Abstract:
Drug-drug interactions (DDis) are one of the crucial concerns in pharmaceutical research. In the past decade, many researchers developed some machine learning based methods, but these ML methods focus on whether two drugs interact with each other not. The research in DDis domain shows us that different subsequent even could be caused by DDis which can be adverse effects or slowing the recovery process of a patient which is consuming multiple drugs at a time so for investigating the hidden mechanism behind the usage of multiple drugs at a time the prediction of drug-associated events can be more useful. Drug-to-drug interaction occurs when a patient consumes more than one drug at a time. Hence, due to using di afferent drugs, any drug can influence the effect of another drug. The drug-to-drug interactions (i.e., DDI) are detected or identified using the pathways and enzymes interactions, therefore machine learning and deep learning techniques are used to find the DDI with each other. The deep learning models i.e., CNN, LSTM's, and RNN are used to analyze the DDI based on the 65 different types of drug interaction and its associated events using the selected database. The inputs used, in our model, out of the 65 types of drugs are smiles of drugs, enzymes, pathways of the drug to target, and the target. Therefore, the different number of layers, activation function, and features of drugs for the multi-model CNN, RNN, and LSTM's is used to achieve better accuracy, as compared to traditional prediction algorithms. We have done different experiments in terms of using different numbers of layers, activation functions, and different features of drugs the multi-model CNN model achieved an accuracy of0.9000, Fl-score of0.8286, AUPR of 0.9478, AUC of 0.9981 the multi-modal LSTM's models achieved an accuracy of 0.8902, Fl-score of 0.7792, AUPR of 0.9407 and AUC of 0.9978 and the multi-modal RNN's model achieved an accuracy of 0.8866, Fl-score of 0.7779, AUPR of 0.9395 and AUC of 0.9979. The various computational experiments show that a combination of various drug features is performing better than one separate feature of drugs. Compared with other proposed methods like DDIMDL, DeepDDI, CNN-DDI, RF (Random Forest), KNN (K-nearest neighbor), LR (Linear Regression) our multi-modal CNN and LSTM's model method has better performance as compared to the other proposed methods mention above. While the RNN model results are better than LR, KNN, RF methods but approximately equal DDIMDL methods