BigHat Biosciences joins AI antibody design with automated experimental data

BigHat Biosciences combines computational antibody design with an automated laboratory and a system for managing experimental data. Its Milliner platform supports repeated cycles of making candidates, measuring their properties and using the results in further design. The company develops therapeutic assets and offers partnerships around antibody engineering and model-training data. That makes its laboratory and data quality as important to evaluate as its AI models.

01 / ProductThe platform links candidate design to traceable measurements

The platform page describes Milliner as an automated wet-lab system built around cell-free protein synthesis. Reccy coordinates data and laboratory workflows, while RADS provides machine-learning design capabilities. The architecture is intended to connect proposed antibodies with measurements that can inform later model training and candidate selection.

The design-to-data explanation describes how experimental designs become machine-readable instructions and how results are associated with the relevant experiment. This is a specific operational problem: when many candidates and assays are running, a measurement is useful only if its identity and context remain intact.

BigHat’s company overview places this platform within both wholly owned and partnered therapeutic programs. Its pipeline includes BHB810, which the company identifies as an antibody-drug conjugate in a Phase 1 study. That establishes a clinical-stage program, not an approved treatment or proof that the platform predicts patient outcomes.

A useful distinction is between discovering a binder and engineering a viable candidate. Binding is one relevant property, but the research process also needs to understand whether a molecule can be produced, characterized and developed for the intended purpose. BigHat’s proposition is to measure several relevant dimensions and return that information to the design process.

02 / AudienceThe commercial fit can be a therapeutic program or a data problem

The partnering page identifies therapeutic collaboration, data generation and model development, and asset licensing as distinct routes. A partner can therefore approach BigHat with different missing capabilities. It may need help improving an antibody, a reliable experimental dataset or rights to a particular development asset.

A therapeutic team should describe the property that limits its candidate, rather than simply request a better antibody. A model developer should define what the dataset needs to represent and how the measurements will be used. These are different scientific jobs, even if they rely on the same laboratory infrastructure.

The fit is less obvious for a user seeking a self-service generative website or an ordinary laboratory-information software subscription. The public pages describe integrated capabilities and collaborative work; they do not establish that each named internal component can be purchased independently.

Absci is a relevant comparison for AI-linked antibody discovery. Generate Biomedicines offers another view of generative protein design and therapeutic development. Compare the actual starting problem, experimental contribution and partnership scope. The shared use of generative AI is not enough to make the offers interchangeable.

03 / WorkflowA proposed evaluation asks what the next measurements will teach

Imagine a proposed project with an antibody that binds a research target but has an undesirable property affecting its usefulness. Begin by separating what is known from what is assumed. The starting molecule, the measured problem and the acceptable tradeoffs should be explicit before any model is asked to generate variants.

Next, choose the evidence needed to assess improvement. A candidate can appear better on one measurement while losing another property the program depends on. In this proposed evaluation, the team would define a small set of relevant outcomes and establish which are essential and which can be traded. This is a decision framework, not a laboratory protocol.

Design the first candidate set to test more than one hypothesis about the limitation. Some variants may preserve familiar features; others may explore a different direction. The useful output is a set whose results can influence the next round, rather than a large volume of similar designs that all succeed or fail for the same reason.

Before interpreting measurements, inspect how sample identity and experimental context travel through the workflow. Reagent or instrument differences can be mistaken for molecular effects if the metadata are lost. A platform’s ability to produce data quickly is valuable only when those data remain comparable and interpretable.

Keep failed expression, ambiguous measurements and clean negative results distinguishable. They carry different information for the next model update. Quietly removing difficult examples can make a dataset look complete while teaching the model only about candidates that were easiest to measure.

Then compare successive rounds against the original decision. Did the new information narrow the design problem, expose a previously hidden tradeoff or improve the candidate sufficiently to justify further development? Preserve the evidence even when the answer is to stop. A well-supported rejection can be a more valuable research outcome than another round of optimistic generation.

The proposed endpoint is an experimentally supported choice and a clear account of what remains unresolved. We did not use Milliner, run RADS or evaluate a BigHat-produced candidate. The public sources explain the platform’s intended mechanism, while a partner must assess performance on its own program.

04 / PricingPartnership prices depend on whether the output is an asset or data

These routes should not be collapsed into a common unit price. A dataset assembled for model training has different completion criteria from a candidate selected for development. A collaboration may also include scientific interpretation and iteration that a price per sample would fail to capture.

For a data project, the estimate should specify what counts as a usable result and how missing or ambiguous measurements are represented. For a therapeutic project, it should define the evidence required at the next decision point. The reviewed site gives a basis for discussing those scopes, while the applicable price and terms require a direct proposal.

Asset licensing is different again: it concerns rights to a particular program and its future development. A program’s presence on the pipeline page should not be interpreted as a public offer on standard terms. Availability and the actual transaction structure need to be established for that asset.

05 / DistinctionsThe data-generation capability has a concrete external example

The October 2026 NanoLab account says BigHat generated experimental data for Lilly TuneLab’s VHH developability model. It describes attention to missing results and ambiguous measurements rather than simply discarding them. This is a useful example of how laboratory work can support an AI product without granting external users access to BigHat’s entire platform.

BigHat’s protein-language-model evaluation also reports that sampling choices and campaign-specific data materially affected results in its antibody-engineering tests. Those are vendor-run findings within a defined task. They should not become a universal ranking of protein models, but they explain why the company emphasizes data generation alongside model selection.

The practical implication is that an organization should examine the measurement system as carefully as the generative model. A larger model cannot recover experimental context that was never captured. Conversely, a smaller or more familiar model may become more useful when the program produces informative data and a clear way to assess its outputs.

That focus also changes what a demonstration should show. The compelling sequence is design, sample, measurement, interpretation and revised decision. A gallery of molecular candidates exposes only the first stage and leaves the rest of the proposed advantage untested.

06 / QuestionsThroughput and clinical stage do not establish transfer to a new project

The platform advertises rapid design-to-data cycles and broad characterization. Treat these as company-reported capabilities, not a contractual schedule for every assay or antibody format. A prospective partner should identify where its work fits the existing system and where new method development is required.

For a training dataset, ask how its coverage matches the intended use. A carefully measured set can still be unrepresentative of the structures or properties a model will encounter later. The acceptance criteria should make that mismatch visible before a large dataset is mistaken for a complete answer.

For a therapeutic program, distinguish computational selection, preclinical evidence and clinical development. BHB810’s Phase 1 status is meaningful company progress, but it does not prove the efficacy of the molecule or validate every other program. The relevant evidence remains specific to the asset and study.

Finally, clarify the boundary between BigHat’s data-generation work and any external platform that uses the results. The NanoLab relationship illustrates that a model may be distributed through a partner. It does not imply that a TuneLab user receives BigHat laboratory services or rights to the underlying experimental corpus.

07 / DecisionSelect the engagement around the research output you need

BigHat Biosciences is most compelling when an organization needs reliable experimental feedback to improve antibody design or train relevant models. The starting point should be a precise research output and the evidence required to accept it. That keeps the discussion grounded in the laboratory and data decisions that make the AI useful.