Scientific claims about mechanisms routinely fail to generalize: a drug target validated in one cell line does not transport to another, a neural circuit identified in one model does not replicate in the next, a foundation model benchmark that ranks models in one tissue ranks them at chance in another. The standard response is to test generalization after the fact. This research program asks whether structural features of data and models can predict transport failure before cross-context evaluation is run.
Two questions organize the work:
Mechanism Transport. When a causal mechanism is established in one system, under what conditions does the finding transfer to a new system? Applied to drug transport (drug mechanism transport across cell lines and data modalities) and genetic interactions (higher-order epistasis unpredictable from molecular wiring).
Evaluation Validity. When a metric or benchmark claims to measure a property of a model, does the measurement actually track that property? Applied to RNA/DNA foundation models (structure awareness after nucleotide composition control) and single-cell foundation models (integration metrics at chance on foundation model embeddings). The shared methodology draws on confound-controlled null models, geometric invariants, and causal evaluation design. The evaluation methodology and evidence standards are developed in the Mechanistic Validity framework.