A novel neural network-based nearest neighbor approach for drug function prediction from chemical structures

Abstract

Drug function prediction is a crucial task in drug discovery, design, and development, which involves the prediction of the biological functions of a drug molecule based on its chemical structure. Misleading drug function is a common reason for adverse drug reactions and drug failures. A computational approach can aid in correctly identifying drug functions during clinical testing. Therefore, this study proposes a neural network-based nearest neighbor approach using the Multi-Label Convolutional Neural Network and Nearest Neighbor (MLCNN-NN) method to identify drug functions from chemical 2D structures. This model is built upon the hypothesis that chemical compounds (drugs) with similar molecular structures are likely to exhibit similar drug functions, and the drug functions that occur together are likely to share similar chemical structures. The findings illustrate that the presented models can accurately predict the functions of drugs and outperform the performance of ResNet50, DenseNet201, MobileNetv2, Inceptionv3, VGG19, Graph Convolutional Network (GCN) and Meyer et al. models. The proposed model is evaluated on a benchmark dataset of drug molecules with known functions and achieves the highest accuracy value of 98.10%. Moreover, the identification and visualization of co-occurring drug functions serve as solid indicators of the effectiveness of the proposed model in detecting potential co-occurrence of drug functions, with detection rates of 84.02%. The results demonstrate the effectiveness of the MLCNN-NN model in drug function prediction. They also highlight the potential of a multi-label neural network-based nearest neighbor approach, which utilizes convolutional neural network and nearest-neighbor methods, in drug discovery, design, and development.