AutoRNSRF-DTI: a recommendation model for predicting drug–protein interactions using reliable negative samples

The drug discovery and drug development tasks involve substantial cost in terms of time and money. Hence, an interdisciplinary effort in this direction from biologists, mathematicians, chemists, and medical experts to computer professionals would help in addressing high drug attrition rates while developing drugs. The long journey of bringing the drug from laboratories to the market involves the prediction of drug–target interaction (DTI) as one of the key steps in drug development. Though the DTI prediction problem is tricky and challenging, it acts as the foundation for a wide range of research tasks like predicting side effects of drugs, drug repurposing, and drug discovery. The costly and time-consuming wet lab experiments can be preceded by computational methods for DTI prediction. These methods intend to narrow the search space of the probable drugs for the query target or vice versa. The resultant predictions between drugs and targets can further be validated using the wet lab experiments.

The DTI datasets contain experimentally verified positive drug–target samples representing the interaction data, while the unknown interactions between the drug–target samples are considered as negatives. The performance of the ML models built using such datasets can be improved by finding reliable negatives. Hence, rather than assuming the drug–target samples with unknown interactions as negatives, we may find reliable negatives and then predict the interactions using the new dataset.

Since majority of the druggable targets are the proteins, hence, in this paper, we propose AutoRNSRF-DTI, a recommendation model for predicting drug–protein interactions. An autoencoder has been used to find the reliable negative samples which affects the overall result of the machine learning classifiers. A comparative analysis of the machine learning (ML) classifiers (Random Forest, KNN, Decision Tree, Logistic Regression, Gradient Boosting, and AdaBoost) constructed using standard DTI datasets was done with the respective classifiers which were constructed using autoencoder-based reliable negative samples from the standard DTI datasets.

The results show that the ML classifiers constructed using the reliable negative samples have an increased performance in terms of AUCROC, Average Precision, Geometric mean, and Mathews correlation coefficient as compared to the classifiers constructed using the standard DTI datasets. The results also indicate that Random Forest (RF) has the highest performance of 0.811, 0.814, 0.638, and 0.621 in terms of AUPR/average precision in each of the Enzyme, Ion channel, GPCR, and Nuclear receptor datasets, respectively. RF also ranked highest in terms of AUC ROC values of 0.965, 0.963, and 0.911 for Enzyme, Ion channel, and GPCR datasets, respectively, but KNN recorded the highest AUCROC score of 0.867 for the Nuclear receptor dataset.

We noted that AutoRNSRF-DTI proves to be an efficient recommendation model in terms of AUPR evaluation metric for DTI prediction problem. It filters the reliable negatives from the set of all drug–protein samples marked as non-interacting pairs/negatives and then constructs the classifier models. When compared to the previously published works, our method ranked highest in terms of AUPR in the Enzyme, Ion channel, and GPCR datasets, respectively.