Complex diseases such as cancer, diabetes, and neurodegenerative disorders are driven by multiple interconnected biological mechanisms, limiting the effectiveness of traditional one-drug-one-target strategies. Controlled polypharmacology, in which a single molecule modulates multiple disease-relevant targets in a selective and coordinated manner, has therefore emerged as a promising therapeutic paradigm. However, current computational drug repurposing and discovery approaches remain largely focused on single-target interactions, use biomolecular modalities in limited ways, and do not directly model multi-target interaction patterns, thereby restricting generalizability. In this study, COMPAS (Contrastive Multimodal Polypharmacology-Aware Screener) is proposed as a new machine learning model for directly predicting interactions between small molecules and multiple targets. COMPAS utilizes molecular embeddings obtained from pretrained multimodal models using SMILES and graph representations, while protein embeddings are derived from pretrained multimodal protein language models using sequence, structure, text modalities. These embeddings are projected into a shared latent space through trainable projection layers. A molecule-centered, multi-positive supervised contrastive learning strategy is then applied to organize this space, so that each molecule is positioned closer to its interacting proteins and farther from irrelevant targets. In this way, coordinated multi-target interaction profiles are learned within a unified multimodal setting. The proposed model is developed using data from public resources: ChEMBL, PubChem, ZINC, and UniProt. Alzheimer's disease is adopted as a use case focusing on AChE and BACE-1, and predicted candidates are intended for molecular docking analysis. COMPAS is expected to improve the identification of biologically relevant multi-target drug candidates and support scalable drug discovery for complex diseases.

Co-authors: Tunca Doğan