Abstract Predicting drug–target interactions (DTIs) is crucial for modern drug discovery. However, existing machine learning models have significant limitations: they are typically designed for a single task—either predicting binding affinity or docking pose—leading to excellent performance on one metric but limited practical utility. These models also often struggle with generalizability to novel molecules and proteins due to their reliance on small, labeled datasets. Furthermore, they frequently ignore the essential information contained within the 3D structure of proteins and molecules. To overcome these challenges, we introduce DTAP, a unified framework that simultaneously predicts both the quality of docking poses and drug–target binding affinity. To boost its generalizability, DTAP leverages pretrained large models to learn rich, contextual representations of drugs and targets from extensive unlabeled data. The framework also directly incorporates 3D structural data from both molecules and proteins, using two graph transformers to learn their joint representations. A shared latent vector and task-specific decoders enable crucial cross-task knowledge transfer, allowing the model to learn from the interconnected nature of these two properties. DTAP significantly outperforms state-of-the-art methods on both tasks, demonstrating superior performance especially in cold start situations where data are scarce. Our interpretability analysis on the model’s attention mechanisms confirms its ability to effectively focus on key binding sites. All results indicate that DTAP is a valuable and practical tool for accurately predicting drug–target affinities and docking poses.
Liu et al. (2026) studied this question.