Stanford's ProteoCraft hits 86.2% pass rate in protein design, 82.9% wet-lab hit rate

机器之心 · wechat · 2026-11-03

Stanford's X-lab and SNAP lab released ProteoCraft, a protein design framework that combines gradient-based proposals, full structure-prediction verification, and two-stage Monte Carlo tree search.

Method

Proposal stage blends sequence priors, geometric constraints and truncated one-step diffusion gradients to control cost;
Verification runs full ESMFold2 predictions on discrete candidates;
Stage A explores multi-residue changes broadly; Stage B refines single positions.

Results

86.2% strict pass rate under independent AF2-IG evaluation vs BindCraft 49.5%, BindCraft v2 58.1%, PXDesign-d 9.5%;
Tree search lifts ESMFold2 strict-criteria pass rate to 54.3%, 2.85× the Best-of-N baseline (19.0%); single-point refinement raises it from 40.5% to 54.3%;
Big config with Anthropic's ESMFold2 inference optimizations gives 3.57× end-to-end speedup with comparable quality;
Wet lab: PD-L1 miniproteins achieved 29/35 binding hits (82.9%), six with Kd<10nM (best 0.879nM); hits span 29 distinct sequence clusters;
IFNAR2 VHH nanobodies hit 44.4%; 94.7% of 75 total designs expressed successfully, plus immunogenicity and off-target screens.

Original post →

Stanford's ProteoCraft hits 86.2% pass rate in protein design, 82.9% wet-lab hit rate

More from Research