Abstract
Drug target discovery is a fundamental aspect of contemporary drug research and development. However, the use of conventional biochemical screening, omics analysis, and related approaches is constrained by substantial technical complexity and significant resource requirements. With the advancement of artificial intelligence-based large language models, notable progress has been achieved in drug target identification. During target mining, large language models with natural language comprehension capabilities can efficiently integrate literature data resources and systematically analyze disease-associated biological pathways and potential targets. Notably, models specifically designed for biomolecular "language" have demonstrated advantages across multiple aspects. The genomics-focused large language model has significantly enhanced the accuracy of pathogenic gene variant identification and gene expression prediction. In transcriptomics, large language models enable comprehensive reconstruction of gene regulatory networks. In proteomics, advancements have been made in protein structure analysis, function prediction, and interaction inference. Additionally, the single-cell multi-omics large language model facilitates data integration across different omics technologies. These technological advancements provide multi-dimensional biological evidence supporting drug target discovery and contribute to a more efficient screening process for candidate targets. The development of these models is generally based on deep neural networks of Transformer architecture, and powerful representation capabilities are obtained through large-scale unsupervised pre-training (such as mask language modeling, autoregressive prediction) combined with task-specific supervised fine-tuning. This review systematically examines recent advancements in the application of large language models in drug target discovery, emphasizing existing technical challenges and potential future research directions.