Phi-Former: A Pairwise Hierarchical Approach for Compound-Protein Interaction Prediction

ZHE WANG, Siyuan Liu, Chun Cai, Zijing Liu, Chencheng Xu, Jia Li, Yuan Yao

Keywords: compound-protein interaction, graph transformer, hierarchy graph, biomolecule affinity prediction

Abstract: Predicting compound-protein interactions (CPIs) is critical for AI-aided drug design. Recent deep learning (DL) methods have successfully modeled molecular interactions at the atomic level, achieving both efficiency and accuracy improvements compared to traditional energy-based methods. However, these models do not always align with the chemical realities of CPIs, as molecular fragments (i.e., motifs) often participate in the interactions dominantly. In this paper, we aim to fill this gap by considering the role of motifs for CPIs. We propose a pair-wise hierarchical interaction representation learning (Phi-former) method. Phi-former represents the compound or protein hierarchically and employs a pair-wise specific pre-training framework for modeling the interactions in a more systematic way~(i.e., atom-atom, motif-motif, and atom-motif). We propose an intra-level and inter-level Phi-former pipeline for learning the pair-wise biomolecular graph representation, making learning the different interaction levels mutually beneficial. We demonstrate that Phi-former can achieve superior performance on CPI-related tasks. Furthermore, a case study indicates that our method can accurately identify the specific atoms or motifs activated in CPIs, and thus provide good model explanations that may give insights into molecular structural optimization.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3235