From Computational Chemistry to Generative Models: A Survey of AI-Driven Small-Molecule Drug Discovery

Abstract

Small-molecule drug discovery requires navigating enormous chemical space to identify candidates that are simultaneously potent, selective, synthetically accessible, and acceptable across pharmacokinetic and safety profiles; modern generative AI extends a long computational medicinal chemistry lineage that began with QSAR, pharmacophore modeling, and early structure-based de novo design. This survey focuses primarily on developments from 2017 to 2025, with the literature updated through 31 January 2026. Using a medicinal-chemistry lens, it first establishes the shared machinery of the field: molecular representations such as SMILES, SELFIES, graphs, fingerprints, three-dimensional coordinates, voxels, and latent vectors; commonly used datasets and benchmarks; and graph neural network encoders, including message-passing, directional, and equivariant architectures. We compare five principal generative mechanisms—autoregressive models, variational autoencoders, generative adversarial networks, normalizing flows, and diffusion models—together with reinforcement learning as a major goal-directed optimization framework. Across these approaches, we show that current methods offer complementary rather than interchangeable trade-offs, and we highlight recurring challenges, including validity versus synthesizability, limited prospective experimental validation, mode collapse and diversity loss, stereochemical and physical realism in three-dimensional design, computational cost, data scarcity, and the emerging convergence toward scaffold-conditioned generation, hybrid architectures, developability-aware optimization, and molecular foundation models.