Circuit-Guided Testing of Biology in Single-Cell Foundation Models

Single-cell foundation models (scFMs) are increasingly interpreted as repositories of biological knowledge, but identifying biologically recognizable genes or features does not show that a model computes the corresponding biological process. This distinction is essential if interpretability is to support discovery rather than annotation. We introduce a five-stage circuit-guided framework that specifies an external biological hypothesis, localizes its representation under shortcut controls, contextualizes how it is routed across cellular states, traces how it is constructed from model components, and tests whether those components influence the model under intervention. To enable the tracing and intervention stages in widely used post-normalized architectures, we extend transcoder circuit tracing to post-norm Transformers. We apply the framework to oligodendrocyte maturation in scGPT, using the known transition from progenitor to myelination programs as a stringent within-lineage case study. In the analyzed checkpoint and dataset, the framework reveals a representation that generalizes across donors, pooled evidence of context-dependent routing, cross-token construction of the program, and model-internal intervention effects that reproduce in donors held out from circuit selection. Together, these results provide a controlled route from biologically plausible model features to falsifiable claims about whether and how an scFM has implemented a biological process.