mBER

mBER (Manifold Binder Engineering and Refinement) is an open-source system for designing de novo protein binders in a specific molecular format, demonstrated on single-domain VHH nanobodies. Backpropagation-based design through structure prediction networks has produced high experimental hit rates for unconstrained "minibinders," but antibody-format design is harder: the immunoglobulin fold requires large conserved framework regions to be held fixed while diversity is concentrated in the complementarity-determining regions (CDRs). mBER shows that this constraint can be imposed entirely at inference time, by supplying AlphaFold-Multimer with an informative structural template and a protein language model sequence prior rather than by training a new antibody-specific network.

The system was developed by Erik Swanson, Michael Nichols, Supriya Ravichandran, and Pierce Ogden at Manifold Bio, and posted to bioRxiv in September 2025. It builds on the ColabDesign framework and inherits most of its loss functions and optimization schedule from BindCraft, with the core contribution being the template-and-prior conditioning that steers AlphaFold-Multimer toward plausible VHH-antigen complexes without multiple sequence alignment input.

mBER appeared contemporaneously with Germinal, which reached similar conclusions with a related backpropagation approach, and it reports experimental success rates comparable to closed-source antibody design systems including Chai-2 and JAM. Its distinguishing contribution is the scale of wet-lab validation behind those numbers.

#Key Features

• Inference-time format control: A partially masked framework sequence plus a matching structural template is sufficient to constrain design to the antibody format, with no additional training or fine-tuning of the underlying folding and language models.
• Language model sequence priors: A single forward pass over the masked VHH sequence yields per-position amino acid logits that both initialize and bias the optimization, keeping designed CDRs within antibody-like sequence space.
• Epitope hotspot targeting: Surface-exposed hotspot residues are selected from the predicted target structure, and the target is truncated to a 25 Å neighborhood around the hotspot to cut compute and sharpen epitope focus.
• Confidence-based filtering: AlphaFold-Multimer's ipTM score is validated as a practical enrichment filter, raising effective hit rates by up to an order of magnitude without increasing off-design binding.
• Modular, swappable pipeline: Template preparation, trajectory optimization, and evaluation are separate configurable modules, so the machinery extends to other scaffolds by supplying a different template and prior.

#Technical Details

An mBER run starts from a UniProt or PDB identifier (structures are pulled from the AlphaFold Protein Structure Database), optional hotspot specifications, and a partially masked VHH framework based on the human IGHV3-23 germline. During template preparation, ESM-2 (650M) generates logits over the masked CDR positions; a sequence is sampled and folded with NanoBodyBuilder2 to produce the binder template, which is merged with the truncated target structure. Design follows ColabDesign's three-stage protocol, optimizing primarily on the predicted aligned error matrix with auxiliary confidence, contact, and sequence losses; the ESM-2 logits are added as a temperature-scaled bias at every step. Four of the five pretrained AlphaFold-Multimer weight sets are used for design and the fifth held out for evaluation, both with three recycles. AbLang2 is available as an alternative sequence-guidance model. The stack needs roughly 9 GB of downloaded weights and an NVIDIA GPU with at least 32 GB of VRAM.

Validation covered 1,153,241 unique VHH designs against 436 human cell-surface targets across two phage-display libraries, screened all-against-all against 145 targets for more than 100 million measured binding interactions. Sixty-five targets (45%) showed significantly elevated on-design hit rates by Fisher's exact test. Unfiltered per-target hit rates ranged from about 0.02% to 8% (median 0.4%), rising into double digits above ipTM 0.7, with the most favorable epitope reaching 38% (8 of 21 designs). In an in-silico head-to-head under matched inputs, mBER produced higher ipTM binders than RFantibody.

#Applications

mBER targets therapeutic and reagent discovery programs that need binders in a clinically established format against a defined epitope. Manifold Bio used it to generate candidate tissue-targeted delivery vehicles against cell-surface receptors, the use case that motivated the 436-target campaign. Because format is specified by template and prior rather than by trained weights, the pipeline extends in principle to scFv or Fab fragments and to non-antibody scaffolds such as DARPins.

#Impact

This is the largest reported de novo protein design and validation campaign, and the all-against-all matrix gives unusual visibility into off-design binding, a property most binder-design papers cannot measure. The MIT-licensed code makes double-digit antibody design hit rates reproducible outside the companies that first reported them. The work remains a preprint that has not been peer reviewed, and its limitations are real: success is strongly epitope-dependent, with productive binders recovered for only a few hotspots per target; phage display establishes specificity but not affinity, and predicted binding modes were not confirmed structurally. No new model weights are contributed and the screening dataset has not been released, so the artifact is the design system and its conditioning recipe rather than a trained model.