Toward Out-of-Domain Binding Affinity Prediction

Predicting binding affinity takes a central role in modern drug design by clarifying how ligands interact with protein targets and supporting more efficient candidate selection. However, many existing methods rely on training and evaluation data drawn from closely aligned domains, which does not adequately capture the diversity observed in real biomedical research settings. In practice, predictive models are frequently challenged by distributional shifts arising from novel chemical scaffolds, previously unseen protein families, differences in experimental conditions and emerging drug resistance, all of which can greatly reduce accuracy. These challenges, known as out-of-domain (OOD) scenarios, increase the risk of overlooking potential compounds or selecting unsuitable substitutes. Although several benchmarks have been proposed to evaluate OOD generalization, many lack accessible and flexible codebases for robust and balanced evaluation. To address this gap, we present a benchmark built upon the DrugOOD dataset, crafted to systematically examine how binding affinity prediction models respond to realistic distributional variations. We also introduce BIOW2D, a new approach that outperforms previous baselines by retaining chemically meaningful substructures as key features, enhancing both accuracy and interpretability. We aim for this study to promote applied research on OOD generalization and aid in advancing reliable binding affinity prediction techniques, especially in contexts where domain shifts cannot be avoided, ultimately supporting more informed compound selection.