LLMDTA: Improving Cold-Start Prediction in Drug-Target Affinity with Biological LLM

Drug target affinity (DTA) prediction can significantly boost the drug development process. Although deep learning-based models perform well on benchmark datasets, their prediction accuracy drops significantly in cold-start scenarios - i.e., when confront with drugs or proteins that have not appeared in the training set. Limited by training dataset scale, features learned by end-to-end DTA models have less generalization. To this end, we propose a novel method named LLMDTA by transferring the biological Large Language Model to solve the cold-start problems in DTA prediction. Specifically, we use molecular pretrained model Mol2Vec and protein language model ESM2 as feature extractors. To apply these pre-trained features to the downstream DTA prediction smoothly, we apply a 1D-CNN-based Encoder to obtain the independent molecular features. We construct a bilinear attention module to model the interactive molecular features between drugs and proteins. Finally, independent and interactive molecular features are fused for final affinity prediction. The results of warm-start and cold-start experiments on three benchmark datasets show that LLMDTA outperforms the start-of-the-art baselines, especially in novel-protein and novel-pair settings. Furthermore, the case study of epidermal growth factor receptor demonstrates the LLMDTA model can identify the novel relationship between unknown drugs and target proteins. Overall, LLMDTA is a promising tool to predict DTA in reality.