BoltzDesign1

BoltzDesign1 turns a structure predictor into a designer. Rather than training a new generative model, it takes Boltz-1 — the open reproduction of AlphaFold 3 — freezes its weights, and backpropagates a design objective through it until an input sequence emerges that the model confidently folds into a binder for a chosen target. This is the "hallucination" strategy first developed on trRosetta and AlphaFold 2, carried over for the first time to an all-atom model that natively handles small molecules, nucleic acids, ions, and covalent modifications alongside protein.

The method was developed by Yehlin Cho, Zhidian Zhang, and Sergey Ovchinnikov at MIT with Martin Pacesa and Bruno Correia at EPFL and the SIB Swiss Institute of Bioinformatics, and posted to bioRxiv in April 2025. Its central trick is to avoid the diffusion module entirely: backpropagating through 200 diffusion steps is expensive, memory-hungry, prone to vanishing gradients, and samples only one structure from a distribution. BoltzDesign1 instead optimizes directly on the distogram produced by Boltz-1's Pairformer — the predicted probability distribution over inter-atomic distances — optionally adding a confidence loss routed through the Confidence module with a stop-gradient on the diffusion step.

BoltzDesign1 should not be confused with BoltzGen, a later and separately trained generative design model. BoltzDesign1 trains nothing: Boltz-1's checkpoint is used exactly as published, and the only learning that happens is gradient descent on the input sequence.

#Key Features

• No fine-tuning required: The frozen published checkpoint is wrapped as a differentiable oracle, so new targets need no retraining, no design dataset, and no extra weights.
• Diffusion-free optimization: Optimizing the distogram rather than sampled coordinates shapes the whole distance distribution, steering designs toward sequences with well-defined energy minima while cutting compute and memory cost sharply.
• Targets beyond protein: Binders were designed against small molecules, zinc and iron ions, B-DNA, and post-translationally modified sites including phosphorylated PCNA Tyr-211, phosphorylated Smad2, and glycosylated CD45.
• Flexible ligand modeling: Because the model predicts binder and ligand jointly, the ligand conformation is re-derived at each iteration instead of being held fixed as in backbone-only diffusion methods, allowing induced-fit effects to be explored.
• Sequence redesign hand-off: Designs pass to ProteinMPNN or LigandMPNN for surface redesign, with interface residues optionally held fixed.

#Technical Details

Sequence optimization proceeds in four stages, mirroring BindCraft's schedule: a warm-up in continuous space at temperature 1.0, a mixed logits/softmax phase with linearly ramped interpolation, a temperature-annealed softmax phase, and a final one-hot stage using a straight-through estimator. Losses are entropy-based contact terms over the 64-bin distogram spanning 2–22 Å, with a 14 Å cutoff for intra-binder contacts and 22 Å for binder–target contacts; a minimum sequence separation of 9 residues on the intra-contact term suppresses trivial all-helix solutions.

Validation is entirely in silico. On the four small-molecule targets used to benchmark RFdiffusionAA, the all-atom diffusion designer built on RoseTTAFold All-Atom — IAI, FAD, SAM, and OQO — BoltzDesign1 beat RFdiffusionAA on AlphaFold 3 success rate (complex pLDDT > 0.7 and interface PAE < 10) for all four ligands, with the Pairformer-plus-Confidence variant ahead of Pairformer alone on three of four. Designs were also more diverse, with a mean pairwise TM-score of 0.36 versus 0.46 for RFdiffusionAA. Gnina docking placed 9.3% of SAM designs, 7.3% of OQO designs, and 4.0% of IAI designs above the wild-type complex score. Best results came from zero Pairformer recycling with the interface sequence fixed. The authors report no wet-lab validation, and the repository labels itself experimental software.

#Applications

The practical draw is targets that protein-only design pipelines cannot address at all: metal-coordination sites, double-stranded DNA, and specific post-translational modifications. A binder that discriminates a phosphorylated residue from its unmodified form is directly useful for biosensors and intracellular signaling probes, and the demonstrated PCNA, Smad2, and CD45 targets are all disease-relevant. Small-molecule binder design likewise feeds sensor engineering and enzyme scaffolding. The pipeline runs from a cloned repository or a hosted Colab notebook, with Boltz-1 weights fetched automatically at setup.

#Impact

BoltzDesign1 established that the AlphaFold 3 generation of all-atom predictors can be inverted for design without retraining, and it did so early enough that the pattern has spread through the open structure-prediction ecosystem. Its reported gains over RFdiffusionAA in both success rate and diversity are notable given that no new model was trained. The caveats are substantial: the results remain a bioRxiv preprint; success is measured by an AlphaFold 3 filter rather than by experiment, and the authors state explicitly that no design has been tested in the laboratory; template inputs and nucleic-acid MSAs are not supported, limiting protein–DNA and protein–RNA work; and using one model family for both design and evaluation risks a shared blind spot. The repository's README declares an MIT license, but no LICENSE file is present in it.