Single-cell foundation models show task-dependent utility for aging biology research
A benchmark across 2.5 million single-cell transcriptomes reveals strengths in detecting rare cell states and gene networks, though simple gene sets predicted age better.
In a computational preprint analyzing more than 2.5 million single-cell transcriptomes, researchers benchmarked ten general-purpose foundation models, three aging-specific models, and conventional methods across five aging-related tasks. Using frozen pretrained representations, Geneformer performed best among foundation models for chronological-age prediction and age–pseudotime concordance. However, a conventional baseline of 2,000 highly variable genes achieved higher mean performance. Several foundation models captured positive molecular-age shifts across three disease contexts. SCimilarity excelled at identifying rare cellular states across out-of-distribution datasets, outperforming both aging-specific models and conventional baselines. Finally, scGPT best recovered reference transcription factor–target interactions, including aging-related regulatory hubs.
Why it matters
Single-cell foundation models can aid longevity researchers in mapping complex regulatory hubs and discovering rare cell types linked to aging. However, their variable performance shows that standard statistical baselines remain preferable for certain tasks like predicting chronological age.
Caveats
This computational study is a preprint that has not yet undergone peer review. Furthermore, model success depended heavily on the specific analytical task, with no single model performing best across all questions.
The paper
Benchmarking single-cell foundation models for aging biology
Ni X, Liang Y, Zhu J et al.
bioRxiv · 27 Sep 2026 · Preprint, not yet peer-reviewed