ColdstartCPI: Induced-fit theory-guided DTI predictive model with improved generalization performance
Qichang Zhao, Haochen Zhao, Linyuan Guo, Kai Zheng, Yajie Li, Qiao Ling, Jing Tang, Yaohang Li, Jianxin Wang
Abstract
Predicting compound-protein interactions (CPIs) plays a crucial role in drug discovery. Traditional methods, based on the key-lock theory and rigid docking, often fail with novel compounds and proteins due to their inability to account for molecular flexibility and the high sparsity of CPI data. Here, we introduce ColdstartCPI, a framework inspired by induced-fit theory, which leverages unsupervised pre-training features and a Transformer module to learn both compound and protein characteristics. ColdstartCPI treats proteins and compounds as flexible molecules during inference, aligning with biological insights. It outperforms state-of-the-art sequence-based models, particularly for unseen compounds and proteins, and shows strong generalization capability compared to structure-based methods in virtual screening. ColdstartCPI also excels in sparse and low-similarity data conditions, demonstrating its potential in data-limited settings. Our results are validated through literature search, molecular docking, and binding free energy calculations. Overall, ColdstartCPI offers a perspective on sequence-based drug design, presenting a promising tool for drug discovery. Existing models for predicting compound-protein interactions (CPIs) struggle with novel molecules and sparse data. Here, authors introduce a framework that combines pre-trained feature extraction with transformer dynamic tuning to improve CPI prediction.