BigHat Biosciences

Partners come to BigHat for the design loop. Milliner is built to serve the models rather than as a discovery operation with analysis added: it synthesizes and characterizes more than 2,000 molecules a week across more than 20 assays so the models can be retrained on fresh measurements, with every instrument feeding a data lake that flows straight into training. Both founders describe the company's thesis as moving machine learning from the end of the research pipeline to the front, and the platform is that argument built out. BHB810, the first asset to reach a patient, is described as the first program the company has taken from AI generated design into human studies.

The Synaffix conjugate chemistry, licensed in November 2024, supplies the payload downstream of the design engine.

The oversight is physical, and the company describes it in its own voice. A model proposal counts as a result only once it has been built and measured: more than 2,000 molecules are synthesized, purified and characterized across more than 20 assays every week on a one week turnaround, every data point passes quality control with controls evaluated, and the models are retrained on what came back. Greenside has stated the underlying position explicitly, that the company works from smaller informative datasets with fast feedback rather than from volume, and has publicly criticized developability models that generalize poorly because they were trained on inconsistent data.

The platform is described as running agent harnesses at every part of the design round workflow; who decides which designs are built, and what happens when a model and an assay disagree, is not published.

Every part of the platform is described and quantified except the models themselves.

Milliner is the automated wet laboratory, built around cell free protein synthesis and end to end automation, reported at 2,000 or more molecules produced and characterized per week, a one week turnaround from design to data, more than 20 assays spanning developability, affinity and function, and tens of thousands of labeled data points a week. Reccy is the data control plane, a bespoke laboratory information management system that captures every instrument's raw output with experimental metadata and quality control into a single data lake. RADS is the production machine learning stack, described as agent automatable, designing millions of molecules a week and training hundreds of models on the company's proprietary data. Eleven publications at machine learning venues give the underlying methods.

No model, architecture or version is named: the production system is called frontier intelligence and a zoo of models managed by agents, and training data is described by throughput rather than size or composition.

The disclosure rests on named and dated external dependencies rather than a policy. Conjugate chemistry is in licensed from Synaffix, a Lonza company, from November 2024, and is identified as the technology behind the lead program. The Lilly TuneLab arrangement discloses the outbound direction, with BigHat generated datasets feeding a developability foundation model the company does not own. The compliance stack is named too: the SOC 2 engagement was performed by BARR Advisory, P.A. using Vanta. And the published papers name the methods the platform builds on and cite the public work behind them, so the intellectual supply chain is traceable in a way a marketing page never makes it.

The model layer itself is in house, and no third party model provider is named or implied. Still withheld: compute and hosting providers, training data provenance for the internal models, an open source component inventory, any third party model license, and anything stating what governs a partner's sequences and assay data once they reach the platform.

A molecule in human trials and a published method record are both now in place, and the trial is the recent arrival.

On 1 September 2026 BigHat announced the first patient dosed in a first in human Phase 1 dose escalation trial of BHB810, a Cadherin-17 directed VHH-Fc antibody drug conjugate for gastric cancer and other advanced gastrointestinal tumors, initially enrolling advanced gastric and gastroesophageal junction tumors and evaluating safety, tolerability, pharmacokinetics and preliminary antitumor activity. The principal investigator is named, Alexander Spira of NEXT Oncology-Virginia. The preclinical package behind it is reported with a denominator: across nearly 30 patient and cell derived tumor models, including gastric cancers with low and heterogeneous target expression, BHB810 drove complete or near complete tumor clearance. A second asset, BHB299, an avidity driven T cell engager targeting CEACAM6, is stated to enter the clinic in 2027.

The method record runs deep. A publications page lists eleven items from 2021 to 2026 at machine learning venues rather than in the biology press: ICML 2021, PMLR 2022, ICLR 2024, 2025 and 2026, and NeurIPS 2025, on Bayesian optimization for antibody design, denoising autoencoder surrogates, generative humanization, guided sequence and structure generative modeling, multimodal reformatting, masked language models for protein engineering and a unification of discrete, Gaussian and simplicial diffusion. The thread is coherent rather than scattered.

Two qualifications matter. No trial registration identifier is published for BHB810 and no readout is dated, so the program rests on an announcement rather than a protocol. And venue labels on the publications page are the company's own, with at least one entry a workshop track rather than a main conference, so the entries are not uniform in standing.

Antibody sequences and biophysical and functional readouts from internal and partnered programs sit in Reccy, the laboratory data system, and a named assessor performed a SOC 2 Type 2 examination covering that system's security, availability and confidentiality. The platform's stated design is that every measurement feeds automatically into model training, and the company now sponsors a Phase 1 trial and so holds human subject data.

Whether one partner's sequences and assay results train models that serve other programs is not stated.

BigHat's website carries no privacy policy, terms of use or data protection notice, and nothing on it addresses protected health information or the health privacy rule. The company now sponsors a first in human Phase 1 trial of BHB810 at a named clinical site, so subject data flows through that trial under the clinical research regime, while the platform itself works on molecules and assay data. The SOC 2 Type 2 examination of its laboratory data system covers security controls rather than privacy.

