Research and Analysis of Drug Target Interaction (DTI) Prediction Using Self-supervised Multi-channel Network Based on Hypergraph

Abstract

Drug-Target Interaction (DTI) prediction is a core task in drug discovery. The newly proposed HSMI-DTI model is the first to interact hypergraph, multimodal, and self-supervision. On mainstream public datasets, HSMI-DTI simultaneously pushes the Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPR) to over 90%. This method nearly aligns the ROC curve with the optimal inflection point in the upper left corner, and the AUPR also rose simultaneously, establishing a benchmark for contemporaneous performance. HSMI-DTI aims to effectively extract hypergraph features and model cross-modal correlations, utilizing hierarchical self-supervised learning to reveal potential associations between different channels. This framework aims to effectively extract hypergraph features, model cross-modal associations, and mine associations between different channels using hierarchical self-supervised learning. This paper focuses on the most innovative "cross-modal information fusion" mechanism of the HSMI-DTI model, reviews similar methods from 2020 to 2024, and points out that multimodal strategies have become mainstream in exploiting heterogeneous biological relationships; although hypergraph convolution can depict high-order associations, information between channels remains isolated; HSMI-DTI explicitly models inter-channel dependencies for the first time through "attention matching + hierarchical contrast", but it is still limited by the prior design of channels and the bias in negative sample construction.