FoldCraft

FoldCraft is a de novo protein binder design pipeline that lets the user dictate the shape of the binder, not just its target. Most deep learning binder design methods produce whatever topology the generator finds convenient — typically compact bundles of regular secondary structure — which leaves large regions of fold space, and most antibody-like architectures, out of reach. FoldCraft addresses this by conditioning the design trajectory on a reference fold, so the resulting binder adopts a specified topology while still forming a high-confidence interface with the chosen target.

The method works by hallucination through a frozen structure predictor rather than by training a new network. A random binder sequence is initialized, AlphaFold-Multimer predicts the binder-target complex, and a single loss — the difference between the predicted contact probability map and a fold-conditioned template contact map — is backpropagated to the binder sequence. Because the loss simultaneously encodes the desired intra-chain geometry and the desired contacts with target hotspots, one term does the work that comparable pipelines split across half a dozen auxiliary losses.

FoldCraft was introduced by Khondamir Rustamov and Artyom Baev of the Center for Advanced Technologies in Tashkent in a bioRxiv preprint posted in July 2025, which has not yet been peer reviewed. It sits alongside hallucination-based design pipelines such as Germinal and diffusion-based ones such as RFdiffusion, and it is the fold-control axis of that landscape rather than a new structure predictor.

#Key Features

• Contact-map similarity loss: Fold conditioning is expressed as an RMSE between the predicted and reference contact maps over a masked subset of residue pairs, computed from AlphaFold's distogram at 14 Å for intra-chain and 21.7 Å for inter-chain contacts.
• Arbitrary target topologies: Binders were generated in six distinct folds — Top7, β-barrel, Ig-like, TIM barrel, β-solenoid, and ankyrin repeat — each defined simply by supplying a reference structure.
• VHH nanobody design: A VHH-specific framework masks the flexible CDR loops when building the conditioning map, letting the pipeline generate single-domain antibody binders without an antibody-specific network.
• Implicit confidence optimization: The single cmap loss correlates with AlphaFold's own pLDDT, PAE, ipTM, and interface-contact metrics, so fold fidelity and interface confidence improve together.
• Runs on frozen public weights: The pipeline uses ColabDesign and stock AlphaFold2 parameters, with MIT-licensed code and Google Colab notebooks for both general and VHH design.

#Technical Details

Design follows ColabDesign's three-stage sequence optimization — 100 logit steps, 100 softmax steps, and 20 one-hot steps — with the target supplied as a fixed structural template and the binder predicted in single-sequence mode. Designs are then refined with ProteinMPNN or its soluble variant SolMPNN at temperature 0.1 and zero backbone noise, either over the full sequence or restricted to residues more than 4 Å from the interface. Against PD-L1, 8–45% of designs per fold landed within 3.5 Å RMSD of the monomeric fold template, and FoldCraft produced higher ipTM and lower interface PAE than RFdiffusion across all six topologies. For nanobodies, 40 backbones and five sequences each were generated per target for PD-1, PD-L1, IFNAR2, and EGFR and scored with AlphaFold 3 as the oracle: FoldCraft reached 0.5–19.5% in silico success (pLDDT > 70, interface PAE < 10, ipTM > 0.5), while no RFAntibody design passed the same interface filters. All results are computational; the preprint reports no experimental validation.

#Applications

FoldCraft is aimed at protein engineers who need a binder with a particular architecture — a solenoid or repeat scaffold for multivalent display, an Ig-like domain for reagent compatibility, or a VHH nanobody for intracellular and diagnostic use — against targets where no natural binder exists. Because it runs on public AlphaFold2 weights through Colab notebooks, small academic groups can attempt fold-specified binder campaigns without diffusion-model training infrastructure.

#Impact

FoldCraft demonstrates that fold control in binder design can be reduced to one geometric loss over a structure predictor's distogram, an approach that generalizes to any topology for which a reference structure exists. Its main limitations are stated plainly in the preprint: the AlphaFold3 oracle frequently places nanobodies in side-docked geometries, an artifact traced to training-data bias, and the method did not extend to scFv antibodies. The open repository has continued to develop past the preprint, including a documented variant that swaps AlphaFold2-Multimer for a JAX implementation of Boltz-2 with iterative ProteinMPNN redesign, and designs from that protocol were submitted for wet-lab evaluation in a binder design competition.