ESMFold2-pipeline Arrives at Ariax Bio: Protein and Antibody Design at Campaign Scale

In May, Biohub released ESMFold2 as part of its new world model of protein biology. The all-atom structure-prediction model is built on the six-billion-parameter ESM Cambrian protein language model, and its performance is striking: ESMFold2 matches or surpasses AlphaFold3 across challenging protein-complex benchmarks, operates efficiently from sequence alone, and was experimentally validated as a design engine for both de novo miniproteins and antibody-derived scFvs.

Today, we are bringing that design method to Ariax as ESMFold2-pipeline, our fourth protein-design engine and our first workflow with first-class scFv design.

ESMFold2-pipeline turns Biohub's research method into a production campaign system: upload a target, select an epitope, choose a miniprotein, VHH, or scFv campaign, and generate a structurally validated shortlist across one or more GPUs. It handles target preparation, parallel design, ESMFold2 scoring, independent Protenix validation, consensus ranking, checkpointing, and result collection.

And uniquely, Ariax is the only hosted platform where ESMFold2-pipeline is available.

If you're new to Ariax, learn about our mission in Introducing Ariax Bio.

ESMFold2 Changes the Structure-Prediction Equation

The original ESMFold showed that a protein language model could predict structure directly from a single sequence, without first constructing a multiple sequence alignment. ESMFold2 advances that idea substantially.

Built on representations learned by ESMC-6B, ESMFold2 combines a recurrent pair-folding architecture with an all-atom diffusion model. It supports both single-sequence and MSA-conditioned prediction, while the streamlined ESMFold2-Fast variant is optimized for high-throughput single-sequence inference.

In the ESMFold2 preprint, the model achieved:

• A 50% DockQ pass rate on antibody-antigen complexes from sequence alone, compared with 47% for AlphaFold3 using an MSA.
• A 70% pass rate on protein-protein complexes and 66% on protein-ligand complexes from single-sequence inputs.
• With MSA context, pass rates of 53% for antibody-antigen and 76% for protein-protein complexes.
• A 1,024-residue prediction in 15.8 seconds on an H100, with ESMFold2-Fast completing the same benchmark in 9.4 seconds under the paper's matched inference settings.

The particularly important result for protein design is not simply accuracy. It is the combination of accuracy and throughput. Protein design is a search problem: every credible candidate gives the model another opportunity to discover a genuinely successful binding solution.

ESMFold2 makes that search unusually scalable.

From Structure Prediction to Experimental Binders

Biohub's team inverted ESMFold2—optimizing candidate sequences against the structures and interfaces predicted by the model—to design binders against five therapeutically relevant targets: PDGFRβ, EGFR, PD-L1, CD45, and CTLA-4.

They tested two modalities: compact de novo miniprotein binders and antibody-derived single-chain variable fragments, or scFvs.

In the higher-compute campaigns, experimentally measured hit rates reached 36–88% for miniproteins and 15–29% for scFvs. Averaged across the five targets, increasing inference compute raised miniprotein hit rates from 53.8% to 70.0% and scFv hit rates from 12.1% to 21.0%.

The resulting binders were not merely detectable hits. Reported affinities ranged from 68 picomolar to 70 nanomolar, with nanomolar binders found for every target and both design modalities. The designs engaged their intended epitopes, discriminated between closely related homologs, and included PD-L1 binders with activity in a cellular immune-checkpoint assay.

One EGFR miniprotein was also characterized by cryo-EM. Its experimentally observed complex agreed with the computational design at 1.204 Å RMSD, providing unusually direct structural confirmation that ESMFold2's predicted binding pose was realized in the laboratory.

These remain preprint results and every computational design still requires experimental validation. Nevertheless, the combination of broad target coverage, high hit rates, strong affinities, functional activity, and structural confirmation makes ESMFold2 one of the most compelling new design systems available.

Turning a Research Method into a Campaign Pipeline

Biohub released the ESM code and model weights, along with a binder-design tutorial demonstrating the underlying method. That notebook is an excellent way to understand and explore ESMFold2 design.

A real campaign, however, requires much more than a notebook.

You need to manage many independent designs, distribute them across GPUs, recover from interruptions, track which outputs are complete, score interfaces consistently, validate top candidates with an independent model, and convert all of that information into a defensible experimental shortlist.

ESMFold2-pipeline provides that missing production layer.

Each campaign proceeds through four stages:

1. Design: ESMFold2 optimizes candidate binders against the target.
2. Critic: Candidates are folded again in complex and scored using interface confidence, particularly ipTM.
3. Validate: Protenix v2 independently predicts selected complexes and reports its own ipTM, ipSAE, and structural agreement with ESMFold2.
4. Rank: Candidates are ranked using confidence from both models together with target-aligned binder RMSD.

Rather than trusting a single model's score, this workflow looks for candidates on which two independently developed predictors agree. That does not replace wet-lab validation, but it provides a stronger basis for deciding which sequences deserve synthesis.

