MEM-DTA: A Multi-Encoder Model for Drug-Target Affinity Prediction Using Sequence and Structural Features

Determining drug-target affinity (DTA) is a critical step in the drug discovery process, as accurate prediction of DTA can significantly accelerate drug screening and reduce research costs. The widespread adoption of computational simulation methods has enabled high-throughput data processing and predictive modeling. As a powerful tool, deep learning technology excels at automatically extracting high-dimensional features from multimodal data representing drugs and targets. To address the limitations of feature extraction caused by relying solely on single-modality data, we propose MEM-DTA, a drug-target affinity prediction model based on multimodal features and a multi-encoder architecture. The model employs two specialized encoders for feature extraction: a sequence encoder, designed to capture the topological relationships and global properties within drug SMILES and protein amino acid sequences, and a structure encoder, which incrementally extracts local features from drug Morgan fingerprints and protein secondary structures. Finally, the integration of multimodal features effectively enhances the accuracy of affinity prediction. Experimental results demonstrate that MEM-DTA achieves significant performance improvements across multiple evaluation metrics compared to existing models.