types despite significant molecular divergence.

Foundation Model Application

The TranscriptFormer model represents a significant advancement for cross-species biological validation. As a generative, multi-species model for single-cell transcriptomics, it was trained on 112 million cells from 12 different species, covering 1.5 billion years of evolution [9]. For spermatogenesis research, TranscriptFormer demonstrates exceptional capability in identifying cell types in species not included in its training data (such as rhesus macaque and marmoset) and accurately transferring labels across related species [9]. This functionality enables researchers to annotate spermatogenic cell types in poorly characterized species using models trained on well-annotated reference datasets, significantly accelerating the mapping of spermatogenesis across mammals.

Table 1: Key Quantitative Findings from Cross-Species Spermatogenesis Studies

Finding	Measurement	Biological Significance
Evolutionary Rate Variation	Rate of expression evolution substantially higher in postmeiotic haploid cell types (rSD and eSD) compared to diploid spermatogenic cells [51].	Explains rapid evolution of the testis; suggests reduced pleiotropic constraints and haploid selection in late spermatogenesis [51].
Primate Analysis Resolution	Evolutionary rates progressively increase from late meiosis (pachytene SC) until the end of spermiogenesis (late eSD) [51].	Provides cellular source for rapid testis evolution and enables fine-grained analysis of primate spermatogenesis [51].
TranscriptFormer Performance	Can identify cell types in species not seen during training (rhesus macaque, marmoset) [9].	Enables translation of biological insights across species and annotation of cell types in unmapped species [9].
Single-Cell Atlas Scale	97,521 high-quality nuclei from 11 species with median of 1,856 RNA molecules per cell [51].	Provides comprehensive resource for investigating testis biology across mammals [51].

Experimental Design and Workflow

The following diagram illustrates the integrated computational and experimental workflow for cross-species validation of spermatogenic cell types using foundation models:

Detailed Experimental Protocols

Protocol 1: Cross-Species Single-Nucleus RNA Sequencing of Testicular Tissue

Purpose: To generate high-quality single-nucleus transcriptome data from testicular tissues across multiple mammalian species for foundational model training and validation [51].

Materials:

• Fresh or frozen testicular tissue samples from target species
• Nuclear isolation buffer (e.g., NP-40 or Triton X-100 based)
• DAPI or Hoechst stain for nuclear quantification
• 10x Genomics Chromium Controller and Single Cell 3' Reagent Kits
• High-throughput sequencer (Illumina NovaSeq or equivalent)

Procedure:

• Nuclear Isolation:
  • Homogenize 20-30 mg testicular tissue in cold nuclear isolation buffer using Dounce homogenizer
  • Filter homogenate through 40-μm cell strainer
  • Centrifuge at 500g for 5 minutes at 4°C
  • Resuspend pellet in 1 mL nuclear wash buffer
  • Count nuclei using hemocytometer with DAPI staining
  • Adjust concentration to 1,000 nuclei/μL

• Library Preparation:

  • Load 10,000 nuclei per sample into 10x Genomics Chromium Controller
  • Follow manufacturer's protocol for GEM generation, barcoding, and cDNA amplification
  • Perform 12 cycles of cDNA amplification
  • Construct libraries with sample indices

• Sequencing:

  • Pool libraries and sequence on Illumina platform
  • Target 50,000 read pairs per nucleus
  • Use 28 bp read 1 (cell barcode and UMI), 90 bp read 2 (transcript)

Quality Control:

• Sequence ≥ 50,000 nuclei per species
• Maintain mitochondrial gene percentage < 10%
• Ensure median genes per nucleus > 1,000
• Include biological replicates (1-3 per species) [51]

Protocol 2: Computational Integration and Evolutionary Analysis

Purpose: To integrate snRNA-seq data across species and quantify evolutionary changes in spermatogenic gene expression programs [51].

Materials:

• High-performance computing cluster
• Python 3.8+ with scanpy, scvi-tools, and numpy packages
• Ortholog mappings between species (Ensembl Compara)
• Custom scripts for evolutionary rate calculation

Procedure:

• Data Preprocessing:
  • Map reads to respective reference genomes
  • Generate count matrices using CellRanger or equivalent
  • Filter low-quality nuclei (genes < 200, mitochondrial reads > 10%)
  • Normalize counts using SCTransform or scran

• Cross-Species Integration:

  • Identify one-to-one orthologs across all species
  • Retain only orthologous genes for integration
  • Apply SCVI or Harmony for batch correction
  • Perform graph-based clustering on integrated space

• Cell Type Annotation:

  • Identify marker genes for each cluster
  • Ancell types using conserved marker genes:
    • Spermatogonia: UTF1, ZBTB16, FGFR3
    • Spermatocytes: SYCP1, SYCP3, TEX15
    • Spermatids: PRM1, PRM2, TNP1 [51]

• Evolutionary Rate Calculation:

  • Generate pseudo-bulk expression profiles for each cell type
  • Build expression phylogenies for each cell type
  • Calculate total branch lengths as measure of expression divergence
  • Compare evolutionary rates across spermatogenic stages [51]

Quality Control:

• Biological replicates should cluster together in PCA
• Known marker genes should show appropriate cell-type specificity
• Phylogenetic trees should recapitulate known species relationships

Protocol 3: Foundation Model Fine-Tuning for Spermatogenesis

Purpose: To adapt foundation models for cross-species annotation of spermatogenic cell types [9].

