A Prospective Experimental Assessment of Structure-Based Drug Design: Are We There Yet?

Artificial intelligence (AI) is reshaping drug discovery, particularly through structure-based drug design (SBDD), which promises pocket-aware generation of ligands tailored to a chosen target. Yet, most evidence for the success of generative structure-based drug design comes from retrospective benchmarks and curated case studies, often without compound realization and experimental validation---leaving the question open whether these models generalize to truly novel targets. This gap is especially relevant in antibiotic discovery, where data is sparse and true hits are rare. The key question of this paper is whether off-the-shelf, structure-aware generators can prospectively deliver active molecules for previously unseen antibacterial targets. Here, we show that a standardized, end-to-end evaluation of six state-of-the-art SBDD models yields extremely few experimentally supported hits. Across three structurally selected bacterial targets, >100000 generated designs were reduced to a small set after standardized post-processing, medicinal-chemistry filtering, and commercial analogue mapping. For the most promising target, MurC, 47 prioritized compounds were acquired and tested, and only one showed weak, measurable inhibition at the screening concentration. Together, these results suggest a substantial gap between plausible-looking, protein-conditioned designs and reproducible biochemical activity under realistic discovery constraints, underscoring the need for prospective, assay-closed benchmarks to guide the next generation of SBDD methods.