Promera

Promera is a unified generative model that combines all-atom biomolecular structure prediction with binder filtering and controllable design in a single system. Modern co-folding models such as AlphaFold3 and Boltz have made structure prediction highly accurate, but two gaps have persisted: their confidence signals are weak at the essential task of telling designed binders apart from non-binders, and dedicated design methods have concentrated on unconstrained binder generation rather than on the controllable, constraint-aware design that real campaigns require. Promera targets both gaps at once, treating prediction, in-silico filtering, and design as facets of one model.

The model performs all-atom co-folding of biomolecular complexes and uses its own confidence metrics to rank candidate binders, distinguishing binders from non-binders more accurately than prior open models for both miniproteins and nanobodies. For design, Promera generates binders by predicting masked protein sequences, optionally guided by epitope, paratope, and template constraints, and it supports both free minibinder generation and VHH (single-domain antibody) design with framework conditioning.

Promera was introduced in June 2026 by Bowen Jing, Mihir Bafna, Daniel J. Diaz, Adam Klivans, and Bonnie Berger, spanning MIT CSAIL and the Center for Generative AI and NSF AI Institute for Foundations of Machine Learning at UT Austin. It continues the MIT line of generative structure work behind AlphaFlow, and is released openly: code under the MIT license, weights on HuggingFace, and a bioRxiv preprint under CC BY 4.0.

#Key Features

• Unified prediction and design: A single model performs all-atom co-folding, confidence-based filtering, and binder design, rather than stitching together separate structure, scoring, and generation tools.
• Accurate binder filtering: Promera's confidence metrics separate binders from non-binders more reliably than prior open models for both miniproteins and nanobodies, addressing a longstanding weakness of co-folding confidence as a design filter.
• Constraint-guided design: Binders are generated by masked protein sequence prediction with optional epitope, paratope, and template constraints, enabling targeted rather than purely unconstrained design.
• Nanobody design: Supports VHH nanobody generation with framework conditioning and integrates LigandMPNN for sequence redesign, with optional AbMPNN weights for antibody inverse folding.
• Open release: Code, pretrained weights, and the preprint are all publicly available under permissive licenses, built on the tinyprot library for structure processing and MSA retrieval.

#Technical Details

Promera performs all-atom co-folding in the AlphaFold3 and Boltz lineage, using a diffusion-based module to generate atomic coordinates from sequence and multiple-sequence-alignment inputs; its implementation draws on the boltz, alphafold3-pytorch, and openfold codebases and uses the tinyprot library for mmseqs2 MSA generation and structure handling. Design is cast as masked protein sequence prediction over the folded complex, with LigandMPNN and optional AbMPNN providing sequence redesign and antibody inverse folding. On co-folding benchmarks, Promera surpasses the open-source models OpenFold3-p2 and Boltz-2 on therapeutically relevant categories. When nanobody designs are filtered under its own co-folding confidence, Promera reaches success rates comparable to backpropagation-based design methods. The authors also propose a scaling law for co-folding models. Inference runs in bfloat16 precision on GPU hardware.

#Applications

Promera is aimed at antibody and protein engineers running de novo binder campaigns, where the ability to triage designs computationally before committing to wet-lab synthesis is as valuable as generating them. Its confidence filter can rank miniprotein and nanobody candidates, while its constraint-guided design mode targets specific epitopes or stabilizes defined conformational states. The authors demonstrate this on therapeutically important problems, including targeting viral glycoproteins and stabilizing the active state of a GPCR, both central to vaccine and drug discovery efforts.

#Impact

Promera reframes co-folding as a design-and-filtering engine rather than a prediction-only tool, directly attacking the binder-filtering bottleneck that limits how much wet-lab experiments can trust in-silico candidates. Its open release, MIT-licensed code and weights alongside a CC BY 4.0 preprint, makes a unified prediction-plus-design system freely available to academic and commercial groups, in the same spirit as Boltz. As a recent preprint, its reported gains over OpenFold3-p2 and Boltz-2 and its nanobody design results are in-silico demonstrations awaiting peer review and experimental validation, and the proposed scaling law for co-folding points toward further capability from larger models trained on more structural data.