BayeshERG: a robust, reliable and interpretable deep learning model for predicting hERG channel blockers

Abstract

Unintended inhibition of the human ether-à-go-go-related gene (hERG) ion channel by small molecules leads to severe cardiotoxicity. Thus, hERG channel blockage is a significant concern in the development of new drugs. Several computational models have been developed to predict hERG channel blockage, including deep learning models; however, they lack robustness, reliability and interpretability. Here, we developed a graph-based Bayesian deep learning model for hERG channel blocker prediction, named BayeshERG, which has robust predictive power, high reliability and high resolution of interpretability. First, we applied transfer learning with 300 000 large data in initial pre-training to increase the predictive performance. Second, we implemented a Bayesian neural network with Monte Carlo dropout to calibrate the uncertainty of the prediction. Third, we utilized global multihead attentive pooling to augment the high resolution of structural interpretability for the hERG channel blockers and nonblockers. We conducted both internal and external validations for stringent evaluation; in particular, we benchmarked most of the publicly available hERG channel blocker prediction models. We showed that our proposed model outperformed predictive performance and uncertainty calibration performance. Furthermore, we found that our model learned to focus on the essential substructures of hERG channel blockers via an attention mechanism. Finally, we validated the prediction results of our model by conducting in vitro experiments and confirmed its high validity. In summary, BayeshERG could serve as a versatile tool for discovering hERG channel blockers and helping maximize the possibility of successful drug discovery. The data and source code are available at our GitHub repository (https://github.com/GIST-CSBL/BayeshERG).

Review of machine learning and deep learning models for toxicity prediction

“…Although 1 and 10 µm have been commonly used as the activity thresholds, there is no widely accepted threshold, and multiple threshold settings are often used to change the compositions of the training datasets. Therefore, many ML and DL models, including graph convolutional neural network (GCN) by Chen et al, 52 DNN by Cai et al, 42 hERG-Att by Kim et al, 53 Deep HIT by Ryu et al 43 and BayeshERG as presented by Kim et al, 53 have been reported for the same training dataset. 47,50,51,54 This is one reason that many models (504 models) have been reported for cardiotoxicity prediction.…”

Interpreting Neural Network Models for Toxicity Prediction by Extracting Learned Chemical Features

“…Typical machine learning methods include k-nearest neighbors, support vector machines, random forest, gradient tree boosting, and more recently deep neural networks (DNN) . Chemical fingerprints are widely used as input features, however, a recent focus has been on the use of different types of descriptors including, for example, string representations such as SMILES, , depictions of chemical structures as inputs to DNNs, and 2D and 3D chemical graphs − which have been used with both classical machine learning methods and with more novel graph-based DNN architectures.…”

In silico prediction of hERG blockers using machine learning and deep learning approaches

“…Deep learning is a subset of machine learning that uses multilayer neural networks for mathematical fitting to learn the intrinsic patterns of sample data, showing good performance in feature extraction and prediction (Deng & Yu, 2014). Recently, some deep neural network models for predicting hERG blockers have also been proposed, such as hERG‐Att (Kim & Nam, 2020), DeepHIT (Ryu et al, 2020), and BayeshERG (Kim et al, 2022), and they all exhibit excellent prediction performance.…”