Materials:

• Pretrained TranscriptFormer model weights
• Annotated single-cell datasets from reference species (human, mouse)
• Unannotated datasets from target species
• GPU cluster with ≥ 4 A100 or equivalent GPUs

Procedure:

• Data Preparation:
  • Format single-cell data using standardized schema
  • Convert expression matrices to ranked gene tokens
  • Split data into training (70%), validation (15%), and test (15%) sets

• Model Fine-Tuning:

  • Initialize model with pretrained TranscriptFormer weights
  • Replace final classification layer with random initialization
  • Fine-tune using labeled data from reference species
  • Apply cross-entropy loss for cell-type classification
  • Use AdamW optimizer with learning rate 1e-5

• Cross-Species Prediction:

  • Extract cell embeddings from fine-tuned model
  • Train k-NN classifier on reference species embeddings
  • Predict cell types for target species using embeddings
  • Evaluate using manual annotation or marker genes

Quality Control:

• Achieve >90% accuracy on reference species test set
• Verify predictions using known marker genes in target species
• Ensure embedding space preserves biological relationships

Signaling Pathways and Molecular Regulation

The molecular regulation of spermatogenesis involves complex signaling pathways that are conserved across mammals. The following diagram illustrates key pathways and their components identified through cross-species analysis:

Research Reagent Solutions

Table 2: Essential Research Reagents for Cross-Species Spermatogenesis Studies

Reagent/Category	Specific Examples	Function/Application
Single-Cell Platforms	10x Genomics Chromium, BD Rhapsody	High-throughput single-cell RNA sequencing of testicular cell populations [51] [53]
Nuclear Isolation Kits	10x Nucleus Isolation Kit, Active Motif	Isolation of intact nuclei from frozen testicular tissues for snRNA-seq [51]
Cell Type Markers	UTF1 (spermatogonia), SYCP3 (spermatocytes), PRM1 (spermatids)	Identification and validation of spermatogenic cell types across species [51]
Foundation Models	TranscriptFormer, Nicheformer	Cross-species cell type annotation and prediction of spatial context [9] [3]
Spatial Transcriptomics	10x Visium, MERFISH, Xenium	Validation of cellular niches and germ cell-soma interactions [3]
Bioinformatics Tools	SCANPY, Seurat, SCVI	Data integration, clustering, and evolutionary analysis [51]

This application note provides detailed protocols for the biological validation of foundation models using spermatogenesis as a case study. The integrated approach combining single-nucleus transcriptomics, evolutionary analysis, and foundation model fine-tuning enables robust cross-species cell type annotation. The conserved yet rapidly evolving nature of spermatogenesis makes it an ideal system for testing the limits of foundation model generalization across species. These protocols establish a framework for validating foundation models in other complex biological systems, with particular relevance for translational research in male infertility and contraceptive development.

The accurate identification of cell types across species is a cornerstone of comparative biology, with profound implications for understanding evolution, developmental biology, and disease mechanisms. The emergence of single-cell foundation models (scFMs) represents a transformative approach for deciphering the "language" of cells by applying large-scale deep learning to vast single-cell transcriptomic datasets [10]. These models, pretrained on millions of cells, learn fundamental biological principles that can potentially generalize across taxonomic boundaries.

However, cross-species prediction faces significant biological and computational challenges. Studies consistently demonstrate that marker gene transferability decreases as evolutionary distance increases [54]. Research on primate embryoid bodies revealed that human marker genes were less effective in macaques and vice versa, highlighting fundamental limitations in direct annotation transfer [54]. Similarly, analysis of primate brain tissues identified that while 76% of genes showed conserved expression patterns, the remaining 24% exhibited extensive differences between human and non-human primates [55].

This Application Note examines the current capabilities and limitations of scFMs in cross-species cell annotation, with specific focus on prediction accuracy from primates to zebrafish. We provide a structured framework for evaluating model performance, detailed protocols for implementation, and practical solutions for overcoming biological divergence in translational studies.

Key Challenges in Cross-Species Prediction

Biological Divergence Across Species

The core challenge in cross-species prediction lies in the fundamental biological differences that accumulate over evolutionary time, manifesting at multiple molecular levels.

Genetic and Regulatory Differences: Analysis of five primate species (human, chimp, gorilla, macaque, and marmoset) revealed that 3,383 out of 14,131 genes (24%) showed extensive expression differences in homologous cell types [55]. These divergent genes were particularly associated with synaptic assembly and function, with nearly half showing expression divergence limited to glial cell types.

Marker Gene Transferability Limitations: A systematic study of embryoid bodies from four primate species demonstrated that the discriminatory power of marker genes decreases with phylogenetic distance [54]. Human marker genes proved less effective for annotating macaque cells, indicating that even between closely related species, direct annotation transfer has limitations that must be accounted for in prediction models.

Technical Limitations in Current scFMs

While scFMs show remarkable potential, several technical constraints currently limit their cross-species applicability.

Data Quality and Integration Challenges: Single-cell data exhibits characteristics of high sparsity, high dimensionality, and low signal-to-noise ratio [56]. When integrating data across species, additional complications arise from batch effects, technical noise, and varying processing steps [10]. These issues are compounded when comparing evolutionarily distant species like primates and zebrafish.

Architectural Constraints: Most scFMs use transformer architectures that require sequential input, but gene expression data lacks natural ordering [10] [56]. Current solutions include ranking genes by expression levels or partitioning them into expression bins, but these arbitrary sequences may not capture biologically meaningful relationships conserved across species.

