Terray Therapeutics

The company's own summary of itself is that it invents its own hardware, builds its own datasets and designs its own AI, and the three are not separable. The microarray generates measurements; the measurements exist to train models; the models decide what gets made next; the automation closes the loop back to the array. Terray states that 100 percent of its small molecule decisions are guided by what it calls full stack AI.

Remove the models and what remains is a very large binding dataset with no design engine attached to it. The named components make the dependence concrete: COATI encodes chemistry into a latent space that can be manipulated and decoded back into real molecules, generative models propose synthesizable candidates from it, TerraBind ranks them for potency, and a selection model with uncertainty quantification decides which handful to synthesize. There is no version of this business that ships a molecule without that chain.

Oversight rests on a real mechanism, and the mechanism is uncertainty rather than a slogan. The selection model does not simply rank molecules by predicted performance: it quantifies the confidence behind each prediction and uses that confidence as an input, running what the company calls a hedged batch selection strategy, and the accompanying preprint reports that this yields six times greater affinity improvement in simulated discovery cycles. Every prediction is then checked physically, because the automation layer synthesizes and assays the selected molecules and feeds the result back into the models. The company frames its tools as instruments in the hands of its own scientists rather than as an autonomous designer.

One part is missing and it is the human one. Nothing published states the point at which a person must decide rather than may, what threshold stops a program, or what happens after a prediction is wrong beyond the model being retrained on the result. The claim that 100 percent of small molecule decisions are guided by full stack AI describes uniform handling rather than a correctness guarantee, so it reads as a scope statement. It sits awkwardly beside the copilot framing, and nothing reconciles the two.

Very little is held back about how these models are built, trained and updated.

The models are named and identified rather than gestured at. COATI is the chemistry foundation model, a multi-modal encoder-decoder trained by contrastive learning across SMILES strings, 2D graphs and 3D conformers, with a SMILES decoder and a 768 dimensional latent space; it is published in the Journal of Chemical Information and Modeling (doi 10.1021/acs.jcim.3c01753) and its code is public. TerraBind is the potency model, published as a February 2026 preprint, and its composition is stated: a Pairformer trunk over a coarse pocket representation of protein C-beta atoms and ligand heavy atoms, combining COATI-3 molecular encodings with ESM-2 protein embeddings, with a diffusion free pose optimization module and a separate affinity likelihood module.

Training data is characterized. COATI-3 is trained on more than one billion diverse molecules. The experimental corpus behind the platform is stated at over 17 billion unique target-molecule interaction measurements on the company's own ultra dense microarray, growing by a billion each quarter.

Versioning and update practice is published too: the company states it has moved from COATI1 to COATI3 over two years, names COATI-3 as the version inside TerraBind, describes the platform's generative model as third generation, and publishes dated model releases. The latent diffusion generator and the selection method are also released as code. What a reader still cannot see is a release note tying a specific model version to a specific date of change in production.

The model layer is disclosed in detail. The models are built in house and named, and the one external model component inside them is named too: TerraBind combines Terray's own COATI-3 encodings with ESM-2 protein embeddings, an outside protein language model. The comparators and adjacent systems the company builds against are named as well, including Boltz-2, AlphaFold, NeuralPlexer, SeedFold and OpenFold. Code for COATI, the latent diffusion generator and the selection method is published, so what produces the output can be inspected rather than inferred.

The compute layer is named: Terray states it worked with NVIDIA and trained the COATI family on DGX Cloud, and NVIDIA's venture arm is an investor in the Series B, a financial tie that sits alongside the technical endorsements from NVIDIA in the company's own announcements.

What is not published is a complete party list with terms. There is no subprocessor list, no statement of what governs partner chemistry once it reaches the platform or the compute provider, and no disclosure of the instrument, synthesis or assay suppliers behind the experimental engine. For a company whose business is running discovery on other pharmaceutical companies' targets, that last gap is the one that matters.

There is no clinical evidence here, because there is no clinical asset. What there is instead is published research with the method available to read, stopping short of independent validation.

The chemistry foundation model COATI is published in the Journal of Chemical Information and Modeling (doi 10.1021/acs.jcim.3c01753) with its code released. The potency model TerraBind is described in a February 2026 preprint with its architecture, its training data and its evaluation protocol, reporting a 16 to 20 percent improvement in Pearson correlation over Boltz-2 on the public CASP16 benchmark and on a proprietary set of more than 25,000 assay readouts across 18 targets, winning on 15 of them, alongside pose accuracy against FoldBench, PoseBusters and Runs N' Poses. Named partner programs give a second, weaker line of evidence: a discovery milestone with Bristol Myers Squibb announced in December 2025, a multi target research collaboration with Gilead from December 2024, and a transcription factor collaboration with Odyssey Therapeutics from October 2024, each a third party staking work on the platform.

