Deep learning is competing random forest in computational docking

Computational docking is the core process of computer-aided drug design; it aims at predicting the best orientation and conformation of a small molecule (drug ligand) when bound to a target large receptor molecule (protein) in order to form a stable complex molecule. The docking quality is typically measured by a scoring function: a mathematical predictive model that produces a score representing the binding free energy and hence the stability of the resulting complex molecule. We analyze the performance of both learning techniques on the scoring power (binding affinity prediction), the ranking power (relative ranking prediction), docking power (identifying the native binding poses among computer-generated decoys), and screening power (classifying true binders versus negative binders) using the PDBbind 2013 database. For the scoring and ranking powers, the proposed learning scoring functions depend on a wide range of features (energy terms, pharmacophore, intermolecular) that entirely characterize the protein-ligand complexes (about 108 features); these features are extracted from several docking software available in the literature. For the docking and screening powers, the proposed learning scoring functions depend on the intermolecular features of the RF-Score (36 features) to utilize a larger number of training complexes (relative to the large number of decoys in the test set). For the scoring power, the DL RF scoring function (arithmetic mean between DL and RF scores) achieves Pearson’s correlation coefficient between the predicted and experimentally measured binding affinities of 0.799 versus 0.758 of the RF scoring function. For the ranking power, the DL scoring function ranks the ligands bound to fixed target protein with accuracy 54% for the high-level ranking (correctly ranking the three ligands bound to the same target protein in a cluster) and with accuracy 78% for the low-level ranking (correctly ranking the best ligand only in the cluster) while the RF scoring function achieves (46% and 62%) respectively. For the docking power, the DL RF scoring function has a success rate when the three best-scored ligand binding poses are considered within 2 Å root-mean-square-deviation from the native pose of 36.0% versus 30.2% of the RF scoring function. For the screening power, the DL scoring function has an average enrichment factor and success rate at the top 1% level of (2.69 and 6.45%) respectively versus (1.61 and 4.84%) respectively of the RF scoring function. keywords: Deep learning; Neural networks; Random forest; Drug discovery; Computational docking; Virtual screening.