📝 Abstract:
Computational prediction of drug–target interactions is central to modern drug discovery, enabling the prioritization of candidate compounds before costly experimental validation. Deep learning methods, particularly graph neural networks (GNNs), have substantially advanced this task. However, reported performance often overestimates real-world utility because evaluation protocols frequently allow chemical and biological information leakage between training and test sets. This thesis investigates the generalization capabilities of deep learning models for drug–target interaction and affinity prediction, with a focus on robustness under distribution shift.First, this thesis presents an open-source toolkit for leakage-aware drug–target affinity modeling. The framework automates target-specific dataset construction from ChEMBL, implements scaffold- and temporal-splitting strategies, performs automated leakage auditing, and supports model training and latent-space analysis through a unified Python library, command-line interface, and no-code web interface.Second, a controlled benchmark study evaluates six GNN message-passing operators across eight protein families under three increasingly challenging data-splitting strategies. By holding the remaining architecture fixed, the study demonstrates that distribution shift—not encoder design—is the dominant limitation in current drug–target prediction systems. Predictive performance decreases significantly when compounds and targets are genuinely novel, while probabilistic calibration deteriorates even more rapidly.Third, this thesis introduces a domain-adaptation framework that combines multimodal molecular and protein representations with multi-level distribution alignment and gradient-aware adaptive loss weighting, improving cross-domain predictive performance and calibration relative to state-of-the-art baselines.Together, these contributions advance more reliable and generalizable deep learning systems for computational drug discovery.