Combining machine learning systems and multiple docking simulation packages to improve docking prediction reliability for network pharmacology

Abstract

Increased availability of bioinformatics resources is creating opportunities for the application of network pharmacology to predict drug effects and toxicity resulting from multi-target interactions. Here we present a high-precision computational prediction approach that combines two elaborately built machine learning systems and multiple molecular docking tools to assess binding potentials of a test compound against proteins involved in a complex molecular network. One of the two machine learning systems is a re-scoring function to evaluate binding modes generated by docking tools. The second is a binding mode selection function to identify the most predictive binding mode. Results from a series of benchmark validations and a case study show that this approach surpasses the prediction reliability of other techniques and that it also identifies either primary or off-targets of kinase inhibitors. Integrating this approach with molecular network maps makes it possible to address drug safety issues by comprehensively investigating network-dependent effects of a drug or drug candidate.

Figure 1. Comparison of prediction accuracy using different docking approaches.

Validation data included the 1300 protein-ligand complexes of PDBbind version 2007. Values were the correlations between calculated docking scores and corresponding experimentally determined binding affinities. Black bars indicate results using default scoring functions equipped with docking tools. Gray bars are those re-scored with external scoring functions (e.g. X-Score and RF-Score) after docking. Red bars represent averages of 25 random test/training partition tests using machine learning systems A + B, and the one with an asterisk is the test using PDBbind version 2012 (2897 complexes) dataset. Error bars  =  ± one s.d. External re-scoring functions improved the correlations compared with the employment of docking simulations alone. The application of machine learning systems A + B was the most effective.

Figure 2. Selectivity scores of 33 kinase inhibitors against 139 kinases.

A comparison was conducted using the screening approach proposed in this study (blue bars; PDB IDs from Table S5) and bioassay results (red bars). The calculation of a predicted selectivity score is “S  =  number of kinases docked with score pKd >5.52/total number of kinases tested”, whereas the experimental selectivity scores is “S  =  number of kinases found to bind with Kd <3 µM/number of kinases tested”. A compound with a lower selectivity score indicates that it actively interacts with a small number of target proteins, implying a lower potential for off-target effects. Trendlines are the 2^nd order polynomial regression functions. In most cases, screening accurately predicted the actual calculated binding constants; however, in some cases, screening predicted significantly higher binding constants than experimental data revealed, while no significant underestimates were observed.

Figure 3. Performance of screening in identifying potential off-targets of 15 high selectivity kinase inhibitors (experimental selectivity score S <0.1).

Off-target proteins are those other than the primary targets that interact with inhibitors with a binding affinity <3 µM (Karaman et al. [30]). Blue bars are off-targets that the screening approach succeeded in finding (docking score >5.52; 25 out of 72 off-targets found). Yellow bars indicate those with a tolerance (docking score >4.52; 7), whereas red bars indicate failure of the screening approach to locate any off-target proteins (46 in total).

Figure 4. Schematic of the signaling network-based screening pipeline.

First, a signaling network is launched by CellDesigner. The identities of proteins involved in the network are retrieved by the CellDesigner plugin API to look up corresponding protein structures in 3D through a protein identity-to-structure mapping system. Second, users submit test compounds for docking simulation. After docking simulation using three docking tools, machine learning system A is then applied to re-score generated binding modes based on features of binding interactions and the test compound's molecular properties, after which, it ranks them. Machine learning system B is subsequently to select a binding mode with the greatest reliability from the three top-score modes. Screening is iterated to assess the next protein until all pathway proteins have been tested. Finally, docking scores are converted into a white-to-red color scale to interpret binding strength, and are projected on the network map for a comprehensive inspection.