Single-Cell Foundation Models (scFMs): A 2025 Guide to Applications, Benchmarks, and Future Directions

Read full text Summarise with VIRAT
Authors: AI Chemical Data Analysis Team
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The emergence of single-cell foundation models (scFMs) has revolutionized computational biology by adapting natural language processing techniques to single-cell genomics. By treating individual cells as sentences and genes as words, transformer-based architectures can learn the complex regulatory grammar of cellular systems through large-scale, self-supervised pretraining on millions of transcriptomes. This comprehensive 2025 guide details the architectural variations among leading models, such as the encoder-dominant scBERT, the decoder-centric scGPT, and the network-focused Geneformer, and their respective tokenization strategies, which convert continuous gene expression values into discrete tokens. Despite their sophisticated designs, recent rigorous benchmarking studies reveal critical insights into scFM performance. While these models achieve state-of-the-art results when fine-tuned for specific applications like cell type annotation, batch integration, and drug sensitivity prediction, they frequently underperform simpler, traditional baselines (such as Highly Variable Gene selection or Harmony) in zero-shot settings. This performance gap underscores the necessity of task-specific model selection based on dataset scale, biological complexity, and available computational resources. To address the challenges of evaluating these black box models, the field is rapidly developing biologically grounded assessment frameworks. Novel metrics like scGraph-OntoRWR and the Lowest Common Ancestor Distance (LCAD) move beyond technical clustering scores to measure how well a model's latent representations align with established biological ontologies. Furthermore, standardization initiatives like the BioLLM framework provide unified interfaces for integrating and benchmarking diverse scFMs, promoting reproducibility and fair comparison. Looking forward, the integration of spatial transcriptomics (as seen in Nicheformer), multi-modal data, and advanced interpretability techniques (such as transcoder-based circuit analysis) promises to enhance the utility of scFMs. For researchers and drug development professionals, successfully leveraging scFMs requires balancing their powerful generalization capabilities with rigorous, biology-aware evaluation, ultimately accelerating target discovery and precision medicine. Source: https://www.cellbiotek.com/posts/cellbiotek.com

Fields of Study & Subjects

Single-cell and spatial transcriptomics Gene expression and cancer classification Bioinformatics and Genomic Networks

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
EN