Generative AI Drives a Paradigm Shift in Drug Discovery: An Analysis of Technical Approaches, Application Cases, and Challenges

Traditional drug research and development has long been characterised by long development cycles, high investment costs and a high attrition rate among candidate compounds, conventional experimental methods for screening candidate compounds are of limited efficiency when faced with the vast chemical space and complex biological mechanisms. Generative artificial intelligence can provide a new technological approach to new drug research and development by generating conditions based on the distribution patterns of molecular structures and their properties, and by enabling de novo molecular design. This paper systematically reviews the main generative models—including RNNs, Transformers, VAEs, GANs, normalising flows and diffusion models—and their technical characteristics. It analyses their applications in drug target discovery, protein structure prediction, candidate compound generation and lead optimisation, with a particular focus on how generative AI, when combined with active learning, computational physics and feedback from wet-lab experiments, reshapes the drug discovery and development process. Research indicates that generative AI, when combined with active learning, physical computing and feedback from wet-lab experiments, is driving a shift in drug discovery from the traditional linear open-loop model towards a data-driven iterative model of 'generation–evaluation–feedback–regeneration–experimental validation'. By examining the technical characteristics of various generative models and their suitability for drug discovery tasks, this paper provides guidance on selecting generative AI technologies, identifying suitable application scenarios and integrating them into drug discovery workflows.