Bridging scales between chemical space and behavioral phenotype

Understanding how molecules affect biology is the core challenge of pharmacology and toxicology. Animal behavior, as a central integrative process, provides a powerful readout of the effects of compounds. High-throughput behavioral screening in zebrafish larvae offers a scalable method to study these effects. Yet, current methods face two limitations: (1) by treating compounds as discrete category without chemical context, they struggle to generalize across chemical space; (2) by relying on coarse behavioral measure (e.g aggregate motion), they offer limited insight into the animal physiological states. Here, we demonstrate the benefits of overcoming these limitations.First, using pretrained molecular embeddings to establish a mapping between drug compounds and behavioral phenotypes. We show that this mapping successfully generalizes to new compounds and captures meaningful relationships in chemical space. Second, using a high-resolution behavioral dataset, we demonstrate that enhanced behavioral quantification significantly improves drug effect detection sensitivity. These improvements highlight the value of integrating advanced molecular and behavioral representations to advance drug discovery.