Attention Patterns Fail Perturbation Prediction in Single-Cell Models
A Feb. 19, 2026 arXiv preprint presents a systematic evaluation framework (37 analyses, 153 statistical tests) for mechanistic interpretability in single-cell foundation models and applies it to scGPT and Geneformer. The authors find attention patterns encode layer-specific biological structure but add no predictive value for perturbation outcomes, where gene-level baselines achieve AUROC 0.81–0.88 versus 0.70. They propose CSSI to improve GRN recovery up to 1.85×.
Scoring Rationale
Strong methodological contribution with clear negative results; preprint status and single-source limits broader confirmation and replication.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
