DRLCI: Dual-view Representation Learning Framework for Predicting Drug-Gene Interactions via Causal Intervention.
Abstract
Investigating drug-gene interactions is essential for advancing personalized medicine and therapeutic innovation. In recent years, Graph Neural Network-based approaches have demonstrated significant potential for this task by utilizing diverse strategies, such as subgraph extraction and hypergraph modeling. However, these associative paradigms remain confined to observational statistics, primarily identifying superficial correlations rather than underlying causal mechanisms. Drug-gene prediction tasks are frequently hindered by hard negative samples acting as confounders, which lead existing methods to overweight biologically irrelevant nodes and introduce substantial noise. Moreover, relying solely on a single topological view restricts the prediction framework to a closed loop of established connections, impeding the discovery of novel interaction patterns. Addressing these deficiencies, we propose a dual-view representation learning framework. Initially, Relation-Aware graph convolution adaptively quantifies the differentiated contributions of neighboring nodes via an attention mechanism to capture key interaction patterns. To refine this procedure, we incorporate a causal intervention mask and a causally-debiased contrastive loss into the feature extraction stage, thereby optimizing the representation space and suppressing confounding effects. Finally, external medical prior knowledge is integrated into the prediction pipeline through a textual modality governed by a gated fusion mechanism. Comprehensive benchmark evaluations demonstrate that our framework achieves state-of-the-art performance on the DrugBank and DGIdb datasets.