Phenotypic drug discovery (PDD) focuses on the observable traits or phenotype of cells or organisms in response to drug treatment, rather than relying primarily on specific molecular targets. Drugs discovered through this approach may have better therapeutic relevance, as they are tested in conditions that closely mimic human disease. In this study, we present PhenoModel, a multimodal molecular foundation model developed using our unique dual-space contrastive learning framework. This model effectively connects molecular structures with phenotypic information. PhenoModel is applicable to a range of downstream drug discovery tasks, including molecular property prediction and active molecule screening based on targets, phenotypes, and ligands. Our results demonstrate that PhenoModel outperforms baseline methods in these areas. Building from this model, PhenoScreen is developed to successfully identify several phenotypically bioactive compounds against osteosarcoma and rhabdomyosarcoma cell lines. These findings highlight the versatility of PhenoModel and its potential to accelerate drug discovery by uncovering novel therapeutic pathways and expanding the diversity of viable drug candidates.

Phenotypic drug discovery  /  Cellular morphological profiles  /  Cell painting  /  Artificial intelligence  /  Contrastive learning  /  Molecular representation  /  Chemical space  /  Virtual screening