Pose Classification Using Three-Dimensional Atomic Structure-Based Neural Networks Applied to Ion Channel–Ligand Docking

Abstract

The identification of promising lead compounds showing pharmacological activities toward a biological target is essential in early stage drug discovery. With the recent increase in available small-molecule databases, virtual high-throughput screening using physics-based molecular docking has emerged as an essential tool in assisting fast and cost-efficient lead discovery and optimization. However, the best scored docking poses are often suboptimal, resulting in incorrect screening and chemical property calculation. We address the pose classification problem by leveraging data-driven machine learning approaches to identify correct docking poses from AutoDock Vina and Glide screens. To enable effective classification of docking poses, we present two convolutional neural network approaches: a three-dimensional convolutional neural network (3D-CNN) and an attention-based point cloud network (PCN) trained on the PDBbind refined set. We demonstrate the effectiveness of our proposed classifiers on multiple evaluation data sets including the standard PDBbind CASF-2016 benchmark data set and various compound libraries with structurally different protein targets including an ion channel data set extracted from Protein Data Bank (PDB) and an in-house KCa3.1 inhibitor data set. Our experiments show that excluding false positive docking poses using the proposed classifiers improves virtual high-throughput screening to identify novel molecules against each target protein compared to the initial screen based on the docking scores.

Enhancing Docking Accuracy with PECAN2, a 3D Atomic Neural Network Trained without Co-Complex Crystal Structures

“…Therefore, the training time is a significant factor to consider. In our previous study [33], we demonstrated that PCN is approximately 30 times faster than a comparable 3D-CNN. We measured the times for both training and testing per pose.…”

“…For training, we employed the widely used BCE (Binary Cross-Entropy) loss function and utilized the Adam optimizer with a learning rate of 0.001. The training was carried out with a minibatch size of 50 and ran for a total of 50 epochs, a choice guided by considerations from our previous research [33].…”

“…Unlike convolutional neural networks, PCN does not rely on a voxel grid, resulting in fewer model parameters and a significantly faster training process. A detailed description of the architecture of the PCN is given in our previous study [33].…”

“…Our method can be trained using only the protein structures, known binding sites, experimental binding affinity data of ligands, and their docking poses. The primary objective of this pose classifier is similar to our original PECAN model [33]. However, our new proposed method does not require any co-complex structures.…”

Integrated Molecular Modeling and Machine Learning for Drug Design

“…Computationally generated decoy poses and training data sets should be generated carefully to prevent the model from having noncausal bias, an issue where the model learns specific data patterns instead of meaningful ligand–protein interactions. , Training data sets tend to have many positive labels that lack diversity and have few or biased negative labels. ,,, Non- and poor-binders are under-reported, and computationally selected nonbinders should be verified experimentally. , There is a lack of quality and standardized data for some targets. Methods such as data set debiasing and introducing specific models to classify and rank actives from inactives ,− are promising regularization techniques to tackle this problem. Furthermore, it is necessary to test models in fair benchmarking data sets akin to real-world conditions. , …”