Benchmarking single-cell foundation models reveals an advantage in the low-data regime for cell-type and disease-state prediction

Preprint (not peer-reviewed): Benhayon et al. (Institute for Medical Research Israel-Canada, Jerusalem) benchmark cell representations from four single-cell foundation models with logistic regression on 868 disease-prediction tasks (82 datasets) and 65 cell-type datasets from CELLxGENE. With full data, the log-normalized-count baseline outperformed the best model (TranscriptFormer-Sapiens) for cell type and matched it for disease; with few labelled cells, TranscriptFormer-Sapiens significantly outperformed the baseline in both tasks up to about 10^2 samples per class.

Authors / organisation

Benhayon H, Danziger MM, Ben-Zvi D, Rosen-Zvi M.

Venue

bioRxiv (preprint, not peer-reviewed)

Date

2026-10-12

Type

preprint

Year

2026

DOI

10.64898/2026.10.12.758421

Verification

bioRxiv API details (title, 4 authors, posted 2026-10-12, v1, new results, CC BY, corresponding institution IMRIC Jerusalem, abstract) fetched 2026-10-20; full text not read; KB id VC0169 (renumbered 2026-10-20)

Benchmark & validationFoundation modelscRNA-seqZero-shot

Open in the searchable library → · JSON