Table 1: Key Challenges in Cross-Species Cell Annotation

Challenge Category	Specific Limitations	Impact on Prediction Accuracy
Biological Divergence	Decreasing marker gene transferability with evolutionary distance	Reduced annotation accuracy for distant species
	Differences in gene co-expression networks (139 genes with human-specific connectivity identified)	Limited functional inference across species
	Variation in cell type-specific gene expression (3,383 genes with primate differences)	Incorrect cell type matching
Technical Limitations	Data sparsity and high dimensionality	Increased noise in model predictions
	Non-sequential nature of omics data	Suboptimal representation learning
	Batch effects across experiments and species	Confounded biological signals

Quantitative Assessment of Current Capabilities

Performance Metrics for Cross-Species Prediction

Evaluating scFM performance requires multiple metrics to capture different dimensions of prediction quality. Based on benchmarking studies, we recommend the following assessment framework.

Table 2: Performance Metrics for Cross-Species scFM Evaluation

Metric Category	Specific Metrics	Interpretation in Cross-Species Context
Cell Type Annotation Accuracy	Overall accuracy, F1-score, Precision/Recall	Measures direct prediction correctness against ground truth
	Lowest Common Ancestor Distance (LCAD)	Ontological proximity of misclassified cell types [56]
	scGraph-OntoRWR	Consistency with known biological relationships [56]
Dataset Integration Quality	Batch integration scores	Separation of biological vs. technical variation
	Label transfer accuracy	Effectiveness of annotation between species
Biological Relevance	Gene ontology term prediction	Functional knowledge capture in embeddings
	Tissue specificity prediction	Conservation of spatial expression patterns

Current Performance Benchmarks

Comprehensive benchmarking of six prominent scFMs against traditional methods reveals a nuanced performance landscape. Under the realistic conditions of cross-species prediction, no single scFM consistently outperforms others across all tasks [56]. The relative performance depends heavily on specific factors including dataset size, biological complexity, and evolutionary distance between source and target species.

For evolutionarily close species (primate-to-primate annotation), scFMs demonstrate robust performance with accuracy metrics often exceeding 0.85 for well-conserved cell types [55]. However, performance degradation occurs with increasing phylogenetic distance, particularly for neuronal and immune cell types that exhibit accelerated evolutionary divergence [55].

Notably, simpler machine learning models sometimes outperform complex foundation models in specific cross-species tasks, particularly when training data is limited or computational resources are constrained [56]. This suggests that scFMs provide the greatest value when leveraging their pretrained knowledge base through transfer learning, rather than applying them in zero-shot scenarios across large evolutionary distances.

Experimental Protocols for Cross-Species Validation

Orthologous Cell Type Identification Protocol

Accurate cross-species prediction requires careful identification of orthologous cell types as a foundation for model training and validation.

Procedure:

• Species-Specific Clustering: Process and cluster cells separately for each species in the study (primates and zebrafish). Generate at least double the number of high-resolution clusters (HRCs) as expected cell types to avoid losing rare populations [54].
• Cross-Species Classification: Use the HRCs of one species (e.g., human) as a reference to classify cells of other species (e.g., zebrafish) using SingleR [54]. Perform pairwise comparisons reciprocally for each species pair.
• Reciprocal Best-Hit Analysis: For each comparison, calculate the fraction of cells annotated as the other HRC in both directions. Perfect orthology is indicated when all cells of HRC-B are assigned to HRC-A (human→zebrafish) and all cells of HRC-A are assigned to HRC-B (zebrafish→human) [54].
• Distance Matrix Construction: Use the resulting annotation fractions to build a distance matrix as input for hierarchical clustering to identify orthologous clusters across species [54].
• Validation: Assess cross-species replicability of identified cell types using MetaNeighbor, which evaluates transcriptomic similarity within and across species [55].

Key Considerations: This protocol requires careful handling of uneven cell type compositions between species. The interactive Shiny application (https://shiny.bio.lmu.de/CrossSpeciesCellType/) provides a practical implementation framework [54].

scFM Fine-Tuning Protocol for Cross-Species Application

Effective adaptation of pretrained scFMs for cross-species prediction requires systematic fine-tuning and validation.

Procedure:

• Model Selection: Choose appropriate scFM based on task requirements. Geneformer [56] excels at gene-level analyses, while scGPT [10] [56] shows strengths in generation tasks, and UCE [56] provides robust cross-species embeddings.
• Orthologue Mapping: Map zebrafish genes to human orthologues using established databases (OrthoDB [55], Ensembl Compare). Only use one-to-one orthologues for initial model fine-tuning to ensure clean signal learning.
• Cross-Species Tokenization: Adapt the model's tokenization strategy to handle cross-species applications. This may involve:
  • Creating combined vocabulary of human and zebrafish orthologues
  • Adding species-specific tokens to distinguish origin [10]
  • Incorporating phylogenetic information into positional encodings
• Staged Fine-Tuning:
  • Phase 1: Fine-tune on primate data only to adapt to specific cell types of interest
  • Phase 2: Continue fine-tuning on paired primate-zebrafish datasets with orthologous cell types
  • Use gradual unfreezing of layers, starting from the output layers
• Validation: Assess performance using the metrics in Table 2, with particular emphasis on biological relevance metrics like scGraph-OntoRWR [56].

Technical Notes: Limit fine-tuning to 10-20% of original pretraining epochs to prevent catastrophic forgetting. Use batch sizes that maintain representation of both species in each update.

