CLAMP: A Mechanistic Probe of Regulatory Structure in Foundation Models under Single-Cell Perturbations

Do single-cell foundation models preserve regulatory structure in a form that supports perturbation prediction? Existing evaluations score these models by predictive accuracy, which cannot distinguish a model that has discarded regulatory structure from one that has not. We invert the problem and develop a mechanistic probe that evaluates regulatory information directly rather than through predictive performance. CLAMP recovers the regulatory network that generates it, clamping each perturbed gene at its observed value and letting the rest relax to equilibrium. Because that network is recovered in gene space, the encoder supplying the loss can be swapped freely, so recovery through a given model measures what that model's learned map preserves, layer by layer. To calibrate that measurement we first fit through the identity encoder, where the recovered network validates against independent ChIP-seq and lineage signatures and predicts held-out two-gene perturbations with no combinatorial training data, confirming that recovery tracks genuine structure. Read layer by layer across scGPT, scPRINT, Stack and Tahoe-x1, the probe splits each model's deficit into regulatory directions its embedding fails to keep separable and directions it keeps but presents non-linearly. Which dominates is set by the pretraining objective, not by scale or data volume: Tahoe-x1, trained by masked reconstruction on a perturbation compendium, collapses the regulatory subspace as completely as scGPT does on observational data. Interventional pretraining data alone does not make regulatory structure recoverable. Code will be released upon acceptance.