BiScale-GTR: Fragment-Aware Graph Transformers for Multi-Scale Molecular Representation Learning
A novel self-supervised framework for molecular representation, named BiScale-GTR, has been unveiled in a paper on arXiv (ID 2604.06336). This framework aims to identify and leverage recurring molecular fragments while addressing the challenge that a common fragment identity may not adequately represent its behavior in different molecular contexts. The same fragment can behave differently based on its atomic surroundings. BiScale-GTR develops a graph Byte Pair Encoding (graph-BPE) vocabulary through Weisfeiler-Lehman (WL)-based fragment identities, chemical validity checks, and recursive out-of-vocabulary (OOV) decomposition. Each fragment token is anchored in its immediate atomic environment, allowing for representations that are both reusable across various molecules and attuned to local atomic contexts. The paper can be found at https://arxiv.org/abs/2604.06336.
Key facts
- BiScale-GTR is a self-supervised molecular representation framework.
- It uses fragment-level representations to capture recurring molecular substructures.
- The framework addresses the issue that shared fragment identity may not fully describe fragment behavior in different atomic environments.
- It constructs a reusable graph Byte Pair Encoding (graph-BPE) vocabulary.
- The vocabulary uses Weisfeiler-Lehman (WL)-based fragment identity.
- Chemical validity filtering is applied.
- Recursive out-of-vocabulary (OOV) decomposition is used.
- Each shared fragment token is grounded in its local atomic context.
Entities
—