FragBERTa: A Fragment-based Molecular Representation Learning Model with Sequential Attachment-based Fragment Embeddings - Models and Datasets

This repository contains the datasets and trained model checkpoints used in the FragBERTa study on fragment-aware molecular representation learning. Datasets pubchem-chembl-zinc-ft__safe___mw-lt-1000__11p08m.parquet: SAFE-encoded molecular sequences used for self-supervised pretraining of FragBERTa. The dataset is constructed from compounds curated from public chemical databases including PubChem, ChEMBL, ZINC, and MoleculeNet after canonicalization, deduplication, SAFE conversion, and molecular weight filtering. finetuning_data.zip: Processed datasets used for downstream evaluation on MoleculeNet benchmarks, including ESOL, FreeSolv, Lipophilicity, PDBBind, BACE, BBBP, HIV, Tox21, and SIDER. Each dataset contains canonical SMILES, SAFE representations, and the corresponding target values. Models pretrained.zip: Pretrained FragBERTa checkpoint. finetuned.zip: Finetuned model checkpoints for individual downstream tasks. Each checkpoint corresponds to the best-performing configuration obtained with hyperparameter optimization. These resources are provided to facilitate reproducibility and further research on fragment-aware molecular representation learning and chemical language models.