gSelformer-MV: Multiview, Subgraph-Augmented Group SELFIES Transformer for Molecular Property Prediction

Data-driven approaches are essential for relating properties to the chemical structure. Atom-focused views of individual compounds are common in molecular representation learning: graph neural networks and chemical language models, the two main algorithm classes, take atomic-level graphs and atom-wise token sequences as input, respectively. However, directly integrating information about functional groups into advanced architectures remains nearly unexplored. To fill this gap, we introduce gSelformer-MV, a transformer that operates on multiple views of Group SELFIES (a SELFIES variant augmented with tokens for functional groups) that enables representation at both the atomic and substructure levels. Unlike prior Group SELFIES approaches that produce a single string per molecule, gSelformer-MV constructs multiple subgraph-partitioned Group SELFIES views and uses them jointly during training and inference. We show that gSelformer-MV is superior in terms of accuracy and explainability to the models trained exclusively on SELFIES strings. Moreover, gSelformer-MV achieves state-of-the-art performance on several regression benchmarks; further gains are obtained when restricting to high-confidence predictions. These results indicate that subgraph augmentation is a simple and effective route for advancing string-based molecular property prediction.