Research Reagent Solutions

Implementing cross-species prediction requires specific computational tools and resources. The following table summarizes essential solutions for scFM-based cross-species annotation.

Table 3: Essential Research Reagents for Cross-Species scFM Implementation

Reagent Category	Specific Tools/Databases	Function in Cross-Species Prediction
Pretrained Models	Geneformer, scGPT, UCE, scFoundation [56]	Provide base models for fine-tuning with biological knowledge
Data Resources	CZ CELLxGENE [10], Human Cell Atlas [10], PanglaoDB [10]	Curated single-cell data for training and validation
Orthology Databases	OrthoDB [55], Ensembl Compare	Map genes across species for model alignment
Evaluation Tools	MetaNeighbor [55], SingleR [54], scGraph-OntoRWR [56]	Validate cross-species cell type correspondence
Annotation Databases	Cell Ontology, Synaptic Gene Ontology (SynGO) [55]	Provide standardized vocabulary for cell types

Cross-species prediction from primates to zebrafish represents both a formidable challenge and tremendous opportunity for advancing comparative biology and translational research. Single-cell foundation models provide a powerful framework for addressing this challenge, but their effective application requires careful consideration of biological divergence, appropriate model selection, and systematic validation.

The protocols and metrics presented here establish a foundation for rigorous cross-species prediction that acknowledges both the capabilities and limitations of current approaches. As scFM architectures evolve and incorporate multi-omic data, their ability to capture conserved biological principles across larger evolutionary distances will continue to improve, potentially bridging the gap between primate and zebrafish biology with increasing accuracy.

Future directions should focus on incorporating protein structure information [57], developing explicit models of evolutionary distance, and creating standardized cross-species benchmarking datasets. These advances will accelerate the deployment of scFMs in both basic research and drug development, where cross-species extrapolation remains a critical challenge.

Cross-species cell annotation represents a transformative approach in single-cell biology, enabling the deciphering of cellular function, development, and disease across evolutionary timescales. By leveraging foundational models trained on vast, evolutionarily diverse datasets, researchers can now identify conserved and divergent cellular states, offering unprecedented insights into fundamental biological processes and accelerating therapeutic development. This paradigm shift moves beyond single-species analysis to a unified framework for understanding cellular biology from a comparative perspective. These foundation models, trained on hundreds of millions of cells, serve as powerful virtual instruments, allowing scientists to ask complex biological questions and test in-silico hypotheses before conducting wet-lab experiments [9]. This document provides detailed application notes and protocols for employing these models to reveal novel biological insights into evolutionary conservation and divergence, framed within the broader thesis of cross-species cell annotation research.

Application Notes: Core Principles and Biological Insights

Conserved Microglial Ontogeny and Divergent Functions

The evolutionary conservation of core developmental programs is exemplified by microglia, the resident immune cells of the central nervous system. Across vertebrate species, microglia share a conserved origin from primitive yolk sac-derived macrophages (or analogous structures like the rostral blood island in zebrafish) that colonize the embryonic brain early in development [58]. This ontogenetic pathway is a conserved hallmark, independent of definitive bone marrow hematopoiesis.

Despite this shared origin, microglia exhibit significant functional and phenotypic divergence across species. Their morphology, gene expression profiles, and responses to stimuli vary considerably, reflecting evolutionary adaptations shaped by factors such as lifespan, regenerative capacity, and overall immune system architecture [58]. For example, in contrast to mammals, yolk sac-derived microglia in birds are transient and are largely replaced by definitive hematopoietic cells later in development [58]. This interplay between conserved origins and species-specific functions makes microglia a prime model for studying evolutionary conservation and divergence using cross-species annotation tools.

Capabilities of Cross-Species Foundation Models

Modern foundational models, pre-trained on massive single-cell transcriptomics datasets encompassing multiple species, have demonstrated remarkable capabilities for cross-species biological discovery. The following table summarizes key insights and performance metrics from state-of-the-art models.

Table 1: Performance and Insights from Cross-Species Foundation Models

Model Name	Training Scale	Key Demonstrated Capability	Performance Highlight	Biological Insight
TranscriptFormer [9]	112 million cells from 12 species (~1.5 billion years of evolution)	Predict cell types in out-of-distribution species (e.g., rhesus macaque, marmoset)	Surpassed baseline models at identifying SARS-CoV-2-infected cells without fine-tuning [9]	Enables translation of gene expression patterns and biological mechanisms across vast evolutionary distances.
CellFM [35]	100 million human cells	Gene function prediction, perturbation prediction, and cell annotation.	Outperforms existing models in gene function prediction and gene-gene relationship capturing [35]	Provides a unified model to represent cellular states, overcoming data noise and sparsity.
LICT [59]	Evaluated on diverse datasets (PBMCs, embryos, gastric cancer)	Objective reliability assessment for cell type annotation using multi-LLM integration.	Achieved a 48.5% full match rate with manual annotations on low-heterogeneity embryo data [59]	Addresses annotation reliability, a critical challenge in single-cell biology, especially for ambiguous cell clusters.

These models function as a "virtual instrument" for researchers. A key application is the prompting of generative models like TranscriptFormer to simulate gene-gene interactions within specific cell types and organisms, thereby identifying co-expressed genes and predicting underlying regulatory networks [9]. Furthermore, their ability to produce contextualized gene embeddings that are cell-specific offers a more granular understanding of gene function compared to static annotations [9].

