ProteoCraft | Yanjie Huang 黄䶮杰

ProteoCraft method overview — Method overview: gradient proposals suggest local sequence edits, while full forward AF3-like verification ranks discrete protein designs.

ProteoCraft application targets — Applications: cellular receptors, common allergens, multi-domain nucleases, TCR binders, and TCR CDR redesign.

ProteoCraft benchmark summary table — Benchmark summary: Rosetta filter pass rates across Type I de novo binder and TCR binder design tasks.

ProteoCraft detailed benchmark table — Detailed benchmark: target-level results for standard non-TCR binder targets and TCR-facing binder targets.

ProteoCraft is a protein foundation-model project for design under frozen diffusion-based structure predictors. It treats protein design as a proposal-verification search problem: truncated diffusion gradients are useful for proposing candidate mutations, but reliable values should come from complete forward prediction on discrete sequences.

The framework is built around a hard-forward / soft-backward sequence bridge, gradient-prior Monte Carlo Tree Search, and region-localized masked losses. This lets the system search over mutable protein regions while keeping the broader molecular context fixed, which is important for binder design, TCR-facing interfaces, and localized CDR or TCR redesign.

What It Enables

Protein sequence design with AF3-like diffusion structure predictors without training a new model.
Local mutation proposals from truncated diffusion gradients while preserving legal all-atom topology.
Full-forward verification of candidate sequences before ranking.
Search over de novo binders, immune-relevant TCR targets, and localized protein redesign tasks.
Region-aware losses that optimize designated mutable regions while respecting fixed structural context.

My Role

I am a co-first author and worked on inference-time search and efficient model evaluation. I built reproducible benchmarks across search strategies and diffusion- and hallucination-based baselines for binder and TCR design, outperforming BindCraft, PXDesign-d, and RFDiffusion3 on success rate. I contributed to architecture, template and domain redesign, and Monte Carlo Tree Search, reducing memory usage by 40%.

Release

The manuscript is under review at ICLR 2027.