Reproduce “TorchCraft: Unified binder design by inverting an all-atom structure predictor” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
TorchCraft Team, Yu Liu, Zhouhanyu Shen, Zhengyi Li, Xikun Huang, Jiaqi Liu, Shuxian Gao, Qilin Yu and 3 more
TorchCraft uses a frozen AlphaFold3-like all-atom structure predictor as a differentiable objective for directly optimizing binder sequences across minibinders, VHH antibodies, cyclic peptides, and ligand-binding proteins. In experimental tests, designs bound all four evaluated protein targets in each category: minibinders reached apparent affinities from 2.04 to 60.2 nM, while VHHs reached 24.3 to 354 nM; cyclic-peptide and ligand-binding results were computationally evaluated only.
Key insights
- Predictor inversion can serve as a common design mechanism across substantially different binder formats, provided that editable regions and structural objectives are configured for each task.
- Sequence priors affect more than likelihood scores: composition regularization improved predicted confidence and was associated with higher expression in a PDGFRβ campaign, while CDR-restricted IgLM guidance shifted VHH sequences toward antibody-like distributions and improved humanness-related measures.
- Optional sequence redesign with ProteinMPNN or antibody-specific redesign often improved computational benchmark pass rates, indicating that direct predictor inversion and downstream sequence refinement can be complementary.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.