Quantitative Quality Control for Multi-Modal Data

Rigorous quality control is a prerequisite for reliable cross-species annotation. This is especially critical for multi-modal data like CITE-Seq, which simultaneously measures gene expression and surface protein abundance. The CITESeQC software package provides a systematic, quantitative framework for this purpose [60]. It employs metrics like Shannon entropy to quantify the cell type-specificity of gene or protein expression across clusters, with lower entropy values indicating higher specificity. It also assesses the correlation between gene expression and the abundance of their corresponding proteins, an expected biological relationship that serves as an internal quality check [60]. This objective assessment is vital for ensuring that downstream analyses and model predictions are built upon high-quality, reliable data.

Experimental Protocols

Protocol 1: Cross-Species Cell Type Annotation and Validation Using TranscriptFormer

This protocol details the use of foundation models for annotating cell types across different species, which is fundamental for identifying evolutionarily conserved and divergent cell states.

Table 2: Research Reagent Solutions for Cross-Species Annotation

Item Name	Function/Explanation
TranscriptFormer Model [9]	A generative, multi-species foundation model for single-cell transcriptomics that serves as the primary annotation engine.
CZ CELLxGENE / ZebraHub / Tabula Sapiens Data [9]	Curated, publicly available single-cell atlases that provide the foundational data for model training and validation.
CITESeQC R Package [60]	A software package for performing multi-layered, quantitative quality control on CITE-Seq (RNA + protein) data prior to analysis.
LICT (LLM-based Identifier for Cell Types) [59]	A tool that leverages multiple large language models to provide objective, reference-free assessment of cell annotation reliability.

Procedure:

• Data Acquisition and Curation: Compile a single-cell transcriptomics dataset from the species of interest. Public data can be sourced from repositories like CZ CELLxGENE.
• Preprocessing and Quality Control: Process the raw data using a standardized workflow. This includes quality control to filter low-quality cells and genes, gene name standardization according to guidelines like those from the HUGO Gene Nomenclature Committee (HGNC), and data normalization [35]. For CITE-Seq data, use CITESeQC to quantitatively assess the quality of both RNA and protein data [60].
• Model Application:
  • Load the pre-trained TranscriptFormer model.
  • Generate cell embeddings for the query dataset. These embeddings are numerical representations that capture the transcriptional state of each cell.
  • Use the model's integrated classifiers or nearest-neighbor matching against reference atlases to predict cell types for the query cells [9].
• Annotation Reliability Assessment: Input the top marker genes for each annotated cell cluster into the LICT tool. LICT will employ its multi-model integration and "talk-to-machine" strategy to provide an objective credibility score for the annotations, helping to identify potentially mislabeled or low-confidence populations [59].
• Biological Interpretation and Validation:
  • Compare the annotated cell states with known states from related species to identify putative conserved and lineage-specific cell types.
  • For critical findings, experimental validation is recommended. This could involve techniques like multiplexed immunofluorescence (mIF) on tissue sections to confirm the presence and identity of predicted cell types [61].

Cross-Species Annotation Workflow

Protocol 2: In Silico Perturbation Prediction to Probe Gene Networks

This protocol uses foundation models to predict the transcriptomic consequences of genetic perturbations, enabling the study of gene regulatory networks across species and conditions.

Procedure:

• Model Selection and Setup: Select a model trained on a large and diverse human dataset, such as CellFM, which is specifically designed for perturbation prediction [35].
• Simulation of Perturbation:
  • For a given cell's transcriptomic profile, specify a target gene to be perturbed (e.g., knocked out or overexpressed).
  • Use the model to predict the new expression levels of all other genes in the presence of this perturbation.
• Analysis of Predicted Response:
  • Compare the predicted post-perturbation profile to the original profile.
  • Identify significantly differentially expressed genes. These genes likely form a network that is functionally related to or regulated by the target gene.
• Cross-Species and Cross-Condition Comparison:
  • Perform the same in-silico perturbation in analogous cell types from different species (e.g., human, mouse, zebrafish) using a cross-species model.
  • Compare the resulting gene networks to identify conserved core pathways and species-specific adaptive responses.

In-Silico Perturbation Analysis

Protocol 3: Automated Cell Classification on H&E Images Using mIF-Derived Labels

This protocol leverages multiplexed immunofluorescence to generate high-quality training labels for a deep learning model that can classify cell types directly from standard H&E-stained histopathology images, enabling scalable spatial biomarker discovery.

Procedure:

• Sample Preparation and Consecutive Staining:
  • Obtain formalin-fixed paraffin-embedded (FFPE) tissue sections.
  • Perform multiplexed immunofluorescence (mIF) staining using a panel of antibodies for cell lineage protein markers (e.g., pan-CK for tumor cells, CD3/CD20 for lymphocytes, CD68 for macrophages, CD66b for neutrophils) [61].
  • Image the mIF-stained section.
  • Destain the section and subsequently perform standard H&E staining on the same tissue section.
  • Image the H&E-stained section.
• Image Co-registration and Label Transfer:
  • Co-register the mIF and H&E images at the single-cell level using an initial rigid transformation followed by a non-rigid registration to achieve subcellular accuracy [61].
  • Use unsupervised clustering (e.g., Leiden algorithm) on the mIF protein intensity data to define cell types objectively [61].
  • Transfer these high-confidence cell type labels from the mIF image to the corresponding, aligned cells in the H&E image.
• Model Training:
  • Extract image patches centered on each annotated cell from the H&E whole-slide images.
  • Train a deep learning model (e.g., a convolutional neural network combining self-supervised learning with domain adaptation) to classify the four major cell types based on the H&E morphology, using the mIF-derived labels as ground truth [61].
• Spatial Analysis and Biomarker Discovery:
  • Apply the trained model to new H&E whole-slide images to generate cell-type maps.
  • Quantify spatial interactions between different immune cell types and tumor cells.
  • Correlate these spatial interaction patterns with clinical outcomes, such as patient survival or response to immunotherapy [61].

