MLCNNF: A Multi-Label Convolutional Neural Network Framework for Predicting Adverse COVID Drug Reactions From the Chemical Structure
Pranab Das, Dilwar Hussain Mazumder · IEEE Transactions on Computational Biology and Bioinformatics · 2025
COVID-19 is a novel disease that currently has relatively few drugs available for treatment. Adverse COVID Drug Reactions (ACDR) are a major concern in the drug development process. Predicting unfavorable COVID drug reactions has become increasingly important and challenging in seeking effective treatments. Many adverse drug responses go unnoticed during clinical testing, making predicting these undetected or unobserved reactions crucial. Previous research has focused on online surveys, biological factors, molecule descriptors, and MEDLINE article information to predict adverse drug reactions. However, the potential of using chemical 2D drug structure characteristics to predict adverse COVID drug reactions has yet to be investigated. Adverse COVID drug reactions are a multi-label prediction problem. Therefore, this research proposes a Multi-Label Convolutional Neural Network Framework (MLCNNF) with an image augmentation method to address this challenge. The aim is to determine whether adverse COVID drug reactions can be predicted from the chemical 2D structure. The RGB color channel is used to represent the features of the chemical structure, and these features are extracted using Convolution2D and MaxPooling2D layers. The models are validated through experiments on the chemical 2D drug structure and chemical conformer images, employing a leave-one-out cross-validation technique. The proposed MLCNNF model achieves superior results and outperforms the performance of transfer learning models such as DenseNet201, VGG19, ResNet50, MobileNetv2, Inceptionv3, AlexNet, EfficientNetV2B2, and Xception, with an accuracy rate of 99.40 $\pm$ .0344%.