Where single-cell foundation models help: a benchmark of annotation, perturbation and regulatory inference against simple baselines

Nicheformer achieves 0.98 accuracy for cross-species cell-type annotation from human to mouse lung, outperforming the best baseline at 0.68.

• Why it matters: It is unclear when single-cell foundation models truly add value versus simpler methods, especially for transferring labels across species and for perturbation or regulatory inference.
• What they did: The authors benchmarked five single-cell FMs (including scGPT, Geneformer V1/V2, Nicheformer, scFoundation), CellOracle, supervised perturbation models (STATE), and task-matched baselines on Perturb-seq K562, primary cells, healthy-donor and isogenic leukodystrophy brain organoids, a public organoid CRISPR screen, and lung cancer biopsies.
• The result: Zero-shot in silico perturbations in K562 shift up to 75% of responses in the expected direction (scFoundation, p < 1e-8), and Geneformer V2 recovers patient-level tumor differential expression with fold enrichment 2.9 versus 2.0 for co-expression.

AI summary of the abstract · may contain errors · not medical advice · read the paper.