Self-Supervised Heterogeneous Graph Model Predicts Drug Target Interactions Without Manual Labels

A computational research team including Xin-Ji Chen, Jian-wei Liu, Chao-dong Tan, and Yuan-long Yue has introduced an adaptive multi-level multi-layer self-supervised contrastive learning model, designated as MMCDTI, designed to predict drug–target interactions across heterogeneous graphs. Reporting in the journal Applied Intelligence, the authors demonstrated that the proposed architecture successfully verifies predictive performance against state-of-the-art baselines. By leveraging representations of candidate compounds and proteins, the model addresses interaction prediction, a computational task that has become increasingly critical for both novel drug discovery pipelines and drug repositioning initiatives in bioinformatics.

However, identifying these interactions through conventional laboratory experiments involves substantial logistical and financial costs. As a result, researchers are increasingly turning to computer-based models for simulation experiments.

Despite the growth of in silico screening, most existing computational models rely predominantly on supervised and semi-supervised learning techniques. These conventional methodologies require substantial quantities of high-quality labeled interaction data to train predictive frameworks effectively. Within the domain of bioinformatics, obtaining labeled data can sometimes be challenging.

To overcome data annotation bottlenecks, complex drug and protein interaction networks can be abstracted into heterogeneous graphs containing diverse node and relationship types. Heterogeneous graph neural networks have emerged as effective tools for processing this structured data, drawing upon architectural advances such as the Heterogeneous Graph Transformer introduced by Hu and colleagues, which employs transformer-based mechanisms to model graph heterogeneity. Building upon these principles alongside the graph contrastive learning paradigm with augmentations proposed by You and colleagues, the researchers developed MMCDTI to infer interactions without relying on manual annotations.

At the operational level, the MMCDTI architecture first constructs a drug–target heterogeneous graph and encodes its structural features using a hierarchical attention mechanism that integrates both node-level and metapath-level attention. Rather than adopting standard graph augmentation strategies—such as random node perturbation, edge alteration, or feature masking—the framework generates two complementary views consisting of a topology graph and a semantic graph. This dual-view contrastive mechanism enables the system to capture richer structural and semantic relationships across the heterogeneous graph without requiring human-generated training labels.

To capture both localized node information and global graph structure, the model incorporates a multi-level, multi-layer self-supervised contrastive framework utilizing feature outputs from both intermediate and final layers. An integrated attention mechanism dynamically balances local and global loss functions during optimization. Through this adaptive multi-layer design, the authors verified that MMCDTI achieves competitive predictive performance compared to established baseline models, offering an automated graph representation strategy that operates independently of labor-intensive labeled datasets.

Subject of Research: Drug–target interaction prediction using self-supervised contrastive learning on heterogeneous graphs

Article Title: Adaptive multi-level multi-layer self-supervised contrastive learning for drug–target heterogeneous graph interaction prediction

Article References: Chen, X.-J., Liu, J.-W., Tan, C.-D., & Yue, Y.-L. (2026). Adaptive multi-level multi-layer self-supervised contrastive learning for drug–target heterogeneous graph interaction prediction. Applied Intelligence, 56(16), Article 557. https://doi.org/10.1007/s10489-026-07591-6

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07591-6

Keywords: drug-target interaction prediction, MMCDTI, heterogeneous graph neural networks, self-supervised contrastive learning, hierarchical attention, metapath-level attention, node-level attention, topology graph, semantic graph, drug discovery, drug repositioning, bioinformatics, Heterogeneous Graph Transformer, graph representation learning, multi-level contrastive learning, multi-layer contrastive learning, loss function balancing