BiScale-GTR combines graph transformers with fragment-aware tokenization to improve molecular property prediction across multiple structural scales

BiScale-GTR addresses limitations of GNN-dominated hybrid architectures by balancing local message passing with global receptive fields using Transformers

The method uses improved graph Byte Pair Encoding (BPE) tokenization to create chemically valid, consistent fragment tokens for molecular representation

Parallel GNN-Transformer architecture enables adaptive multi-scale reasoning to capture molecular patterns spanning different structural granularities

Framework supports self-supervised learning for more effective molecular representations in property prediction tasks