FAVAR-GAT: A Hybrid Temporal-structural Model for Drug-target Binding Affinity Prediction

The accuracy of drug-target binding affinity prediction is a critical component of modern drug discovery, as it accelerates the efficient identification of compounds with high therapeutic potential. Despite recent advancements, many traditional computational approaches have failed to fully capture the complex temporal dynamics and structural dependencies that control protein- ligand interactions, thereby restricting their predictive accuracy and stability.

To resolve these constraints, we propose FAVAR-GAT a hybrid framework combining Factor-Augmented Vector Autoregression (FAVAR) and Graph Attention Networks (GAT). To model temporal interaction patterns, FAVAR extracts low-dimensional latent factors, while GAT encodes molecular graph structures through adaptive attention. This hybrid architecture captures structural, temporal, and relational features of drug-target interactions.

The hybrid FAVAR-GAT methodology is estimated on various benchmark datasets: Davis, KIBA, BindingDB, and PDBbind. Our experimental results show that FAVAR-GAT outperforms the state-of-the-art baseline models across various evaluation metrics, including Concordance Index (CI), Mean Squared Error (MSE), and R², demonstrating improved predictive accuracy.

The proposed hybrid framework combines latent factor modeling with attention-based graph learning to improve drug-target binding affinity prediction.

The study demonstrates that integrating graph attention mechanisms with latent feature extraction provides a robust and scalable approach for computational drug discovery. The framework shows strong potential for improving drug-target interaction analysis and supporting future AIdriven pharmaceutical research.