A new framework called TorchCraft designs binder proteins by running a structure predictor in reverse.
Most structure prediction models, like AlphaFold 3, take a protein sequence and guess its 3D shape. TorchCraft flips that process: it freezes a pretrained AlphaFold 3 model and optimizes a candidate sequence against it, adjusting sequence logits until the predictor is confident the resulting structure will bind a chosen target. The framework, built on a system called TorchFold, blends confidence, contact, geometric, and sequence-prior objectives into one optimization loop. It works across four different binder formats: minibinders, VHH antibody fragments, cyclic peptides, and ligand-binding proteins. The researchers report experimentally measured binding for minibinders and VHHs across four targets each, with no post hoc sequence redesign needed to get there.
The "no redesign" detail is the actual news here. A lot of AI-driven protein design tools still lean on a second-pass optimization step to clean up sequences before they're tested in the lab, which is a tell that the generative model alone isn't fully trustworthy. TorchCraft skipping that step, and still producing binders that worked experimentally, suggests the underlying structural priors in AlphaFold 3 are usable directly for design, not just prediction.
The bigger idea is reuse: instead of training separate generative models for every binder type, TorchCraft treats one frozen predictor as a general-purpose design engine across formats. That's a meaningfully different bet than the RFdiffusion-style approach of training dedicated generative diffusion models for binder design. Whether inversion beats purpose-built generation will come down to hit rate and affinity in follow-up wet-lab work, which this paper only partially provides -- the cyclic peptide and ligand-pocket results are computational benchmarks, not lab-confirmed binders.