A Length-Adaptive LLM-Enhanced Method for Drug–Target Interaction Prediction

Drug–target interaction (DTI) prediction is central to drug repurposing. Although existing methods integrate multi-dimensional features and have achieved certain progress, they rely heavily on molecular structure or graph structure information. This makes it difficult to capture complex biomedical semantic associations, leading to insufficient semantic information. To address this issue, this paper proposes BioLA-DTI, a large language model (LLM)-enhanced method for drug–target interaction prediction. The method leverages LLMs to provide semantic information for drugs and proteins. Furthermore, existing methods apply uniform weighting strategies to proteins with significantly different lengths. This introduces noise for short proteins and under-represents long ones. To tackle the feature imbalance problem caused by protein-length variation, this work designs a length-adaptive fusion module. This module dynamically weights the LLM-derived semantic features against structural features based on the actual protein-sequence length. Experiments on two benchmark datasets show that BioLA-DTI outperforms existing baseline methods across multiple evaluation metrics, particularly in cold-start scenarios. Case studies demonstrate that this model can provide valuable references for drug discovery. Ablation studies further confirm the critical roles of LLM enhancement and length adaptation modules.