No position on business associate status, or on what the platform processes, is published.

BigHat announced a SOC 2 Type 2 examination covering the security, availability and confidentiality of data in Reccy, its laboratory infrastructure management system, performed by BARR Advisory, P.A. with Vanta as the compliance automation platform. The scope fits the business: Reccy is the database holding antibody sequences and biophysical and functional readouts from both internal and partnered programs, so the audited system is where a partner's data sits.

The announcement is dated 29 September 2022, and a SOC 2 Type 2 report is conventionally treated as current for about twelve months from its report date; no renewal has been announced since.

The models and laboratory serve the company's own programs and its partners, so no device claim or clearance attaches to them.

Regulatory status belongs to the company's medicines. BHB810, a Cadherin-17 directed antibody drug conjugate, entered a first in human Phase 1 dose escalation trial with the first patient dosed on 1 September 2026, which implies an investigational new drug application cleared ahead of it, and BHB299, a CEACAM6 directed T cell engager, is planned for the clinic in 2027. The conjugate chemistry is licensed from Synaffix, a Lonza company.

No trial registration identifier is published for BHB810, so its protocol, doses and enrollment criteria cannot be checked publicly.

Model behavior is evaluated in reviewed venues, on the questions this platform turns on. The published set runs from ICML 2021 through ICLR 2026 and includes Effective Surrogate Models for Protein Design with Bayesian Optimization, Accelerating Bayesian Optimization for Biological Sequence Design with Denoising Autoencoders, Antibody Design with Constrained Bayesian Optimization, Generative Humanization for Therapeutic Antibodies, Guided Sequence-Structure Generative Modeling for Iterative Antibody Optimization, Is Sequence Information All You Need for Bayesian Optimization of Antibodies, Improved Therapeutic Antibody Reformatting through Multimodal Machine Learning, and How To Make the Most of Your Masked Language Model for Protein Engineering. Papers at these venues carry benchmarks and ablations, so the methods behind the platform have been through outside review.

The training data sits in Reccy, whose security, availability and confidentiality controls were examined in a SOC 2 Type 2 engagement by a named assessor, which bears on the integrity of that data.

No results are published for the production system by target class, format or program, and no model card or outside audit of model behavior exists.

Every designed molecule is made and characterized across more than 20 assays before it advances, so a prediction is tested rather than acted on, and Reccy captures each instrument's raw output with metadata and quality control before results reach the models; that system's controls were examined in a SOC 2 Type 2 engagement by a named assessor. A partner can trace a claim back to the measurement behind it.

No success or failure rate with a denominator is published for the design models, and no guarantee or remedy covers a design that does not work.

BigHat works on molecules in its own laboratory and delivers results to partners, so no clinical record system is involved, and the Phase 1 trial of BHB810 runs through its clinical site rather than through any integration BigHat offers. Reccy, the laboratory information management system it names, is its own internal data plane rather than a connection a partner can use.

No connection to a partner's notebooks, data platforms or cloud environments is offered.

The platform runs for BigHat and its partners rather than being deployed to customers: the laboratory and the company are at 1900 Alameda de las Pulgas in San Mateo, California, and the wet lab, the data plane and the models are one stack the company built and operates itself. The data control plane, Reccy, was examined in a SOC 2 Type 2 engagement covering security, availability and confidentiality, infrastructure and continuous integration are described as cloud native, and the company lists a chief information officer among its leaders.

The cloud provider and region holding partner data are not named.

Collaborations, rather than licenses, are what the company sells, and its partnering section routes to a contact form. Where deal terms are published they are specific: the AbbVie collaboration was disclosed at $30 million upfront against up to $325 million in milestones. The Eli Lilly collaboration of April 2025 covers up to two antibody programs with an equity investment and support for BigHat's own gastrointestinal cancer program through Lilly Catalyze360, leaving BigHat full global rights; a second Lilly project in January 2026 has BigHat generating datasets for a developability foundation model inside Lilly TuneLab; and the Merck agreement covered up to three programs and was announced as completed in August 2026.

A $75 million Series C on 24 September 2026 brought total funding to $223 million, following an $80 million Series B in 2022, and three of its pharmaceutical partners, Amgen, Eli Lilly and Merck, are also among its investors. The Amgen and Johnson & Johnson terms are not published.

Coverage is broad across formats and therapeutic areas while staying inside biologics. Formats run to multispecifics, single domain antibodies, single chain variable fragments and antibody drug conjugates, and the company both discovers and engineers, so a partner can bring an existing molecule for optimization or start from a blueprint. Therapeutic reach covers oncology, immunology, inflammation, infectious disease, neuroscience through the Johnson & Johnson work and chronic disease through the Eli Lilly collaboration.

Stage coverage reaches the clinic. BigHat follows a molecule from design toward patients rather than handing off at the bench: BHB810 went from AI generated design through nearly 30 patient and cell derived tumor models to a first patient dosed on 1 September 2026, and BHB299 is planned for the clinic in 2027, both in gastrointestinal and solid tumor oncology.

Clinical reach so far is one asset in one indication with a second behind it, and most partnered programs disclose no indication.