Beyond Replicate Matching: Self-Supervised Behavioral Representations for Phenotypic Drug Discovery

Phenotypic drug discovery provides a functional view of compound activity that is complementary to molecular structure and curated pharmacological annotations. We present Taxicon, a protocol-aware representation-learning framework for cross-compound behavioral geometry in stimulus-elicited zebrafish locomotor assays. A shared stimulus-window CNN and cross-window Transformer are pretrained by masked autoencoding with a plateau-driven masking curriculum and then optimized by metric learning on replicate-as-positive pairs. Mode-of-action (MoA) and activity annotations serve as external probes of cross-compound pharmacological correspondence and do not enter training. Under these external probes, Taxicon improves cross-compound MAP@$R$ by 19--28\% relative to PCA over Motion Index traces on 660 CNS-active compounds and outperforms embeddings from the released supervised Twin-NN checkpoint of~\citet{gendelev_deep_2024} on every cross-drug retrieval metric. The released Twin-NN embeddings do not outperform PCA under these external probes, so strong replicate matching need not yield equally strong correspondence with curated cross-compound pharmacology. As internal checks, learned positional embeddings yield three clusters broadly aligned with stimulus types, and seed-stable attention statistics place three windows consistently among those with the highest between-class variation. The resulting behavioral representation permits measurement of correspondence with known pharmacology and subsequent comparison with structural representations.