COLDLNA: Enhancing long-range node features extraction to improve robust generalization ability of drug-target binding affinity prediction in cold-start scenarios.

Abstract

Recent advances in deep learning have driven significant progress in drug-target affinity (DTA) prediction. However, many models do not effectively utilize drug molecular graphs or capture long-range protein features, limiting their predictive accuracy. To address these limitations, a novel COLDLNA model is designed for robust DTA prediction. The model employs the Long-range Node Attention Module to refine drug structure representations, while leveraging the Convolutional Attention Module to elucidate critical binding sites by extracting pivotal long-range information from protein amino acid sequences. Compared with the baseline model GraphDTA, COLDLNA reduced the MSE by 12.2% and 11.5% on the Davis and KIBA datasets, respectively. Additionally, its strong generalization ability was further validated on the Human dataset, C. elegans dataset, and in cold-start scenarios.