ABSTRACT
Protein-protein interactions (PPIs) are essential for many biological processes and diseases. Small molecule modulators of PPIs have emerged as promising drug candidates. However, existing computational methods for identifying PPI modulators rely on similarity to known ones, while neglecting PPI target information and the principles of molecular interaction. Consequently, these methods cannot handle PPI targets with few or no known modulators. To address this challenge, we construct a benchmark dataset that contains PPI targets, modulators, and their interaction relationship. We propose MultiPPIMI, a deep learning framework for predicting the interactions between any PPI-modulator pair. Our model integrates multimodal representations of PPI targets and modulators, and uses a bilinear attention network to capture inter-molecular interactions. We demonstrate that MultiPPIMI achieves an average AUROC of 0.7866 under three cold-start scenarios, and 0.994 under warm-start scenario. Furthermore, we show that combining MultiPPIMI with molecular simulations improves the hit rate for screening inhibitors of Keap1/Nrf2 PPI interactions. This work provides a new way to discover PPI modulators for existing and emerging diseases.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
Title and abstract