Herbal ingredient-target interaction prediction via multi-modal learning

Abstract

The computational prediction of herbal ingredient-target interactions (ITIs) is essential for understanding the mechanisms of action (MoA) of herbal medicine. However, many existing computational methods have yet to fully utilize the multi-modal knowledge of herbs, and the potential noise in literature-mined ITI data has been overlooked. To address these challenges, we propose Multi-ITI, a multi-modal learning framework to learn molecular biological and network topological features for ingredients and targets from multi-modal herbal data, including ingredient SMILES sequences, target protein sequences, ingredient SMILES sequence similarity, target protein sequence similarity, and ingredient-target interactions. Multi-ITI consists of a biological feature learning module and a heterogeneous graph learning module. The biological feature learning module integrates pre-trained models to build deep feature representations for ingredients and targets, while the heterogeneous graph learning module leverages a heterogeneous graph neural network with dynamic attention mechanisms to capture ingredient-target network interactions and mitigate the impact of noisy connections. Experimental results on three public datasets demonstrate that Multi-ITI outperforms six state-of-the-art methods. Additionally, we validate the effectiveness of Multi-ITI through molecular docking simulations and comparisons with recent studies, further highlighting its superior predictive performance and practical applicability.