Generative Artificial Intelligence for De Novo Small-Molecule Drug Discovery: an Innovative Strategy in Modern Drug Development

Abstract

Modern drug discovery is limited by the enormous size of chemical space, the high cost of experimental screening, and the frequent failure of promising candidates during development. Generative artificial intelligence has emerged as an innovative strategy for de novo small-molecule drug discovery because it can learn relationships between molecular structures and desired biological or physicochemical properties and then propose previously unreported molecular structures. Unlike conventional screening, which primarily searches existing chemical libraries, generative approaches can explore new regions of chemical space while applying constraints related to potency, selectivity, physicochemical behavior and synthetic feasibility. The workflow begins with high-quality chemical and biological data, followed by molecular representation, model training, generation of candidate structures, multi-property filtering, synthetic assessment and experimental validation. Recurrent neural networks, variational approaches, generative adversarial approaches, graph-based models and transformer-based architectures have been investigated for molecular generation. Early studies demonstrated the ability of generative systems to produce focused libraries and biologically active compounds, while more recent work has progressed to candidates entering human clinical trials. The discovery and development of rentosertib, a smallmolecule inhibitor of TRAF2- and NCK-interacting kinase for idiopathic pulmonary fibrosis, provides an important translational example of an artificial-intelligence-driven program progressing from target discovery and molecule generation through preclinical evaluation and phase 2a clinical testing. Nevertheless, generative artificial intelligence does not eliminate the need for medicinal chemistry, pharmacology, toxicology and clinical evidence. Data quality, model bias, chemical validity, synthetic accessibility, uncertainty, interpretability, reproducibility and regulatory acceptance remain important challenges. Properly integrated with experimental science and human expertise, generative artificial intelligence can function as a powerful hypothesis-generation and candidateprioritization strategy for modern drug development.