A Deep-Learning Approach toward Rational Molecular Docking Protocol Selection

Abstract

While a plethora of different protein–ligand docking protocols have been developed over the past twenty years, their performances greatly depend on the provided input protein–ligand pair. In this study, we developed a machine-learning model that uses a combination of convolutional and fully connected neural networks for the task of predicting the performance of several popular docking protocols given a protein structure and a small compound. We also rigorously evaluated the performance of our model using a widely available database of protein–ligand complexes and different types of data splits. We further open-source all code related to this study so that potential users can make informed selections on which protocol is best suited for their particular protein–ligand pair.

Predicting Accurate Lead Structures for Screening Molecular Libraries: A Quantum Crystallographic Approach

“…Table S17: Population distribution of the 30 poses of the ligands of ERβ belonging to 4 types. Tables S18-S27: Interaction energies (kCal•mol −1 ) of ERβ with GoldScore poses of ligand 15,25,27,29,40,44,57,62, 68 and 70, respectively. Table S28: Comparison of ranking based on the experimental IC 50 values, the average IE from KEM-CP and fitting score of the best poses of the ERβ ligands.…”

“…For ERβ, the GoldScore predicts Type 1 pose as the best pose for majority of the ligands (six: 15, 40, 44, 57, 62 and 68), Type 4 as the best pose for three ligands (25, 27 and 29), Type 3 as the best pose for ligand 70 only and none from Type 2. Interestingly, KEM-CP also predicts Type 1 pose as the best pose for six ligands (25,29,40,57,62, and 68), Type 3 as the best pose for three ligands (15,44,70), Type 2 as the best pose for ligand 27 only and none from Type 4. Further, the comparison of RMSDs (Table 7) of these four types of poses of the ligands with the crystal geometry of 4NA suggests that Type 1 (RMSD < 0.4 Å) and Type 3 (RMSD < 1.3 Å) are the correctly predicted poses for these ligands.…”

“…In general, the docking and scoring functions are assessed using these two parameters and seldom based on the binding affinity. Additionally, the employment of deep-learning approach [23,24], the graph matching method [25], followed by the traditional docking process have shown to improve the accuracy of protein-ligand binding mode. However, this approach could be target specific [26].…”

Opportunities and challenges in application of artificial intelligence in pharmacology

“…Also, deep learning (DL) a subset of ML used to improve docking results has been extensively implemented in drug design and development [ 55 , 71 ]. For example, a study developed a DL neural network architecture that the input data were protein voxels and ligand fingerprints, and the output linear data were RMSD min , RMSD ave , and nRMSD by DockBench [ 78 ]. Previously, introduced, DeepVS, based on CNN has been introduced which also achieved good results (the best AUC ROC has ever reported) without human-defined parameters [ 79 ].…”

Drug Discovery for Mycobacterium tuberculosis Using Structure-Based Computer-Aided Drug Design Approach

“…According to the Himar1 transposon mutagenesis study conducted by DeJesus in 2017 [200], the majority of the enzymes targeted by this method are enzymes encoded by crucial genes, with the exclusion of antigens BioA, NarL, 85c, EthR, and LipU. Although this technique assigns nonessentiality to genes based on in vitro growth, it cannot be relied on to determine whether genes are nonessential in vivo [201,202]. For instance, the NarL enzyme is necessary for anaerobic survival throughout infection, while BioA is crucial for biotin synthesis during the latency phase of Mycobacterium TB infection [203,204].…”