Recently much effort has been invested into using convolutional neural networks (CNN) models trained on 3D structural images of protein-ligand complexes to distinguish tightly binding from non-binding ligands. However, the dearth of reliable binding structure and affinity data has required the use of constructed datasets for the training and evaluation of CNN molecular recognition models. Here we outline sources of dataset bias and have constructed and performed tests to investigate whether CNN models are properly learning the underlying physics of molecular recognition or are instead learning biases inherent in the datasets themselves. We find that superior enrichment efficiency in CNN models can be attributed to the analogue and decoy bias hidden in the DUD-E datasets used to train and test the model rather than successful generalization of the pattern of protein-ligand interactions. Comparing additional deep learning models trained by PDBbind datasets, we found that their enrichment performances using the DUD-E dataset are not superior to the performance of the docking program AutoDock Vina.

Hidden bias in the dataset leads to misleading performance of deep learning in structure-based virtual screening

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.