CaSBRE: Causality-inspired Semi-supervised Biomedical Relation Extraction

ICLR 2026 Conference Submission22624 Authors

Keywords: Semi-supervised Learning, Biomedical Relation Extraction, Spurious Correlation, Do-calculus

TL;DR: A novel causality-inspired semi-supervised learning method for learning biomedical relations

Abstract: Biomedical interaction relations, such as chemical-protein interactions (CPIs) and gene-disease associations (GDAs), are crucial for advancing drug discovery and clinical treatments. However, the vast diversity of biomedical entities and the limited availability of labeled data pose significant challenges to accurately modeling these interactions using traditional supervised learning approaches. These methods often overfit to spurious feature-label correlations in the scarce labeled relations, leading to poor generalization to unseen biomedical entities. To overcome these challenges, we introduce CaSBRE, a causality-inspired semi-supervised learning framework designed to disentangle and mitigate the impact of such spurious correlations. CaSBRE includes two core components: (i) Feature Disentanglement, which separates causal from spurious features by identifying and exploiting discrepancies between their correlations in labeled and unlabeled data; and (ii) Do-calculus Interaction Inference, which marginalizes the influence of spurious features on relation predictions. Through extensive experiments on CPI and GDA tasks, we demonstrate that CaSBRE substantially outperforms state-of-the-art methods, particularly in generalizing to previously unseen biomedical entities, thereby providing a robust and scalable solution for biomedical relation extraction.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 22624