HUBioDataLab/SELFormer

Implements a RoBERTa-based transformer encoder pre-trained on masked language modeling with SELFIES notation—a 100% chemically valid alternative to SMILES—enabling robust molecular representation learning. Provides pre-trained models, fine-tuning capabilities for property prediction tasks (classification/regression), and downloadable molecular embeddings from ChEMBL and MoleculeNet datasets. Outperforms graph-based and SMILES-language approaches on solubility and toxicity prediction benchmarks.

107 stars. No commits in the last 6 months.