BoltzProt-1

BoltzProt-1 is a de novo protein binder design pipeline released by Boltz on June 16, 2026, alongside the structure-prediction model BoltzMol-1 and the commercial Boltz API. It is the successor to BoltzGen, Boltz's open, generalizable protein-design foundation model, and represents a substantial step up in the ability to generate experimentally promising binders from scratch.

The central problem in computational binder design is not only generating plausible candidate structures but reliably ranking them so that the few selected for synthesis are the ones most likely to bind in the lab. Most design pipelines rank candidates by structural confidence, a signal that correlates only weakly with actual binding affinity. BoltzProt-1 improves both halves of the problem—generation and scoring—and its key innovation lies in how it prioritizes candidates.

Built on the diffusion-based Boltz lineage, BoltzProt-1 is positioned as a production tool for therapeutic discovery, with an emphasis on nanobody (single-domain antibody) design. Boltz is careful to note that its outputs are not finished therapeutic biologics; they are strong starting points that still require downstream optimization and experimental validation.

#Key Features

• Boltz-PPI scoring: A custom protein-protein interaction (affinity) model, trained on structural and patent-derived data, scores proposed interactions directly—mirroring the affinity approach used in Boltz-2—rather than relying on structural confidence as a binding proxy.
• Binding-oriented ranking: By prioritizing candidates with a criterion closer to experimental success, BoltzProt-1 makes large-scale generation tractable when screening tens of thousands of candidate binders.
• De novo nanobody design: Specialized for generating single-domain antibody binders against challenging targets without requiring an existing binder as a template.
• Strong developability profile: 58% of its designed binders passed a stringent multi-criteria developability panel in full, ahead of BoltzGen (40%) and clinical-stage VHH controls (21%).
• Accelerated inference: NVIDIA cuEquivariance kernels power efficient inference, supporting predictions from roughly $0.025 each via the Boltz API.

#Technical Details

BoltzProt-1 builds on the diffusion-based generative architecture of the Boltz/BoltzGen family for candidate generation, and pairs it with Boltz-PPI, a separately trained interaction-scoring model. Boltz-PPI is trained on a combination of structural data and patent-derived interaction data to predict binding signal directly, an approach analogous to the affinity head introduced in Boltz-2. Across ten novel targets, ranking candidates with Boltz-PPI raised the confirmed-binder hit rate from 3.3% to 8.0%. On published benchmark targets, BoltzProt-1 produced binders for 7 of 10 targets, ahead of Chai-2's 6 of 10. On developability, 58% of designs passed the full multi-criteria panel, compared with 40% for BoltzGen and 21% for clinical-stage VHH controls. The pipeline is documented in a June 2026 bioRxiv preprint and delivered through the commercial Boltz API rather than as open weights.

#Applications

BoltzProt-1 targets therapeutic and research workflows that need new binders against difficult protein targets, particularly nanobody discovery campaigns. It is served through the Boltz API (api.boltz.bio) with Python and JavaScript SDKs, first-party Claude Code integration, and support for the Codex and Gemini CLIs. Boltz lists partner platforms including Benchling, Phylo, Amazon Bio Discovery, Rowan, Tamarind, Kiin Bio, Pauling.ai, Mirror Physics, and Cultivarium, situating the model within existing drug-discovery and lab-automation stacks. Generation can be scaled to tens of thousands of candidates, with Boltz-PPI narrowing them to the most promising for synthesis.

#Impact

BoltzProt-1 advances the practical frontier of de novo binder design by addressing the long-standing gap between in silico confidence and in vitro binding. More than doubling the confirmed-binder hit rate over unranked generation and surpassing clinical-stage nanobodies on developability suggests computationally designed binders are moving closer to direct utility in therapeutic pipelines. Notably, BoltzProt-1 marks a shift in Boltz's release strategy: where Boltz-1, Boltz-2, and BoltzGen shipped as open-source models, BoltzProt-1 is delivered API-only, reflecting the commercial value of high-quality binder design. Its limitations are clear—designs are starting points requiring downstream optimization, and the lack of open weights or training code constrains independent replication despite the published preprint.