Research & models Nov 8, 2026

A preprint accepted at NeurIPS 2026 says that of 55 designs judged successful by static checks on the existing protein-design benchmark MotifBench, only 8 passed every check after molecular dynamics simulation.

(paper dated November 5)

The authors' simulations suggest that judging AI protein designs by static structure matching alone may overstate the share of usable designs, and no lab has synthesized the designs yet. In this kind of task, a generative model is given a few functionally critical amino acids and grows a whole protein backbone around them. The authors set the 55 designs moving in simulation: 31 roughly held up, and only 8 passed every check. Compact, self-contained functional sites held up more often, and sites that depend on their surroundings failed more often. The paper has 12 authors, was accepted to the NeurIPS 2026 Datasets and Benchmarks track, and its data and code are public (arXiv 2611.07723, 2026-11-05). How we checked →