Out-Of-Distribution Test (OOD Test) is a novel dataset split of PDBbind v2020, developed to minimise dataset leakage between the train and test sets. OOD Test was introduced in https://chemrxiv.org/engage/chemrxiv/article-details/6675a38d5101a2ffa8274f62
The split can be found in the "split" column of index_oodtest.csv. OOD Test has the following properties:
- Test set complexes are from the Refined set of PDBbind v2020.
- All ligands in the training data have less than 0.5 Tanimoto similarity to the test set ligands. Tanimoto similarity was calculated using ECFP6 fingerprints.
- All proteins in the training data have less than 50% sequence identity to test set proteins. Sequence identity was calculated with MMseqs2.
- The test set excludes heavy ligands (molecular weight > 1000 Da) and ligands with more than 20 rotatable bonds.
- There is no Pocket Pfam cluster overlap between the training and test datapoints. The Pocket Pfam clusters are defined by Zhu et al. in Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening (JCIM, 2022).
- The test and training sets include 295 and 18,402 complexes respectively.
Maximum similarity of training complexes to test complexes in CASF v2016 (orange) and OOD Test (blue), in terms of ligand Tanimoto similarity (left), protein sequence similarity (middle), and protein-ligand interaction similarity (right). We use Tanimoto similarity of PLEC fingerprints to estimate protein-ligand interaction similarity.
