Modern drug discovery primarily concentrates on identifying and understanding drug-target interactions. Traditional techniques, limited by factors like throughput, precision, and cost, struggle to efficiently identify these potential drug-target interactions. Hence, there's a pressing need for advanced computational methods to verify these drug-target relationships.
We constructed a deep learning model for drug-target interaction prediction. The features of target proteins were extracted and associated with drug molecular substructure fingerprints to form feature vectors of drug-target pairs. The features of compounds and target proteins were subsequently compressed into a unified vector space using sparse principal component analysis. Finally, we used a long short-term memory (LSTM) neural network to make predictions. Five-fold cross-validation was employed to evaluate the performance of our model.
Upon evaluation, our model showcased satisfactory performance in drug-target interaction prediction. Specifically, it achieved accuracies of 86.7%, 84.7%, and 73.4%. These scores were obtained from three different drug-target datasets, highlighting the model's robustness and generalizability. The slight variation in accuracy scores across datasets suggests that, while the model is highly effective, there might still be room for further optimization, particularly for datasets with unique characteristics. These findings indicate that the method is competitive with other contemporary drug-target prediction tools.