IPM-DTI: An Interaction-Pattern-Driven Multimodal Framework for Drug–Target Interaction Prediction via Knowledge Graphs

Accurately predicting drug-target interactions (DTIs) is a foundational challenge in computational drug discovery. While existing multimodal approaches successfully integrate protein sequences and molecular structures, they frequently ignore the latent, high-order interaction patterns that reflect complex biological mechanisms. Knowledge graphs (KGs) provide a structured framework for identifying these indirect associations; however, traversing their large search spaces remains computationally challenging. To address these challenges, we propose a multimodal DTI prediction framework driven by interaction patterns (IPM-DTI). By integrating heterogeneous modal data, including protein sequences, molecular graphs, and explicit KG structures, it can accurately identify underlying interactions through learned interaction patterns. Specifically, an explicit DTI KG is first established from the DTI databases. We then utilize multimodal encoders to map drugs and targets into a unified latent space to extract cross-modal interaction features. To capture heterogeneous interaction patterns, we decompose the global DTI KG into several semantic subspaces based on these features. This enables more focused modeling of local interactions. Furthermore, we incorporate an RL-based reasoning module that maintains a global reasoning horizon. This module exploits a heuristic transition matrix based on relation-type statistics to guide the agent in navigating across subspaces and identifying latent associations. The framework concludes with a multisource fusion module that incorporates structure- and RL-derived results to provide a comprehensive DTI prediction. Experimental results demonstrate that our model outperforms single-modal and single-embedding baselines, validating the effectiveness of the overall framework.