H&E Cell Classification via mIF

A fundamental challenge in computational biology is developing models that perform reliably on data from species not encountered during training, known as out-of-distribution (OOD) species. This capability is crucial for creating truly generalizable biological foundation models that can accelerate discovery across the tree of life. Recent research has revealed a significant generalization gap where models excel on familiar species but fail to maintain predictive performance when applied to evolutionarily distant organisms [62]. This application note examines the current state of OOD generalization in cross-species cell annotation models, provides experimental protocols for evaluation, and offers visualization tools to guide research in this emerging field.

Performance Metrics and Quantitative Comparisons

Key Performance Indicators for OOD Evaluation

Robust evaluation requires multiple metrics to assess different aspects of model generalization. The table below summarizes the primary quantitative measures used in recent studies.

Table 1: Key Performance Metrics for OOD Species Generalization

Metric	Definition	Interpretation	Typical Performance Range
Cell Type Annotation Accuracy	Percentage of cells correctly classified in unseen species	Measures basic transfer learning capability	40-85% depending on evolutionary distance [9]
Neural Predictivity Score	Correlation between model predictions and actual neuronal responses to OOD stimuli	Quantifies how well models generalize to novel visual patterns	Varies significantly across model architectures [62]
Disease State Prediction AUC	Area Under Curve for identifying infected/diseased cells in new species	Evaluates clinical or pathological relevance	>0.75 in state-of-the-art models [9]
Gene-Gene Interaction Accuracy	Precision in predicting conserved genetic interactions	Tests understanding of fundamental biological mechanisms	Higher for evolutionarily conserved pathways [9]

Comparative Performance Across Model Architectures

Recent benchmarking studies reveal substantial differences in how various architectures handle OOD generalization. TranscriptFormer has demonstrated state-of-the-art performance, accurately identifying cell types in unseen species like rhesus macaque and marmoset without species-specific training data [9]. In comparative analyses, adversarially robust models often yield substantially higher generalization in neural predictivity, though the degree of robustness itself doesn't directly predict performance [62]. Surprisingly, performance on common computer vision OOD benchmarks does not correlate with OOD neural predictivity performance, suggesting domain-specific evaluation is essential [62].

Table 2: Model Architecture Comparison for OOD Generalization

Model Type	OOD Cell Type Accuracy	Training Data Scale	Strengths	Limitations
Transformer-based (TranscriptFormer)	70-85% for closely related species [9]	112 million cells across 12 species [9]	Cross-species transfer, generative capabilities	Computational intensity for training [10]
Adversarially Robust Models	Improved but unquantified neural predictivity [62]	Varies by implementation	Resistance to synthetic OOD stimuli	Limited single-cell implementation
Encoder-Based (scBERT)	Moderate for in-domain species [10]	Millions of single-cell transcriptomes [10]	Effective for classification tasks	Limited generative capacity
Decoder-Based (scGPT)	Good interpolation, limited extrapolation [10]	Diverse single-cell corpora [10]	Strong generative performance	May learn ecologically implausible relationships [63]

Experimental Protocols for OOD Generalization Testing

Cross-Species Cell Type Annotation Protocol

Purpose: To evaluate model performance on cell type identification in evolutionarily distant species not included in training data.

Materials:

• Pre-trained cross-species foundation model (e.g., TranscriptFormer, scGPT)
• Single-cell RNA-seq data from target OOD species
• Reference cell type markers for validation
• Computational environment with appropriate GPU resources

Procedure:

• Data Preprocessing: Normalize target species gene expression data using the same methodology applied during model training.
• Gene Orthology Mapping: Map gene identifiers between species using established orthology databases to ensure feature alignment.
• Embedding Generation: Process cells through the model to generate latent embeddings without fine-tuning.
• Cell Type Prediction: Apply the model's classification head or use k-NN classification against reference embeddings.
• Validation: Compare predictions with ground truth labels where available, using manual validation with marker genes when absent.

Validation Methods:

• Cross-species marker expression: Verify predicted cell types express appropriate orthologous marker genes.
• Cluster coherence: Assess whether predicted cell types form coherent clusters in visualization.
• Manual curation: Have domain experts review predictions for biological plausibility.

OOD Neural Predictivity Assessment Protocol

Purpose: To measure how well model representations predict neural responses to novel, out-of-distribution visual stimuli.

Materials:

• Primate electrophysiology data or equivalent neuronal recording data
• Deep neural network models with varying architectures
• Synthetically generated OOD visual stimuli
• Standardized neural predictivity benchmarking pipeline

Procedure:

• Stimulus Selection: Choose OOD stimuli that differ substantially from training distribution.
• Neural Response Recording: Collect neuronal population responses to stimuli in relevant brain regions.
• Model Response Extraction: Generate model activations for the same stimulus set.
• Predictive Model Fitting: Train linear regression to predict neural responses from model activations.
• Generalization Assessment: Evaluate prediction performance on held-out OOD stimuli.

Key Considerations:

• Adversarially robust models often show better OOD generalization [62].
• Current computer vision benchmarks may not correlate with neural predictivity [62].

