New benchmark exposes persistent weaknesses in AI molecular design tools

Generative artificial intelligence models promise to accelerate drug discovery by designing small molecules directly from protein structures, but a new benchmark suggests the technology still falls short of practical use. Researchers report the creation of MolGenBench, a standardized testing framework for evaluating generative models in structure-based drug design, covering both de novo molecular generation and hit-to-lead optimization.

The benchmark is built on a large, structurally diverse dataset, an intentional departure from earlier evaluations that relied on simplified datasets and computational surrogates. According to the authors, those simplified setups poorly capture the heterogeneity and constraints of real-world drug discovery, potentially overstating how well current models would perform in actual research settings.

In the study, published in Nature Communications, the team evaluated 17 leading generative methods across the benchmark’s tasks. The results, they report, reveal persistent limitations across the field rather than isolated failures of individual models.

Among the weaknesses identified were weak virtual-screening performance, limited coverage of target-specific bioactive space, and the frequent appearance of high-risk structural motifs in generated molecules. The models also showed poor target awareness and reduced generalization when applied to proteins they had not encountered during training.

Not every finding was negative. Some models did improve the generation of three-dimensional molecular conformations, indicating genuine progress in producing geometrically plausible structures. Even so, the authors note that these models still struggle to produce active molecules reliably, a gap between structural plausibility and biological function that remains a central obstacle.

The benchmark’s designers frame the work as a bridge between two parts of the drug discovery pipeline that have often been evaluated separately: protein structure prediction and small-molecule ligand design. Advances in protein structure prediction have created opportunities for structure-based drug design, the paper notes, yet translating that structural knowledge into actionable small-molecule ligands remains difficult.

To address that gap, MolGenBench provides standardized data, metrics, and analyses intended to connect structural prediction with practical molecular design. By applying the same tests to many models under the same conditions, the benchmark aims to give developers a clearer picture of where their methods succeed and where they fail, and to guide the development of more effective structure-based generative models.

The evaluation covered computational performance only; the study reports benchmark results rather than experimental validation of generated molecules in the laboratory. Readers should therefore interpret the findings as a measure of current model capabilities against standardized computational tasks, not as evidence about the clinical or therapeutic value of any generated compound. The work is published as open access, and the authors describe it as a resource intended to support continued development in the field as generative approaches mature.

Subject of Research: A benchmark evaluating generative AI models for structure-based drug discovery and lead optimization

Article Title: Evaluating generative models for structure-based drug discovery from de novo design to lead optimization

Article References: Cao, D., Fan, Z., Yu, J., Chen, M., Jiang, X., Sheng, X., Wang, X., Zeng, C., Luo, X., Teng, D., & Zheng, M. (2026). Evaluating generative models for structure-based drug discovery from de novo design to lead optimization. Nature Communications. https://doi.org/10.1038/s41467-026-78105-8

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78105-8

Keywords: MolGenBench, generative models, structure-based drug design, de novo molecular generation, hit-to-lead optimization, virtual screening, protein structure prediction, benchmark, drug discovery, machine learning, Evaluating, generative