Fig. 2
From: Applying atomistic neural networks to bias conformer ensembles towards bioactive-like conformations
PDBbind data processing and splitting for modelling. Step 1. The PDBbind dataset is limited to complexes with known Uniprot ID of proteins and ligand names in LigandExpo. Step 2. Only ligands that matches the LigandExpo reference (chirality included) are kept. Step 3. Up to 250 conformers for each unique ligand are generated (a), leading to a dataset of conformers. The ARMSD to the closest bioactive conformation is computed for each conformer (b). Step 4. The dataset containing bioactive and generated conformers is split using a random or scaffold splitting