When perturbation queries fail: static gene and protein embeddings outperform simulated scFM perturbation responses

Single-cell foundation models (scFMs) are increasingly used to predict transcriptional responses to genetic perturbations, and recent work has shown that regulatory structure can be extracted from a pretrained scFM by querying how its output changes when one gene's expression is computationally altered. We ask whether the features such queries produce carry information about the response actually measured when that gene is knocked down. On two genome-scale CRISPRi Perturb-seq screens, in HepG2 and Jurkat cells, we extract zero-out, virtual-value and gradient-trajectory features from three released scFMs (scGPT, scPRINT and TranscriptFormer-sapiens), and compare them against each model's static gene embedding table and against baselines that use no foundation model at all: ESM2 protein embeddings, control-cell co-expression, gene-level expression statistics, and dimension-matched random embeddings. Every representation passes through the same small translator, which predicts only the perturbation-specific residual left once the response shared across knockdowns is removed, and is scored on whole held-out groups of perturbations with similar measured responses. In both cell lines the static gene tables and the protein embeddings substantially outperform the query-derived features, which sit near co-expression, and fine-tuning scGPT on the perturbation data itself does not change this. Reading the same queries from the hidden states of two of the three models, rather than from their output heads, shows that the perturbed gene's identity is decodable at every depth, and the hidden-state features beat their output-head counterparts, though they still fall short of a static gene table. Under our evaluation, much more of the perturbation-relevant signal accessible to the decoder comes from static gene representations than from the tested simulated responses, suggesting that strong performance on regulatory-edge benchmarks does not, by itself, establish that the same query representation predicts the consequences of a perturbation.