Department of Computer Engineering

The discovery and development of effective anticancer therapies face high failure rates, often due to inadequate target efficacy, unintended off-target toxicity, and molecular heterogeneity among tumors. To address these challenges and enhance precision oncology, we introduces an integrated computational framework that combines multi-omics data, prior knowledge networks, and deep learning architectures for predicting drug sensitivity in cancer cell lines. We leverage transcriptomic data to infer transcription factor activities using univariate linear models and incorporate curated biological networks to capture regulatory interactions. These representations are refined using MOON (Meta-fOOtprint aNalysis), ensuring context-specific, biologically consistent network embeddings. On the compound side, we represent drugs using Extended Connectivity Fingerprints (ECFPs), capturing essential structural features in a machine-readable format. A Graph Neural Network (GNN) is employed to learn from these contextually enriched prior knowledge networks, while a feed-forward neural network processes the ECFP vectors. By merging these learned representations, we predict half-maximal inhibitory concentrations (IC50) for various cell line–drug combinations. We benchmark our approach against classical regression and machine learning models, comparing the predictive performance of raw gene expression, transcription factor activities, and MOON-refined representations.