J Mol Graph Model. 2025 Sep 13;142:109170. doi: 10.1016/j.jmgm.2025.109170. Online ahead of print.
ABSTRACT
Drug-target affinity (DTA) prediction facilitates accelerated drug screening and reduces development costs. To enhance prediction performance and generalization capability, this paper proposes a DTA prediction model based on discrete curvature, named Ricci-GraphDTA, which integrates molecular graph and protein sequence modeling for efficient and accurate DTA prediction. The model consists of three parts: feature encoding, input representation learning, and affinity prediction. In the feature encoding stage, drug molecules are modeled as graphs, where Forman curvature is introduced to adjust the weights of neighbor information aggregation. A GIN residual network is then used to capture the local geometric and topological features of molecules. Protein sequences are modeled using BiLSTM to extract global dependency features, enhanced by an attention mechanism to capture long-range dependencies and key residue interactions-overcoming the limitations of traditional CNNs in handling long-range dependencies. In the input representation learning stage, the high-level representations of drugs and proteins are concatenated and passed through multiple nonlinear transformations to extract cross-modal interaction features, which are then used for affinity prediction. Experimental results demonstrate that Ricci-GraphDTA exhibits significant performance across various evaluation metrics on the Davis and KIBA datasets. Further cold-start experiments demonstrate the strong generalization ability of Ricci-GraphDTA in scenarios involving unseen drugs or targets, highlighting its potential in real-world drug discovery applications. On average, it achieves a 22.5% reduction in MSE across three cold-start tasks, with over 42% reduction in the dual cold-start setting, showcasing excellent structural modeling capability and robustness.
PMID:40966797 | DOI:10.1016/j.jmgm.2025.109170
