PDBbind CleanSplit is a refined training dataset for binding affinity prediction models that is based on PDBbind and has been filtered to reduce redundancy and train-test data leakage into the CASF benchmark datasets. The composition of PDBbind CleanSplit can be found in PDBbind_data/PDBbind_data_split_cleansplit.json.
We provide PyTorch datasets of precomputed interaction graphs for PDBbind CleanSplit, for the complete PDBbind database (v.2020) and for the CASF benchmarks on Zenodo (https://doi.org/10.5281/zenodo.14260171). Each PyTorch dataset is available in five versions containing different combinations of language model embeddings in the graph features.
pytorch_datasets_00AEPL- ChemBERTa-77M includedpytorch_datasets_B0AEPL- ChemBERTa-77M and ankh_base includedpytorch_datasets_06AEPL- ChemBERTa-77M and ESM2-T6 includedpytorch_datasets_B6AEPL- ChemBERTa-77M, ankh_base and ESM2-T6 includedpytorch_datasets_B6AEPL- Ablation: ChemBERTa-77M, ankh_base and ESM2-T6 included, protein nodes deleted from graph
In addition, we provide GEMS models that have been trained on each of these datasets:
model/GATE18e_00AEPL_d0100- No embedding included (trained with GATE18e architecture, neglects ChemBERTa)model/GATE18d_00AEPL_d0100- ChemBERTa-77M includedmodel/GATE18d_B0AEPL_d0600- ChemBERTa-77M and ankh_base includedmodel/GATE18d_06AEPL_d0500- ChemBERTa-77M and ESM2-T6 includedmodel/GATE18d_B6AEPL_d0500- ChemBERTa-77M, ankh_base and ESM2-T6 includedmodel/GATE18d_B6AE0L_d0100- Ablation: ChemBERTa-77M, ankh_base and ESM2-T6 included, protein nodes deleted from graph
For each model, we provide five stdicts corresponding to the models originating from 5-fold cross-validation. Depending on the language model embeddings incorporated, these model showed different performance on benchmark datasets:
