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

TL;DR

This work tackles compound-protein interaction (CPI) prediction by explicitly incorporating motif-level information through a hierarchical, pairwise representation learning framework called Phi-former. It introduces two encoders for atoms and motifs plus a third encoder that treats motifs as priors, trained via distance-based SSL with three losses $L_V$, $L_M$, and $L_{V|M}$ during pre-training and a downstream fine-tuning stage for CPI tasks. Empirical results on binding-affinity prediction (PDBBind/CASF) show Phi-former achieves strong RMSE and Pearson performance, with ablations confirming the benefit of pre-training and motif-level supervision; a case study on π–π interactions demonstrates chemical consistency and interpretability. Overall, the approach improves interpretability and alignment with chemical principles, offering a path toward more reliable and rational drug design and precision medicine applications, with potential extensions to DDI and PPI in future work.

Abstract

Drug discovery remains time-consuming, labor-intensive, and expensive, often requiring years and substantial investment per drug candidate. Predicting compound-protein interactions (CPIs) is a critical component in this process, enabling the identification of molecular interactions between drug candidates and target proteins. Recent deep learning methods have successfully modeled CPIs at the atomic level, achieving improved efficiency and accuracy over traditional energy-based approaches. However, these models do not always align with chemical realities, as molecular fragments (motifs or functional groups) typically serve as the primary units of biological recognition and binding. In this paper, we propose Phi-former, a pairwise hierarchical interaction representation learning method that addresses this gap by incorporating the biological role of motifs in CPIs. Phi-former represents compounds and proteins hierarchically and employs a pairwise pre-training framework to model interactions systematically across atom-atom, motif-motif, and atom-motif levels, reflecting how biological systems recognize molecular partners. We design intra-level and inter-level learning pipelines that make different interaction levels mutually beneficial. Experimental results demonstrate that Phi-former achieves superior performance on CPI-related tasks. A case study shows that our method accurately identifies specific atoms or motifs activated in CPIs, providing interpretable model explanations. These insights may guide rational drug design and support precision medicine applications.

Figures (5)

• Figure 1: (A) The hierarchical levels of protein and compound. The purple surface represents the protein, while the green ligand denotes the compound. (B) The purple ball-and-stick model illustrates a residue within the protein, showcasing the weak interaction between yellow C and red N atoms guided by the affinity between carbonyl and pyridine functional groups. (C) An erroneous binding result predicting weak interaction between the yellow C and red O atoms when functional group constraints are ignored.
• Figure 2: (a) the pre-training process. In this phase, a complex structure is represented using atom graphs and motif graphs, while the distances between the nodes of the compound and protein are manually masked. (b) the fine-tuning process, during which complete information is provided for some downstream task. The output representations of the atom and motif graphs are then utilized to generate the final prediction.
• Figure 3: Motif graph generation rule
• Figure 4: Graph transformer architecture as encoder
• Figure 5: Case study showing motif constraints enable correct non-covalent interaction modeling.