Phylo × Boltz: design and screen with Boltz, end-to-end from a prompt
By Phylo
Biomni Lab now connects to Boltz! Bring your own Boltz API key, and Biomni can fold complexes, design protein and small-molecule binders, and screen libraries against your targets. All from a single prompt, running on your Boltz account.
Why Boltz
Boltz is building open frontier AI for biomolecular discovery. Their new release brings API inference for their latest models, Boltz-2 for structure and affinity prediction and BoltzGen for de novo binder design. Boltz-2 approaches FEP-level accuracy for protein-small-molecule binding affinity at roughly 1000x the speed, while BoltzGen has been validated across 26 targets and 8 wet-lab campaigns. The release also includes two new pipelines, BoltzProt-1 and BoltzMol-1 for protein and small molecule design. Together, they help teams predict complexes, rank binders, and design new proteins, peptides, antibodies, and nanobodies much faster.
How the integration works
Add your Boltz API key in Biomni Lab’s integrations settings. Once connected, Biomni can call Boltz when a task involves structure prediction, affinity scoring, binder design, ligand design, or virtual screening.
From a single prompt, Biomni can help teams:
- Fold and score molecular complexes
- Design proteins, peptides, antibodies, or nanobodies with constraints
- Screen libraries against a target
- Generate synthesizable small molecules
- Evaluate SMILES libraries
Boltz returns structures, predicted poses, and confidence metrics. Biomni brings those results into the broader workflow: analyzing datasets and literature, preparing inputs, tracking jobs, organizing outputs, and turning results into candidate tables and summaries. The result is a smoother path from target definition to candidate comparison.
Examples
A note on scale. Real design campaigns often generate 20,000+ candidates and can run for hours or days. The examples below are intentionally smaller so they’re easy to replay, but the same workflows can scale up by changing a single number.
1. Genetically Supported Nanobody Target Discovery for Ulcerative Colitis
Using OpenTargets and GWAS Catalog, identify the top 5 genetically-supported targets for ulcerative colitis with druggable extracellular or surface domains. For the highest-scoring target, use boltz to design 50 nanobody binders, focusing on accessible epitopes from the AlphaFold structure. Return the target shortlist with genetic evidence scores (L2G, MR, coloc) alongside the top 10 nanobodies ranked by binding confidence, with CDR sequences and interface metrics.
The agent ranks UC targets from Open Targets and GWAS Catalog by genetic support, filters for extracellular or surface-accessible proteins, and returns IL12B, CD274, IL10, IL23R, and GPR35 as the top five. Although IL12B ranks first overall (composite 0.955), the agent selects IL23R for nanobody design because it is a structured cell-surface receptor with a high-confidence extracellular domain and no approved antibody directly targeting the receptor itself. It pulls the AlphaFold v6 IL23R structure, focuses on the D1 IL-23-binding domain, and sends the epitope to Boltz. The top nanobody: iPTM 0.858, min PAE 3.49 Å, with a 19-aa CDR3 predicted to reach into the IL-23 binding cleft.
2. Single-Cell-Guided Nanobody Design Against Tph-Specific Surface Markers in Rheumatoid Arthritis
Run a single-cell analysis on the AMP-RA synovial dataset to identify surface markers enriched on pathogenic CD4+ peripheral helper T cells (Tph) versus other T cell subsets and healthy controls. Pick the most Tph-specific surface protein with a clean expression profile. Use boltz to design 50 nanobody binders against its extracellular domain, then counter-screen the top 20 against the closest expressed homolog on bystander T cells. Return the top 10 candidates ranked by on-target affinity and off-target selectivity margin.
On the AMP-RA Phase 2 synovial T cell atlas, the agent identifies Tph populations, runs differential expression against non-Tph T cells, and ranks accessible surface proteins by specificity, effect size, and significance. TNFRSF18/GITR emerges as the most Tph-specific surface receptor, ahead of OX40, CTLA4, LAIR2, and LAG3. Boltz designs 50 nanobodies against the GITR extracellular domain, then counter-screens the top 20 against OX40, the closest expressed bystander TNFR-family receptor. The top candidate, NB-001, has the strongest on-target binding confidence (BC 0.0378), very low OX40 off-target signal (BC 0.000028), ~1,330× selectivity, and strong structural support (iPTM 0.890).
Get started
1. Sign up for Boltz Lab and grab an API key, or apply for free credits from Biomni via this form!
2. In Biomni Lab, go to Settings → Integrations → Boltz and paste your key.
3. Try one of the prompts above, or bring your own target.