Critical Assessment of Artificial Intelligence Methods for Prediction of hERG Channel Inhibition in the “Big Data” Era

Abstract

The rise of novel artificial intelligence methods necessitates a comparison of this wave of new approaches with classical machine learning for a typical drug discovery project. Inhibition of the potassium ion channel, whose alpha subunit is encoded by human Ether-à-go-go-Related Gene (hERG), leads to prolonged QT interval of the cardiac action potential and is a significant safety pharmacology target for the development of new medicines. Several computational approaches have been employed to develop prediction models for assessment of hERG liabilities of small molecules including recent work using deep learning methods. Here we perform a comprehensive comparison of prediction models based on classical (random forests and gradient boosting) and modern (deep neural networks and recurrent neural networks) artificial intelligence methods. The training set (~9000 compounds) was compiled by integrating hERG bioactivity data from ChEMBL database with experimental data generated from an in-house, high-throughput thallium flux assay.We utilized different molecular descriptors including the latent descriptors, which are real-valued continuous vectors derived from chemical autoencoders trained on a large chemical space (> 1.5 million compounds).The models were prospectively validated on ~840 in-house compounds screened in the same thallium flux assay. The deep neural networks performed significantly better than the classical methods with the latent descriptors. The recurrent neural networks that operate on SMILES provided highest model sensitivity. The best models were merged into a consensus model that offered superior performance compared to reference models from academic and commercial domains. Further, we shed light on the potential of artificial intelligence methods to exploit the chemistry big data and generate novel chemical representations useful in predictive modeling and tailoring new chemical space. File list (2) download file view on ChemRxiv Critical Assessment of Artificial Intelligence Methods for Pr... (1.91 MiB) download file view on ChemRxiv Supporting Information.docx (495.71 KiB)

Machine learning models for classification tasks related to drug safety

“…Here, we have collected 15 works from the past five years that employ machine learning-based classification approaches to predict hERG inhibition [38][39][40][41][42][43][44][45][46][47][48][49][50][51]. All of these works apply training datasets of more than 1,000 molecules (and up to tens of thousands in some cases [47,48]), and an overall majority presents two-class (active vs. inactive) classification (with the notable example of the 2015 study of Braga et al, who have introduced a third class of "weak blockers") [38].…”

CRNNTL: Convolutional Recurrent Neural Network and Transfer Learning for QSAR Modeling in Organic Drug and Material Discovery

“…To demonstrate the versatility of our strategy, two latent representations were generated from one AAE [ 40 ] and another VAE [ 9 ] (named DDC). Then, their QSAR performance were studied based on CRNN and other baseline methods as well as transfer learning for the modeling improvement on a small dataset.…”

“…Due to a prior distribution of the training, AAEs facilitate the generation of novel structures. In this study, two VAEs [ 6 , 9 ] and one AAE [ 40 ] were used for latent representation generation. Their description is provided in Supplementary Materials .…”

High-Throughput Chemical Screening and Structure-Based Models to Predict hERG Inhibition

“…The accuracy of the best performing ensemble model on the external set ( n = 407) was found to be 0.79. Using the benchmark models developed in [ 47 ], models are available and performed better on the dataset built using the assays AID588834 with an MCC equal to 0.87. However, this set is fully included in its training set.…”

“…To further develop and test the QSAR models, the Tox21 dataset was enriched using the ChEMBL database (Version 27) [ 46 ], processed using a similar approach as that applied in [ 47 ], Figure 1 . First, activities on the hERG target (ChEMBL240) were extracted from the database.…”