MolRL: Self-supervised molecular image representation learning via graph structure bootstrapping

Dongjing Shan

doi:10.1016/j.patcog.2025.112773

MolRL: Self-supervised molecular image representation learning via graph structure bootstrapping

2025-11-22

0

PRE

AI

DOI：10.1016/j.patcog.2025.112773

Original

Original request for help

Save

Abstract

En 中文

• Innovative Molecular Image Representation: This study introduces a groundbreaking approach to molecular data representation by structuring molecular images into graphs. This unique perspective enriches the methodological landscape and offers a fresh dimension to data handling in the field of computational chemistry. • Self-Supervised Fusion Training Paradigm: A novel fusion training paradigm is formulated, seamlessly integrating a graph learner with branches of anchor views and learner views within the robust BYOL self-supervised learning framework. Tailored molecular data augmentation techniques enable comprehensive and accurate molecular representation. • Hierarchical Graph Attention Mechanism: An innovative graph attention mechanism, rooted in weighted graph isomorphism networks, is proposed for processing graph-structured data. This mechanism gracefully adjusts its focus across varying levels of information, from superpixel fragments to functional groups, enhancing model performance and interpretability.

MolRL: Self-supervised molecular image representation learning via graph structure bootstrapping

Abstract

Journal

Organization