The weakness is the gap between benchmark and medicine. No molecule from this platform has entered a clinical study, no development candidate or program is named on the pipeline page, and the company's own 2024 guidance of a first compound in the clinic by 2026 has no published clearance or trial start behind it. The platform level figures that describe program outcomes rather than model accuracy, more than 90 percent of difficult targets yielding unique starting points and eight programs progressed for undruggable targets in four years, carry no denominator, comparator or definition of a difficult target.

No patient data enters this platform. The inputs are chemical and experimental: binding measurements generated on the company's own microarray, protein sequences and structures, and public chemistry. There is no clinical program, no trial and no record feed, so protected health information does not arise in its usual form.

What is published is a general privacy notice, last updated 10 August 2026, covering the company's websites, sales, marketing and events. It is substantive on personal information, setting out categories collected, third party analytics and security partners, disclosure situations and the criteria used to determine retention. It does not state a retention schedule, and it says nothing about the data that matters most here.

That gap is partner chemistry. Terray runs discovery against other pharmaceutical companies' targets under collaborations with Bristol Myers Squibb, Gilead and Odyssey Therapeutics, and nothing published states whether partner structures, targets or assay results are retained, whether they are used to train the models that serve other programs, or how they are separated between programs. The models improve by feeding every experimental result back into training, which makes the question sharper here than at a company whose models are static. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

A substantive privacy document exists, and no status is stated anywhere. The notice was last updated on 10 August 2026 and is current. It defines its scope as the company's services, naming both of its websites, sales, marketing and events, and any other related service it operates, and it covers the ground a privacy document is expected to cover: categories of information, processing purposes, third party disclosure, retention criteria, security, and rights under United States state laws and the European regime.

What it does not do is take a position. Business associate status is neither claimed nor disclaimed, protected health information is not mentioned, and no statement says whether the platform processes patient data or not. That statement, plainly made and explained, is what is missing. On the facts the answer would be simple, since this is a preclinical chemistry platform with no clinical program, no trial and no record feed, and saying so would be enough.

No certification, attestation, trust center or security page is published. There is no SOC 2 report, ISO/IEC 27001 certificate or HITRUST certification, and no independent penetration testing on a stated cadence. The only security material is a section of the privacy notice, which states that the company has implemented reasonable technical and organizational security measures and that no transmission over the internet can be guaranteed secure. These are controls asserted with nothing independent behind them.

The exposure is commercial rather than clinical, and it is real. Terray holds target and chemistry data from Bristol Myers Squibb, Gilead and Odyssey Therapeutics under research collaborations, and trains models on a proprietary corpus it describes as the largest binding dataset of its kind. Nothing published describes the controls protecting either, and that is the first assurance a prospective partner would need.

As a preclinical small molecule company, Terray makes no device claim and scopes the platform accordingly. The discovery platform is not sold as a product, is not presented as a medical device, and carries no clearance. That is the appropriate position for what it is.

There is no regulatory milestone of any kind. No investigational new drug clearance, no clinical trial start and no development candidate is named, and the pipeline page describes internal work in immune system diseases and partnered work across broad therapeutic areas without naming a single program. The company told the trade press in October 2024 that it expected its first compound in the clinic by 2026, and nothing published confirms that happening. The guidance remains outstanding rather than met.

Terray publishes its model evaluations broken out by the group its models actually serve, which is protein targets, and it publishes where they lose.

TerraBind, the potency model, is evaluated on two held out sets: the public CASP16 binding affinity challenge and a proprietary set of more than 25,000 assay readouts spanning 18 distinct targets. Against Boltz-2 as the named baseline it reports a 16 to 20 percent improvement in Pearson correlation, and it states the per target result rather than only the aggregate: it beat the baseline on 15 of the 18 targets, which tells a reader it did not on three. Pose accuracy is assessed separately against three named public benchmarks, FoldBench, PoseBusters and Runs N' Poses. The model's affinity uncertainty estimates are reported as well calibrated, which is a checkable claim about when the model should not be trusted rather than a claim that it is right. The accompanying preprint gives the architecture and the evaluation protocol in full.

