Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal

Abstract Background : Single-cell foundation models such as scGPT and Geneformer are increasingly used for gene regulatory network (GRN) inference, with attention-derived edge scores routinely interpreted as regulatory proxies. However, whether attention patterns capture causal regulatory relationships—rather than statistical associations already present in expression data—has not been systematically tested. This gap is critical because the NLP interpretability literature has established that attention weights do not reliably indicate feature importance, yet this finding has not been rigorously evaluated in biological foundation models. Results : We present a systematic evaluation framework comprising thirty-seven analyses, 153 statistical tests, four cell types (K562, RPE1, T cells, iPSC neurons), and two perturbation modalities (CRISPRi, CRISPRa). Attention patterns encode layer-specific biological structure—protein–protein interactions in early layers, transcriptional regulation in late layers—but this structure provides no incremental value for perturbation prediction: trivial gene-level baselines outperform both attention and correlation edges (AUROC 0.81–0.88 versus 0.70), pairwise edge scores add zero predictive contribution beyond gene-level features (∆AUROC = −0.0004 to −0.002; 559,720 observations), and causal ablation of regulatory heads produces no degradation across three independent intervention channels. The attention–correlation relationship is context-dependent (equal in K562, worse in CRISPRa, better in RPE1), but gene-level dominance is universal. Cell-State Stratified Interpretability (CSSI) addresses an attention-specific scaling failure, improving GRN recovery up to 1.85×. Conclusions : Attention patterns in single-cell foundation models encode structured biological information but not the causal regulatory signal they are commonly interpreted as capturing. The evaluation framework establishes reusable quality-control standards for the field, and CSSI provides an immediately deployable tool for improved edge recovery from heterogeneous populations.