ChemAnchor-DG: Robust Molecular Property Prediction via Chemical Anchor-Guided Domain Generalization
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Accurate molecular property prediction is vital for drug discovery and materials science. However, real-world molecular data inherently suffers from significant challenges, including domain shift from diverse sources and pervasive label noise stemming from experimental inaccuracies. These issues severely compromise the generalization capability and robustness of existing models, leading to unreliable predictions in unseen domains. To address these critical limitations, we propose ChemAnchor-DG, a novel framework designed to learn robust and domain-invariant molecular features. ChemAnchor-DG integrates three core components: Chemical Anchor Generation provides stable, class-specific prior chemical knowledge; Feature Space Anchor Alignment utilizes contrastive learning to guide molecular embeddings towards these anchors; and Adaptive Noise Weighting dynamically adjusts sample contributions based on anchor consistency to mitigate label noise. Extensive experiments on benchmark datasets such as Tox21, Clin-Tox, and HIV, comparing ChemAnchor-DG against state-of-the-art baselines under simulated label noise, consistently demonstrate its superior AUC-ROC performance across all datasets, particularly under challenging conditions involving significant domain shift and label noise. This research offers a promising new direction for developing more robust and generalizable intelligent models for molecular property prediction.
Fields of Study & Subjects
Computational Drug Discovery Methods
Machine Learning in Materials Science
Machine Learning in Bioinformatics
Publication details
- Journal / source
- ChemRxiv
- Publication date
- 31 March 2026
- Language
- EN
- Document type
- Preprint