Protein Large Language Models (pLLMs) Assist Cross-Species Single-Cell Transcriptome Integration: A New Paradigm for Gene Homology Mapping

Title: Protein Large Language Models (pLLMs) Assist Cross-Species Single-Cell Transcriptome Integration: A New Paradigm for Gene Homology Mapping Abstract: This article introduces a cross-species single-cell transcriptome integration method based on Protein Large Language Models (pLLMs), which achieves gene homology mapping through protein sequence embedding, providing a new tool for comparative genomics and evolutionary biology research. Keywords: Protein language model, cross-species integration, single-cell transcriptome, gene homology mapping, ESM-2, computational biology, comparative genomics Original author/maintainer: KKzhongyi Source platform: GitHub Original title: pLLM-cross-species-integration Original link: https://github.com/KKzhongyi/pLLM-cross-species-integration Source release time/update time: 2026-05-27T14:14:22Z

Core观点: This project proposes a cross-species single-cell transcriptome integration method based on protein large language models (e.g., ESM-2), which achieves gene homology mapping through protein sequence embedding. It solves the problems of traditional methods relying on databases and ignoring functional conservation, providing a new tool for comparative genomics and evolutionary biology.

Background and Challenges

Cross-species single-cell transcriptome integration is a core problem in computational biology. When comparing the same cell type across different species, the biggest obstacle is inconsistent gene naming—different names for homologous genes make direct alignment difficult. Traditional methods rely on pre-built homology databases (e.g., Ensembl Compara, NCBI HomoloGene), but they have issues like incomplete coverage and lagging updates, and only based on DNA sequence similarity, ignoring protein functional conservation.

Method Framework

Two-Stage Mapping Strategy

1. Protein embedding generation: Extract the protein sequence encoded by each gene, use a pre-trained pLLM (e.g., ESM-2) to generate embedding vectors, capturing biochemical properties and structural tendencies.
2. Cross-species nearest neighbor matching: Calculate cosine similarity/Euclidean distance to establish one-to-one gene mapping. Advantages include discovering distant homologs, strong robustness, and high efficiency.

Single-Cell Data Integration Pipeline

1. Data preprocessing: Standardization and feature selection
2. Gene alignment: Construct a shared gene space
3. Batch correction: Eliminate technical batch effects (scVI, Harmony, etc.)
4. Joint embedding: Learn a unified representation of cell types
5. Downstream analysis: Cell type annotation, differential expression analysis, evolutionary conservation assessment

Technical Advantages and Biological Significance

• Functional-level alignment: Captures functional similarity, enabling identification of homology even with large sequence divergence.
• Handling gene duplication: Distinguishes orthologs from paralogs, providing precise mapping.
• Support for non-model organisms: No need for complete annotations—embeddings can be generated as long as protein sequences are available, lowering research barriers.

Application Scenarios

• Evolutionary developmental biology: Compare embryonic development trajectories and identify conserved programs.
• Disease model research: Map human disease cell states to mouse models and evaluate effectiveness.
• Drug target discovery: Identify cross-species conserved targets and assess translational medicine feasibility.
• Comparative immunology: Study the conservation and diversity of immune cell types.

Limitations and Future Directions

Limitations:

• Currently only supports one-to-one mapping, not handling complex many-to-many relationships.
• Only applicable to protein-coding genes; integration of non-coding RNAs requires other strategies.
• Performance differences between different pLLM models need systematic evaluation.

Future Directions:

• Integrate multiple sequence alignment information.
• Develop fine-tuning strategies for specific species groups.
• Joint modeling with single-cell pre-trained models (e.g., scGPT).

Conclusion

The pLLM-cross-species-integration project applies cutting-edge protein language models to comparative genomics problems, connecting evolutionary biology and deep learning through semantic embedding of biological sequences. With the popularization of single-cell technology and the advancement of pLLMs, such methods will become standard tools for multi-species cell atlases, promoting the understanding of life's diversity and unity.