AI models for drug discovery are trained on curated datasets where measurements are consistent, labeled, and comparable. Real biological measurements are not like that. The same sample can read differently across days, instruments, and labs; cell lines drift; assays vary in sensitivity; and results are often recorded in inconsistent formats. This mismatch is a structural limit, not a temporary one: a model that learned from a tidy table meets a world that does not produce tidy tables.
Can AI Run a Drug Discovery Lab on Its Own?