Cyril Pillai

Background.  Generative artificial intelligence (AI) is increasingly used for de novo design of small-molecule drug candidates, with the promise of accelerating early discovery and therapeutic targets^1,2. However, many studies still define success using computational metrics such as novelty, validity, docking scores, and predicted drug-likeness, which do not directly establish biological or clinical relevance^1,3. As AI-generated compounds begin to advance into wet-lab studies, animal models, and early clinical trials, a clearer standard is needed for describing what counts as meaningful validation^2,4. This study examines recent primary studies and proposes a staged framework for classifying validation strength in de novo AI drug discovery.

Methods.  A literature search was conducted using PubMed, Google Scholar, and the Texas A&M Library databases. Search terms included “generative AI,” “de novo drug design,” “deep generative models,” and “AI-enabled clinical trial.” Primary studies were selected if they used generative methods for small-molecule therapeutics and included prospective experimental, preclinical, or clinical evaluation of AI-generated compounds^1-4. Four recent studies spanning different points along the discovery-to-clinic continuum were selected for review. For each study, the evidence treated as successful validation was synthesized and mapped to a proposed validation tier.

Results.  Validation endpoints across the four studies ranged from computational and biochemical evidence to animal and human clinical data^1-4. DrugGEN combined generative design with docking and molecular dynamics, representing strong computational validity¹. DRAGONFLY generated compounds that were synthesized and evaluated using biochemical assays, selectivity profiling, and ADME testing, representing experimental validation beyond computational benchmark-only performance². Liu et al. used deep transfer learning to generate compounds for colorectal cancer and Alzheimer’s disease and demonstrated activity in mouse models, corresponding to preclinical translational validation³. Finally, Rentosertib, an AI-enabled candidate for idiopathic pulmonary fibrosis, advanced to a randomized, placebo-controlled phase 2a trial, providing safety, pharmacokinetic, and lung-function data in humans and representing the strongest level of validation⁴.

Conclusions.  Generative AI drug discovery now extends beyond algorithmic plausibility to practical human clinical evidence^1-4. The major gap is not the absence of downstream validation stages, but inconsistency in how validation strength is described across studies^1-4. A four-tier framework—computational, experimental, preclinical, and clinical—may provide a practical vocabulary for comparing evidence behind AI-designed drug candidates and guiding future translational research^1-4.

Works Cited

1. Wang S, Li Y, Zhang J, et al. DrugGen enhances drug discovery with large language models and structure-based optimization. Sci Rep. 2025;15(1):98629. doi:10.1038/s41598-025-98629-1.
2. Atz K, Cotos L, Grisoni F, Schneider G. Prospective de novo drug design with deep interactome learning. Nat Commun. 2024;15(1):3408. doi:10.1038/s41467-024-47613-w.
3. Liu Y, Zhang H, Chen Q, et al. De novo generation and identification of novel compounds with drug efficacy based on machine learning. Adv Sci. 2024;11(10):e2307245. doi:10.1002/advs.202307245.
4. Wang L, et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med. 2025;31(8):2602‑2610. doi:10.1038/s41591-025-03743-2.