Reliable Drug–Target Interaction Prediction Using Convolutional Neural Networks with Robust Negative Sample Generation

Abstract

Proteins, including receptors, enzymes, and ion channels, represent primary biological targets whose interactions with small-molecule drugs play a critical role in therapeutic discovery and development. Accurate identification of drug–target interactions (DTIs) remains a fundamental challenge in drug discovery due to the high cost, time requirements, and scalability limitations of experimental validation. Consequently, computational approaches have emerged as efficient alternatives for large-scale DTI prediction. This study proposes a convolutional neural network (CNN)–based framework for predicting drug–target interactions, with a particular focus on reliable negative sample generation to address the inherent data imbalance and uncertainty present in DTI datasets. The proposed method incorporates feature projection techniques to effectively capture meaningful representations of drug and protein features while reducing noise and redundancy. By constructing more reliable negative instances, the framework improves model robustness and mitigates bias commonly introduced by randomly generated negative samples. The proposed model is evaluated on a benchmark DTI dataset, where it achieves a classification accuracy of 0.9800, demonstrating strong predictive capability. To further assess generalization performance, the model is tested on an independent external dataset derived from DrugBank. On this dataset, the framework attains an accuracy of 0.8814 and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.9527, indicating effective transferability across datasets. Experimental results confirm that the integration of CNN-based feature learning with reliable negative instance generation significantly enhances DTI prediction performance. The proposed framework offers a robust and generalizable computational tool for drug–target interaction prediction and has the potential to support early-stage drug discovery by reducing experimental search space and accelerating candidate prioritization.