Visualization of Cross-Species Model Architectures

Tokenization and Processing Workflow

Figure 1: Single-Cell Foundation Model Architecture

Cross-Species Generalization Evaluation

Figure 2: OOD Species Evaluation Workflow

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Research Reagents for Cross-Species Single-Cell Studies

Reagent/Resource	Function	Example Sources/Platforms
Cross-Species Cell Atlases	Training data for foundation models	CZ CELLxGENE, Tabula Sapiens, ZebraHub, Human Cell Atlas [9]
Orthology Mapping Databases	Gene identifier conversion across species	Ensembl Compara, OrthoDB, HGNC, MGI
Single-Cell Foundation Models	Base models for transfer learning	TranscriptFormer, scBERT, scGPT, scFMs [10] [9]
Adversarial Training Frameworks	Improving model robustness	PyTorch Adversarial, TensorFlow Robustness
Contrast Enhancement Networks	Image preprocessing for morphological data	FCE-Net for biomedical optical images [64]
Spatial Transcriptomics Data	Adding spatial context to single-cell data	10X Genomics Visium, MERFISH, seqFISH+
Cell Type Annotation Tools	Reference-based cell labeling	scPred, SingleR, SCINA [16]

Discussion and Future Directions

The development of models that generalize effectively to out-of-distribution species represents both a significant challenge and opportunity in computational biology. Current evidence suggests that scale and diversity of training data are crucial factors, with models like TranscriptFormer demonstrating that training on evolutionarily diverse corpora (covering 1.5 billion years of evolution) enables better generalization [9]. However, simply increasing model size or training data may be insufficient if not paired with architectural innovations specifically designed for OOD robustness.

A promising direction is the integration of adversarial training techniques, which have shown benefits for neural predictivity on OOD stimuli despite not improving performance on standard computer vision benchmarks [62]. This suggests that biological relevance requires specialized approaches beyond those developed for general computer vision tasks. Additionally, developing better tokenization strategies for non-sequential biological data remains an active research area, with current approaches including gene ranking by expression, value binning, and incorporation of biological metadata [10].

Future work should focus on standardized benchmarking for cross-species generalization, development of biologically-motivated regularization techniques, and integration of multi-modal data to provide additional constraints that could improve OOD performance. As these models mature, they hold the potential to transform comparative biology and accelerate the development of therapies tested across appropriate model organisms.

Conclusion

Cross-species cell annotation foundation models represent a paradigm shift in computational biology, successfully integrating evolutionary divergence with deep learning to create powerful tools for deciphering cellular function across species. The synthesis of findings reveals that models like GeneCompass, TranscriptFormer, and CAME demonstrate remarkable capabilities in transferring cell type annotations, predicting disease states, and uncovering conserved regulatory mechanisms. Future directions will involve expanding to non-model organisms, integrating multi-omic data, and enhancing model interpretability for clinical translation. For biomedical research, these models promise to accelerate drug target discovery, improve translation from model organisms to humans, and ultimately enable predictive biology across the tree of life, significantly advancing our ability to understand and treat human disease.

