Simple models outperform single-cell foundation models on downstream benchmarks

Simple, interpretable pipelines using normalization and linear methods achieve state-of-the-art or near state-of-the-art performance on scRNA-seq downstream benchmarks, outperforming foundation models on out-of-distribution novel cell types and organisms.

• Why it matters: Foundation models like TranscriptFormer are computationally intensive, yet it is unclear whether their reported gains reflect essential biological signal or modeling complexity.
• What they did: The authors compare simple normalization plus linear approaches against transformer-based generative gene-expression models that embed genes into a latent space, evaluating multiple standard downstream benchmarks.
• The result: The linear pipelines match or exceed foundation-model performance across tasks including cell-type classification, disease-state prediction, and cross-species learning, and they do better on out-of-distribution settings with unseen cell types and organisms.

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