Unifying molecular structure and cellular morphology to enhance drug-target interaction modeling in cancer

Accurate drug-target interaction (DTI) prediction is critical for accelerating drug discovery and uncovering therapeutic mechanisms. However, the vast combinatorial space of chemical compounds and protein targets, coupled with the complex nonlinear relationships that govern their interactions, presents significant experimental challenges. Existing computational approaches rely heavily on molecular features, often overlooking perturbation-induced morphological phenotypes and lacking a unified representation that connects molecular properties to responses. To address these limitations, we propose a two-stage contrastive learning framework that integrates structural information of drugs and proteins with cellular morphological profiles derived from the Cell Painting high-content screening into a unified multimodal embedding space. In stage one, two modality-specific contrastive learning models were trained independently: a structure-based model that aligns embeddings of drug structures with protein sequences, and an image-based model that aligns morphological embeddings derived from Cell Painting assays of drug and gene knockout perturbations. In stage two, these modality-specific embeddings were further integrated through cross-modal contrastive learning to construct a shared embedding space that jointly encodes drug structures, protein sequences, and their corresponding cellular morphological profiles. Both embedding spaces from stage 1 (structure- and image-based models) effectively aligned annotated drug-target pairs more closely while pushing non-interacting pairs apart. These models yielded correlation differences of 0.4 and 0.3, respectively, as measured by cosine similarity. In the final unified embedding space, coherent clusters emerged among drugs, targets, and their associated morphological embeddings for known DTI pairs, demonstrating successful cross-modal alignment. Among the top-ranking DTI pairs, gedatolisib and PIK3CB showed a similarity of 0.95, consistent with the drug's known activity as a PI3K/mTOR pathway inhibitor. Overall, high-similarity DTI pairs often involved aromatic and heterocyclic compounds whose physicochemical properties closely matched the binding preferences of targets like G protein-coupled receptors (GPCRs) (e.g., CHRM4 ) and kinases (e.g., PIK3CB ). These patterns align with studies showing that GPCR and kinase ligands share characteristic structural features that facilitate binding. In contrast, low-similarity ones involved molecules whose size, polarity, or geometry posed challenges for targets such as metabolic enzymes and heme-binding proteins. Overall, this study highlights the promise of integrating molecular and morphological representations via contrastive learning, providing a powerful framework for advancing DTI modeling and precision cancer therapeutic discovery.