Miniproteins, VHH Nanobodies, and scFvs

One ESMFold2-pipeline interface supports three distinct design modalities:

• Miniproteins: Fully de novo binders with configurable lengths.
• VHH nanobodies: Single-domain antibody binders built on structurally defined clinical frameworks.
• scFvs: Paired VH-linker-VL antibody fragments with both heavy- and light-chain CDR design.

The scFv capability is especially notable. Flexible antibody CDR loops remain one of the hardest problems in computational protein design, and ESMFold2's antibody-antigen performance is a central part of what distinguishes it from earlier sequence-based folding models.

On Ariax, antibody campaigns can sweep across bundled clinical framework structures rather than relying on sequence-only scaffolds. CDR-scoped attraction directs optimization toward the antibody's intended binding loops, while an optional framework-contact penalty can discourage solutions that depend on unintended framework interactions.

Design Against the Biology You Actually Care About

ESMFold2-pipeline accepts PDB and mmCIF target structures, including multichain complexes and assemblies. Ariax can also search the RCSB Protein Data Bank directly by identifier or keyword.

After loading a structure, you can select the relevant chains, crop the target to specific domains, and define hotspot residues corresponding to the epitope you want to engage. Hotspots influence design and are also used during selection, helping distinguish candidates that merely contact the target from candidates that address the intended site.

The target's known geometry can be supplied to ESMFold2 as a structural prior through distogram conditioning. Partial experimental structures are supported, so unresolved residues do not prevent the available structural information from being used.

This makes the pipeline suitable for realistic design problems: multimeric receptors, peptide-MHC complexes, domain-specific epitopes, and targets where preserving an experimentally observed conformation matters.

Independent Validation Comes Standard on Ariax

Every ESMFold2-pipeline campaign on Ariax includes Protenix v2 validation.

By default, candidates reaching an ESMFold2 ipTM threshold of 0.6 are sent to Protenix. The final ranking considers ESMFold2 ipTM, Protenix ipTM and ipSAE, and the RMSD between the two predicted binder poses after aligning their targets. Designs for which both models predict a strong interface and a similar geometry rise to the top.

Ariax automatically handles structural templates and can prepare and reuse target and binder MSAs during validation. The resulting project contains both models' structures, complete metrics, the final ranked shortlist, and diagnostics explaining why individual candidates were excluded.

For details on inputs, model choices, ranking, and results, see the ESMFold2-pipeline project guide.

Campaign-Scale Compute Without Campaign-Scale Infrastructure

The preprint's central design lesson is straightforward: more computational search improved experimental outcomes in nine of ten target-modality comparisons. That does not mean every campaign needs to reproduce the paper's maximum-search experiments. For practical work, we recommend piloting dozens to a few hundred designs, reviewing the ranked results, and scaling a production campaign accordingly.

ESMFold2-pipeline was built to make that progression easy. Campaigns distribute designs across multiple GPUs, resume without repeating finished work, and checkpoint their state throughout execution. On Ariax, Turbo Mode can parallelize a campaign across multiple private GPU instances, trading additional parallel compute for a shorter wall-clock time.

You do not need to install custom CUDA kernels, provision an 80 GB GPU, maintain model checkpoints, coordinate workers, or reconstruct an interrupted run. Ariax handles the infrastructure while your structures, parameters, and results remain isolated to your project.

As with our other engines, there are no subscriptions or platform fees. You pay only for the compute your campaign uses, and the resulting sequences remain your intellectual property.

Free for Research and Commercial Use—Uniquely Hosted on Ariax

Biohub's ESM models are open-source, while ESMFold2-pipeline is source-available. Neither requires a separate commercial software license for academic or commercial use, whether you run the pipeline locally or through Ariax.

Researchers who want to manage their own infrastructure can install ESMFold2-pipeline from its public GitHub repository. For everyone else, Ariax is the only platform offering the complete ESMFold2-pipeline workflow as a hosted service.

ESMFold2 model inference is also available through Biohub; the exclusivity here applies specifically to the complete ESMFold2-pipeline binder-design workflow.

A Complementary New Design Engine

ESMFold2-pipeline joins BindCraft, BoltzGen, and PXDesign on Ariax.

Each engine brings a different search strategy and set of strengths:

• BindCraft remains our most extensively validated engine for de novo miniprotein and helical-peptide design.
• BoltzGen provides broad all-atom design across VHH antibodies, peptides, helicons, and non-protein molecular targets.
• PXDesign offers a fast diffusion-based route to structurally diverse miniprotein binders.
• ESMFold2-pipeline combines protein-language-model inversion with exceptionally fast complex prediction and is our first engine with dedicated scFv design.

For difficult targets, these methods should be viewed as complementary. Different model architectures explore different regions of sequence and structure space, and agreement across independent campaigns can provide valuable evidence when selecting candidates for experimental testing.

Available Now

ESMFold2-pipeline is live on Ariax for miniprotein, VHH, and scFv campaigns. Upload or search for your target structure, choose an epitope and design modality, select your campaign size, and launch.

Start an ESMFold2-pipeline campaign →

For setup and result interpretation, see the Ariax project guide. For local installation and technical details, visit the ESMFold2-pipeline repository. For the model, benchmarks, and experimental validation, read the ESMFold2 preprint, the official Biohub announcement, and the Biohub ESM repository.