Comparative assessment of machine-learning scoring functions on PDBbind 2013

Computational docking is the core process of computer-aided drug design (CADD); 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. An effective scoring function should produce promising drug candidates which can then be synthesized and physically screened using high throughput screening (HTS) process. Therefore, the key to CADD is the design of an efficient highly accurate scoring function. Many traditional techniques have been proposed, however, the performance was generally poor. Only in the last few years the application of the machine learning (ML) technology has been applied in the design of scoring functions; and the results have been very promising.In this paper, we propose 12 scoring functions based on a wide range of ML techniques. We analyze the performance of each on the scoring power (binding affinity prediction), 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. We compare our results with the recently published comparative assessment of scoring functions (CASF-2013) of 20 classical scoring functions most of which are implemented in main-stream commercial software. For the scoring and ranking powers, the proposed ML scoring functions depend on a wide range of features (energy terms, pharmacophore, geometrical) that entirely characterize the protein-ligand complexes (about 108 features); these features are extracted from several docking software available in the literature; a feature-space reduction technique, namely, principal component analysis is then applied and the performance is studied accordingly. For the docking and screening powers, the proposed ML scoring functions depend on the geometrical features of the RF-Score (36 features) to utilize a larger number of training complexes (relative to the large number of decoys in the testing set). For the scoring power, the best ML scoring function (RF) achieves a Pearson correlation coefficient between the predicted and experimentally determined binding affinities of 0.704 versus 0.614 achieved by the best classical scoring function ( X-Score HM ). For the ranking power, the best ML scoring function (RF) achieves a Spearman correlation coefficient between the ranks based on the predicted and experimentally determined binding affinities of 0.697 versus 0.626 achieved by the best classical scoring function ( X-Score HM ). For the docking power, the best ML scoring function (BRT) has a success rate in identifying the top best-scored ligand binding pose within 2¿ root-mean-square deviation from the native pose of 13.8% versus 81.0% achieved by the best classical scoring function (ChemPLP@GOLD). For the screening power, the best ML scoring function (SVM) has an average enrichment factor and success rate at the top 1% level of 3.76 and 6.45% versus 19.54 and 60% respectively achieved by the best classical scoring function (GlideScore-SP).