A Systematic Literature Review of Single-Cell Foundation Models for Transcriptomics: Design, Adaptation, Benchmarks, and Future Directions
Abstract
Large single-cell atlases have enabled reusable representation learning across tissues, diseases, platforms, and species. Single-cell foundation models (scFMs) build on these resources by learning cell- and gene-level representations from large-scale transcriptomic data. However, current scFMs vary substantially in input encoding, tokenization, backbone architecture, pre-training objective, biological prior knowledge, adaptation strategy, and evaluation protocol. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement, this review synthesizes 166 unique studies represented by 175 eligible reports published between 1 January 2022 and 9 June 2026. It addresses five research questions covering architecture, gene-expression representation, pre-training, downstream adaptation, generalization and robustness. The evidence is organized into four technical lineages: scale- and architecture-driven modeling, knowledge-enhanced modeling, adaptation and representation reuse, spatial, multimodal, multiscale, and contextual extension. The review also distinguishes claims reported in original model studies from evidence provided by independent benchmark and re-evaluation studies. The findings provide stronger support for supervised or fine-tuned annotation-like tasks than for zero-shot embedding reuse, while perturbation-response prediction remains more sensitive to data splitting, overlap between pre-training and test data, comparator strength, and metric choice. Independent benchmarks also provide limited support for treating model scale as a proxy for biological understanding or robust transfer. Study-level appraisal judged 10 studies to be at low risk of bias, 140 to have some concerns, and 16 to be at high risk; 78 studies had strong reproducibility support. Future work should prioritize leakage-resistant out-of-distribution evaluation, stronger biological grounding and interpretability, context-aware modeling, and reproducible and secure deployment.