Abstract:The accurate prediction of drug-target binding affinity (DTA) is crucial for drug development and virtual screening. However, existing deep learning methods generally have problems including insufficient utilization of protein 3D structures, limited mining of similar sequence information, and a lack of model generalization ability. To this end, this study proposes DTA prediction model FusionLM-DTA based on the language model and multi-scale feature fusion. This model initializes features based on the protein 3D structures predicted by AlphaFold, thereby enhancing its ability to represent protein spatial conformation. Meanwhile, it extracts high-dimensional semantic features of proteins and drugs by adopting pre-trained molecular sequence models ESM-2 and ChemBERTa, and further captures bidirectional sequence dependency and potential associations between similar sequences via the external attention-enhanced Bi-Mamba module. Furthermore, the designed multi-scale gated linear attention mechanism further enhances the modeling ability of sequence features. Experimental results show that FusionLM-DTA significantly outperforms baseline methods on multiple evaluation metrics on Davis and KIBA datasets, thereby validating the effectiveness and superior performance of the model.