BigHat Biosciences

The models are the reason every one of these partnerships exists. Milliner is not analysis bolted onto a discovery operation; the wet lab was built to serve the design loop, synthesising and characterising hundreds of variants weekly specifically so the models can be retrained on fresh measurement.

The company's founding thesis, stated by both founders, was that machine learning had to move from the end of the research pipeline to the front of it, and the platform is that argument built out. The Synaffix conjugate chemistry licence is payload technology sitting downstream of the design engine and does not dilute the grade.

The oversight is physical and the company describes it in its own voice, which is why this sits above the lane. Nothing the models propose is treated as a result until it has been built and measured: hundreds of variants are synthesised, purified and characterised for biophysics and function every week, and the models are updated from 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 criticised developability models that generalise poorly because they were trained on inconsistent data. A company willing to say in public that models in its own category are used for ranking rather than for improvement is describing the limits of its own tooling. What remains undocumented is the human layer: who decides which designs are built, and what happens when a model and an assay disagree.

The first top grade on this axis in this lane, and it rests on something no neighbour offers: a dedicated publications page carrying eleven items from 2021 through 2026, at ICML, ICLR and NeurIPS, describing the actual methods rather than gesturing at them.

Bayesian optimisation for biological sequence design, denoising autoencoder surrogates, constrained optimisation for antibody design, generative humanisation, guided sequence and structure modelling, masked language models for protein engineering, and a unification of discrete, Gaussian and simplicial diffusion. The lineage is continuous and current rather than historical, so a reader is looking at the methods behind the system as it runs now, not at its ancestry.

Two advisors are among the researchers whose Bayesian optimisation work the papers build on, which makes the technical claims traceable to a named body of public work. Withheld: no model card, no weights, no training data composition, and no description of Milliner as configured in production. Also worth checking rather than assuming, the venue labels are the company's own and at least one entry is a workshop track.

The best disclosure of this kind in the lane, on the strength of three 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 used in the lead programme. The Lilly TuneLab arrangement discloses the outbound direction, with BigHat generated datasets feeding a foundation model the company does not own.

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. Still withheld: compute and hosting providers, training data provenance for the internal models, open source component inventory, and any third party model licence.

The deepest evidence base in this lane by volume, held below the top grade by what the evidence is about. Eleven publications across 2021 to 2026 at ICML, ICLR and NeurIPS establish that the methods work as machine learning; five named pharmaceutical partners establish that sophisticated buyers keep paying; the AbbVie agreement carries a disclosed $30 million upfront, which is money actually received rather than a milestone ceiling. What is missing is evidence at the level of a drug.

No molecule has entered a human, the publications report algorithmic advances rather than therapeutic outcomes, and the one partner statement attesting to a completed program outcome carries a conflict that has to be recorded: Amgen's vice president of research praised the platform for significantly optimising single domain antibodies against the original repertoire, and that quote appeared in the same funding announcement in which Amgen Ventures invested. It may well be accurate. It is not independent. What would move this grade is the lead conjugate reaching the clinic, or a non investor partner attesting on the record.

No protected health information is in scope today. The partner separation question is live here in its sharpest form anywhere in this lane, because of what the platform is for. Five pharmaceutical partners run programs through a system whose stated advantage is that models are continuously retrained on laboratory feedback, and several of those partners compete directly.

The January 2026 Lilly project goes further still: BigHat generates datasets specifically to train a generalisable antibody developability foundation model, and generalisable is the whole point of it, meaning a model built to carry what it learns beyond the context that produced it. Nothing published states whose data enters that model, what a partner's measurements are permitted to train, how programs are segregated, or what survives the end of a collaboration. The architecture makes the question unavoidable and the disclosure does not reach it.

Graded neutrally because the obligation has not yet arisen. The platform operates on protein sequences, structures and laboratory assay measurements, with no protected health information in scope and no covered entity relationship.

Recording a finding in advance so a later reader does not have to rediscover it: this vendor states it is moving a conjugate into the clinic during 2026, and a company running its own trials will handle patient data, which is a different posture from the one graded here. Nothing published sets out what that posture will be. If a trial opens, this axis should be regraded on the arrangements then in force rather than left at a neutral mark earned while the company was preclinical.

The thinnest legal surface graded in this lane, and the lane was already thin. Two dedicated passes on 29 August 2026 across a targeted domain search and a direct review of all eight navigation sections and the site footer found no security page, no trust centre, no certification, no attestation, no penetration testing statement, no incident policy and no privacy policy or terms of any kind. The footer carries a postal address, a navigation list, two social links and a copyright line.