References

• 1. Cross-species cell-type assignment from single-cell RNA-seq ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9977153/]
• 2. integrating single-cell RNA-seq datasets across species ... [https://www.nature.com/articles/s41592-024-02191-z]
• 3. Nicheformer: a foundation model for single-cell and spatial ... [https://www.nature.com/articles/s41592-025-02814-z]
• 4. Single-cell foundation models: bringing artificial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12586647/]
• 5. Biology-driven insights into the power of single-cell ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12492631/]
• 6. Cross-species imputation and comparison of single-cell ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03493-x]
• 7. Benchmarking cell type and gene set annotation by large ... [https://www.nature.com/articles/s41467-025-64511-x]
• 8. Get, set, annotate - an overview of cell type prediction ... [https://www.cellkb.com/blog/cell_type_annotation_review]
• 9. TranscriptFormer: AI Model for Cross-Species Biology [https://chanzuckerberg.com/blog/transcriptformer-model-overview/]
• 10. Single-cell foundation models: bringing artificial ... [https://www.nature.com/articles/s12276-025-01547-5]
• 11. Evolutionary divergence in CTCF-mediated chromatin ... [https://www.nature.com/articles/s41467-025-58275-7]
• 12. Color contrast - Accessibility - MDN Web Docs - Mozilla [https://developer.mozilla.org/en-US/docs/Web/Accessibility/Guides/Understanding_WCAG/Perceivable/Color_contrast]
• 13. Understanding Success Criterion 1.4.3: Contrast (Minimum) [https://www.w3.org/WAI/WCAG21/Understanding/contrast-minimum.html]
• 14. Text has enhanced contrast | ACT Rule | WAI [https://www.w3.org/WAI/standards-guidelines/act/rules/09o5cg/]
• 15. Elements must meet minimum color contrast ratio thresholds [https://dequeuniversity.com/rules/axe/4.8/color-contrast]
• 16. Cross-species single-cell transcriptomic atlas of retinal ... [https://www.sciencedirect.com/science/article/pii/S2211124725009350]
• 17. Single-cell transcriptomic and genomic changes in the ... [https://www.nature.com/articles/s41586-025-09435-8]
• 18. Comparison and development of cross-study normalization ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11386461/]
• 19. Single-species versus multiple-species models [https://www.sciencedirect.com/science/article/abs/pii/S0304380003002278]
• 20. Chapter 7 Cell type annotation | Basics of Single- ... [https://bioconductor.org/books/3.13/OSCA.basic/cell-type-annotation.html]
• 21. Annotation of cell types (ACT): a convenient web server ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10623726/]
• 22. Beyond single-species models: leveraging multispecies ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11846506/]
• 23. PCLDA: An interpretable cell annotation tool for single- ... [https://www.sciencedirect.com/science/article/pii/S2001037025002818]
• 24. Automated Annotation of Single Cell Data (EXPERIMENTAL) [https://cellxgene.cziscience.com/docs/05__Annotate%20and%20Analyze%20Your%20Data/5_5__Automatic%20Annotation]
• 25. Transformer-based Single-Cell Language Model: A Survey [https://arxiv.org/html/2407.13205v1]
• 26. scGREAT: Transformer-based deep-language model for gene ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10951644/]
• 27. Learning the natural history of human disease with ... [https://www.nature.com/articles/s41586-025-09529-3]
• 28. scTab: Scaling cross-tissue single-cell annotation models [https://www.nature.com/articles/s41467-024-51059-5]
• 29. GeneCompass: deciphering universal gene regulatory ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11615217/]
• 30. deciphering universal gene regulatory mechanisms with a ... [https://pubmed.ncbi.nlm.nih.gov/39375485/]
• 31. TranscriptFormer | v0.6.0 - Virtual Cells Platform [https://virtualcellmodels.cziscience.com/model/01966441-339f-77f7-aa06-f67636f865dc]
• 32. Genomic language models: opportunities and challenges [https://www.sciencedirect.com/science/article/abs/pii/S0168952524002956]
• 33. Urban Institute Data Visualization style guide - GitHub Pages [http://urbaninstitute.github.io/graphics-styleguide/]
• 34. Scale-free and unbiased transformer with tokenization for ... [https://www.sciencedirect.com/science/article/abs/pii/S003132032500384X]
• 35. CellFM: a large-scale foundation model pre-trained on ... [https://www.nature.com/articles/s41467-025-59926-5]
• 36. Scale-free and unbiased transformer with tokenization for cell ... [https://scholars.duke.edu/publication/1675794]
• 37. Article BioLLM: A standardized framework for integrating ... [https://www.sciencedirect.com/science/article/pii/S2666389925001746]
• 38. Identifying critical States of complex diseases by local ... [https://www.nature.com/articles/s41598-025-94521-0]
• 39. Time-independent disease state identification defines ... [https://pubmed.ncbi.nlm.nih.gov/40348636/]
• 40. Make Mine a Double: Mapping Gene Networks to Decode ... [https://biosciences.lbl.gov/2025/11/07/make-mine-a-double-mapping-gene-networks-to-decode-function/]
• 41. A global genetic interaction map of a human cell reveals ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12236771/]
• 42. strategies for cross- Benchmarking of single-cell... species integration [https://www.nature.com/articles/s41467-023-41855-w?error=cookies_not_supported&code=86570e17-22f0-46e0-bc93-88b193f2c4a3]
• 43. deep learning methods for biologically Benchmarking ... conserved [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03869-z]
• 44. Efficient large-scale protein sequence comparison and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3315314/]
• 45. An approach of orthology detection from homologous ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2553584/]
• 46. Mixing genome annotation methods in a comparative analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9346927/]
• 47. Benchmarking strategies for cross-species integration of ... [https://www.nature.com/articles/s41467-023-41855-w]
• 48. Annotating the genome at single-nucleotide resolution with ... [https://www.nature.com/articles/s41592-025-02881-2]
• 49. Reproducible single-cell annotation of programs ... [https://www.nature.com/articles/s41592-025-02793-1]
• 50. Understanding and Mitigating Batch Effects in Histopathology [https://www.trillium.de/en/journals/trillium-pathology/archive/2025/tp-1/2025/computational-pathology/image-analysis-understanding-and-mitigating-batch-effects-in-histopathology.html]
• 51. The molecular evolution of spermatogenesis across ... [https://www.nature.com/articles/s41586-022-05547-7]
• 52. Decoding the Spermatogenesis Program: New Insights from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10722372/]
• 53. Single-cell RNA sequencing uncovers dynamic roadmap ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-025-11944-9]
• 54. Identification and comparison of orthologous cell types ... [https://elifesciences.org/reviewed-preprints/105398]
• 55. Comparative single-cell transcriptomic analysis of primate ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10627823/]
• 56. Biology-driven insights into the power of single-cell foundation ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03781-6]
• 57. Rapid key gene discovery for bacterial shape [https://pmc.ncbi.nlm.nih.gov/articles/PMC12581320/]
• 58. Microglia across evolution: from conserved origins to ... [https://www.nature.com/articles/s41423-025-01368-6]
• 59. Evaluation of cell type annotation reliability using a large ... [https://www.nature.com/articles/s42003-025-08745-x]
• 60. Quantitative measures to assess the quality of cellular ... [https://www.frontiersin.org/journals/bioinformatics/articles/10.3389/fbinf.2025.1630161/full]
• 61. Automated cell annotation and classification on ... [https://www.nature.com/articles/s41467-025-61349-1]
• 62. How well do models of visual cortex generalize to out ... [https://pubmed.ncbi.nlm.nih.gov/38820563/]
• 63. A comparison of machine learning and statistical species ... [https://www.sciencedirect.com/science/article/pii/S0304380023000819]
• 64. FCE-Net: a fast image contrast enhancement method ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9208612/]