Bhat, P.Patil, N.2026-02-032025IEEE Transactions on Computational Biology and Bioinformatics, 2025, 22, 5, pp. 2001-2012https://doi.org/10.1109/TCBBIO.2025.3580202https://idr.nitk.ac.in/handle/123456789/20609Predicting protein interaction sites is crucial for tasks such as constructing protein interaction networks, analyzing protein functions, studying molecular-level pathology, and designing novel drugs. However, the restricted predictive performance of sequence-based computational approaches has led to the rise of structure-oriented approaches. Existing cutting-edge methods mostly focus on the secondary structural features, leaving significant scope for further performance improvement. This study incorporates additional structural features from a tertiary-level perspective to derive composite features using graph and temporal convolutions. A hybrid weighted loss function efficiently handles the class imbalance. A fully connected neural network generates the final predictions. The outlined model is tested on various publicly accessible datasets, showing a substantial improvement in performance over leading models. Comparative analysis with the best models from the literature reports enhancement in the Matthews Correlation Coefficient(MCC) and Area under the precision-recall curve (AUPRC) by 4.8% and 4.1% on the Test_60 dataset, 9.8% and 11.2% on the Test_315 dataset, 10.4% and 11.5% on the Dtestset72 dataset, 12.6% and 13.9% on the PDBtestset164 dataset and 10% and 13.1% on the Test_84 dataset. Finally, the statistical t-test showcases the significance of the proposed model in the protein interaction site prediction task. © 2025 IEEE.ForecastingStatistical testsEvolutionary profileGraph and temporal convolutionPacking densityPerformancePropertyProtein interaction sitesRelative solvent accessibilityResidue flexibilitySolvent accessibilityStructural featureConvolution