Every neighbour in this lane publishes at least one legal document, however narrow. This company publishes none, while holding programme data for five large pharmaceutical partners and listing both a general counsel and a chief information officer among its leadership. The absence is not for want of someone to write it.

This is the first record in this lane where the axis has something real to measure and still cannot express it. BigHat is not a company with nothing to clear: it states a gastrointestinal cancer conjugate entering the clinic during 2026 and a T cell engager with an investigational new drug application in the same year, and it has in licensed conjugate technology to support the lead programme.

But as of 29 August 2026 two dedicated passes located no clearance, no trial registration, no first dosing announcement and no designation, and a third party pipeline database still places the lead programme at the enabling stage rather than in the clinic. No regulatory strategy, manufacturing readiness statement or designation filing is published. So the obligation is imminent and unincurred, the plan is stated and unconfirmed, and the grade lands in the same place as a company with no pathway at all.

No governance framework, model card, responsible use policy or evaluation charter appears across the eight sections of the company website as of 29 August 2026. What partially stands in its place is stronger than elsewhere in this lane. Eleven peer reviewed papers describe evaluation methodology in the detail those venues require, including how uncertainty is handled in the Bayesian optimisation work that runs through the whole programme.

And the chief executive has publicly named a limitation of the model class her own company works in, stating that antibody developability models generalise poorly to new sequences because they are trained on limited formats and inconsistent data. Published method scrutiny and a public statement of where models fall short are the behaviours governance is meant to produce, but they are not a policy, and no reader can tell from them how the company decides a model is fit to run a programme.

There is nothing to grade. No terms of service, no warranty position, no limitation of liability, no indemnity, no intellectual property allocation and no service level exists on any public surface as of 29 August 2026, because the site carries no contractual document at all.

For a company whose commercial model is designing molecules a partner will own and develop, the allocation of rights and risk over a machine designed sequence is the central legal question of the business, and it is settled entirely inside private agreements.

The one visible fragment is a fact of a different kind: the Lilly arrangement was announced with BigHat retaining full global rights and development control over its own conjugate programme, which tells a reader how one negotiation went and nothing about how any other would.

No electronic health record surface exists and none would be appropriate for a preclinical biologics company, so this is graded neutrally under the convention for this lane. Recorded in its place, because it is unusually substantive here: the integrations this vendor has are laboratory and scientific rather than software.

Site specific conjugate chemistry was in licensed from Synaffix in November 2024 and applied to the lead programme, and the January 2026 Lilly project routes BigHat generated datasets into an external foundation model inside Lilly TuneLab. Those are real technical dependencies in both directions. What does not exist is any application programming interface, connector or licensable software, because the platform is not sold as a product.

Nothing is published: no hosting arrangement, no tenancy model, no residency commitment, no retention or segregation statement, and no subprocessor list. Graded neutrally under the convention for this lane, which is now the fifth consecutive record where the convention rather than the evidence is setting this grade. One detail makes the silence more pointed here than at the neighbours.

This company lists a chief information officer on its leadership page, which is rare at roughly 80 people and means the function is owned by someone senior internally. Whatever arrangements exist are therefore deliberate rather than absent, and none of them has been made visible outside.

Sits at the top of this band on substance and is held there by hygiene. On substance it leads the lane: five pharmaceutical partners named rather than described, an actual upfront figure separated from a milestone ceiling at $30 million against $325 million with AbbVie, twelve investors named including two pharmaceutical companies, a partner equity investment disclosed, an in licensing transaction with the licensor and date stated, and an acquisition acknowledged.

Nobody else graded in this lane separates an upfront from a ceiling. Against that, the company's own About page reports more than $100 million raised while its January 2026 announcement says more than $140 million, so the published figures disagree with each other. The same page carries unreplaced placeholder Latin text in its awards section and points several award links at a legacy domain the company no longer operates. None of that is deception, but a reader checking the numbers finds two answers.

Materially wider than the rest of this lane 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 optimisation or start from a blueprint. Therapeutic reach covers oncology, immunology, inflammation, infectious disease and neuroscience through the Johnson & Johnson work.

The stage coverage is what earns the grade: named leadership in discovery medicine, preclinical development and program management means this company follows a molecule toward an investigational new drug application rather than handing off at design, which none of its neighbours here do.