Application of convolutional neural network and evolutionary algorithm to predict the outcome of adverse drug interactions

The article delves into the complexities surrounding the comprehensive utilization of pharmaceuticals through the lens of computational pharmaceutics. It underscores the significant cost implications associated with diagnosing the outcomes of such interactions using conventional laboratory techniques. As the medication repertoire expands, it becomes increasingly impractical to identify all potential interaction outcomes. The study explores established computational methodologies and highlights the challenge of ensuring the completeness and accuracy of results, stemming from the diverse sets of resulting interactions employed in machine learning. A hypothesis is posited regarding the potential of deep learning technologies to discern the sought-after function from numerous interaction outcomes, thereby enhancing forecasting accuracy. To address this, a solution is proposed, leveraging a combined approach of a convolutional neural network (CNN) and an evolutionary algorithm. The research outlines the development results of the neural network architecture and elucidates its operational principles. Furthermore, it entails the creation of a computational model and the preparation of datasets for training and testing the neural network. Training is conducted, and comparative analysis results of various forecasting modes using a range of known algorithms are obtained. The findings demonstrate that employing the particle swarm optimization algorithm leads to enhanced accuracy, sensitivity, selectivity, and reliability compared to conventional methods on identical datasets. Despite significant advancements in computational pharmaceutics, existing approaches inadequately meet modern requirements for forecast completeness and accuracy, primarily due to the extensive medication repertoire and the multitude of features necessary for training computational models. However, deep learning technologies offer the potential to extract localized features from a plethora of unknowns, thereby augmenting diagnostic accuracy. The proposed method for predicting drug interactions, based on deep learning technology implemented through a CNN, adjusts key network parameters using the particle swarm optimization evolutionary algorithm. The architecture of the proposed CNN minimizes memory requirements by placing weights in the network's channel layers, potentially improving computational efficiency and expediting computation time. In the proposed prognostic model, key variables for predicting drug interaction include the size of the convolution layer window, the subsampling layer window size, and the number of neurons in the hidden layer of the fully connected output neural network.