The Development of Target-Specific Machine Learning Models as Scoring Functions for Docking-Based Target Prediction

Abstract

The identification of possible targets for a known bioactive compound is of the utmost importance for drug design and development. Molecular docking is one possible approach for in-silico protein target prediction, whereas a molecule is docked into several different protein structures to identify potential targets. This reverse docking approach is hampered by the limitation of current scoring functions to correctly discriminate between targets and nontargets. In this work, a development of target-specific scoring functions is described that showed improved prediction performances for the correct target prediction of both actives and decoys on three validation data sets. In contrast to pure ligand-based approaches, that are in general faster and include a greater target space, docking-based approaches can cover also unknown chemical space that lies outside the known bioactivity data. These target-specific scoring functions are based on known bioactivity data retrieved from ChEMBL and supervised machine learning approaches. Neural Networks and Support Vector Machines (SVMs) models were trained for 20 different protein targets. Our protein–ligand interaction fingerprint PADIF (Protein Atom Score Contributions Derived Interaction Fingerprint) represents the input for training, whereas the PADIFs are calculated based on docking poses of active and inactive compounds. Different data sets of previously unseen molecules were used for the final evaluation and analysis of the prediction performance of the created models. For a single-target selectivity data set, the correct target model returns in most of the cases the highest probabilities scores for their active molecules and with statistically significant differences from the other targets. These probability scores were also predicted and successfully used to rank the targets for molecules of a multitarget data set with activity data described simultaneously for two, three, and four to seven protein targets.

Molecular Docking: Shifting Paradigms in Drug Discovery

“…Integration of docking with more sophisticated ML-based methods has also very recently been explored for target predictions [138,139]. An example is a study of Nogueira and Koch [139], in which SVM and Neural Networks (NN) models were used to improve the results of reverse docking screenings, pre-processed with the PADIF (Protein Atom Score Contributions Derived Interaction Fingerprint) method [82]. In particular, the authors firstly built datasets of compounds for twenty biological targets with already reported bioactivity data on ChEMBL [140] and X-ray crystal structures.…”

“…Then, for each target, they developed SVM and NN machine learning models able to discriminate active from non-active compounds based on the docking-derived PADIFs. Finally, they retrospectively validated their models, achieving notable prediction performances, both in terms of target ranking and on multi-target selectivity predictions [139]. Although these structure-based approaches might present several advantages over ligand-based methods and standard docking, they are time and computationally demanding.…”

From machine learning to deep learning: Advances in scoring functions for protein–ligand docking

“…In addition to the SFs mentioned above, SVMs were also used in other target-specific methods that were served as postdocking filters, such as PESD-SVM, 108 Pharm-IF, 109 MIEC-SVM, 110 PLEIC-SVM 111 and PADIF-SVM (NN). 112 One common characteristic of these methods is that they all adopt a per-residue decomposition to generate some interaction fingerprint-like features to represent protein-ligand interactions. PESD-SVM refers to the SVM model where the molecular shapes and property distributions on protein and ligand surfaces are featured.…”

Supercomputer Docking

“…The more sophisticated use of deep learning is presented in [49] where RAVE method (the reweighted autoencoded variational Bayes for enhanced sampling) [74] is used for calculating absolute proteinligand binding free energies. Other recent examples of machine learning application to docking and scoring are presented in [58,59,69,99]. In connection with this dawn of new era of machine learning application we should make a remark that databases of experimentally measured protein-ligand binding affinities contain many hidden uncertainties and sometimes errors, and before their use for macine learning, these databases should be strongly and cleverly filtered.…”