The chemistry foundation model behind it, COATI, was published in the Journal of Chemical Information and Modeling (doi 10.1021/acs.jcim.3c01753) and its code is released publicly, as are the latent diffusion and selection codebases. Outside groups can therefore reproduce the work rather than take the figures on trust.

What is not published is a governance framework or an independent audit: the evaluations are the company's own, run against public benchmarks and released code. No bias evaluation in the demographic sense applies to a model of this kind, and the composition of the 17 billion measurement internal training corpus is characterized by size rather than by chemical or target coverage.

Performance is published in falsifiable form, with its method, its denominator and its named comparator. TerraBind's accuracy is stated as a 16 to 20 percent improvement in Pearson correlation over Boltz-2 across two held out sets, one of them the public CASP16 benchmark and the other a proprietary set of more than 25,000 assay readouts across 18 targets, with the per target win rate given as 15 of 18. The speed figure is stated with the hardware and settings behind it: a 26 fold reduction in inference time per complex on a single A6000 GPU at 196 tokens and 10 samples. The model's affinity uncertainty estimates are published as well calibrated, and the code for the foundation model, the latent diffusion generator and the selection method is released, so the claims can be tested by someone outside the company.

What is absent is any commitment running toward a partner or anyone else. Nothing published states what Terray is answerable for when a predicted molecule fails, no guarantee, indemnity or remediation obligation attaches to a design delivered under a research collaboration, and no correction route is described for a partner who acts on a wrong prediction. Several headline platform figures also stand without the method the model papers supply: more than 90 percent of difficult targets screened yielding unique chemistry starting points, a selection method beating conventional approaches by 3 times in benchmarking, and better than 2 times faster evolution of a novel chemotype are each published without a denominator or a named comparator set.

The platform serves Terray's own programs and its pharmaceutical partners' discovery programs rather than health systems, so no clinical record integration exists. There is no electronic health record touchpoint, no clinical workflow surface and no provider connection.

On the research systems of record a partner would run, nothing is named either. No electronic laboratory notebook, laboratory information management system, research data platform or partner cloud environment is named as connected, and no interoperability standard or data exchange method is described for the collaborations with Bristol Myers Squibb, Gilead or Odyssey Therapeutics, where target and result data necessarily moves between two organizations. The integration Terray does describe is internal: a custom software layer connecting model output to its own laboratory automation, synthesis and assay systems, which is a closed loop inside one company rather than a connection others could use.

The platform runs for Terray and its partners rather than being deployed to a customer, so there is no tenancy model, hosting option or isolation commitment for anyone to evaluate. One element is published: the company states that it worked with NVIDIA and trained the COATI family of models on DGX Cloud, which names a single compute environment with its location implied rather than committed. The experimental half of the platform is physical and sits in the company's own facility in Los Angeles.

Nothing published states where partner data rests, whether processing crosses a border, or how one collaboration's data is separated from another's. Three named pharmaceutical collaborations run on this infrastructure, and none of them has a published residency or isolation position attached.

Nothing is sold and nothing is licensed, so there is no price and no cost for anyone to size. The business model is discoverable: Terray earns through pharmaceutical research collaborations and its own internal pipeline, and the page inviting partners gives an address to contact rather than a rate card or a deal structure.

The partners are named and dated: a discovery milestone achieved with Bristol Myers Squibb in December 2025, a multi target research collaboration with Gilead from December 2024, a transcription factor collaboration with Odyssey Therapeutics from October 2024, and an earlier multi target collaboration with Calico announced in October 2022. No upfront payment, milestone value, royalty rate or term is disclosed for any of them, so the economic weight of the partnership book cannot be judged from published material. Capital is disclosed in full: a seed round, a $60 million Series A led by Madrona in 2022, and a $120 million Series B in October 2024 led by Bedford Ridge Capital and NVentures, with total equity financing stated at more than $200 million.

Coverage is stated clearly and nothing validates it yet. The internal pipeline is described as de novo medicines for immune system diseases, and the partnered work as de novo medicines across broad therapeutic areas, with the named collaborations pointing at age related disease, transcription factor targets and undisclosed targets for two large pharmaceutical companies. No indication, program or target is named on the pipeline page, so a reader cannot establish where the platform has been shown to work as against where it is being applied.

The platform level scope claim is about target difficulty rather than therapeutic area: Terray positions itself against targets with no known starting points and states that more than 90 percent of difficult targets screened yielded unique chemistry starting points, and that eight programs for undruggable targets were progressed in four years. Neither figure comes with a denominator or a definition of a difficult target, and no program from either group has reached a stage that would validate the claim.