Terray Therapeutics

Generative modelling is the product, and the company is unusually clear about the dependency running the other way as well. The COATI foundation model and the latent diffusion design work sit at the centre of tNova, and the experimental platform exists to feed them rather than the reverse. What makes the grade unambiguous is that the scale claim only means anything through a model: five billion measured interactions is not a resource a chemist can read, it is a training corpus. Remove the models and what remains is an assay factory with no way to act on its own output.

The stated position is modest and specific in a category prone to grandiosity: the tools are described as co pilots that the company's medicinal chemists use every day to explore chemical space, which places a human at the design decision rather than at the end of it. The structural check is stronger still.

An organisation that can make several million molecules a month tests model output empirically at a rate that makes an unchecked prediction unlikely to survive, so oversight here is enforced by the experimental loop rather than by policy. It falls short of the top grade because no published process describes where a chemist must intervene, what a model output is permitted to decide on its own, or how disagreement between prediction and assay is resolved.

The strongest transparency record in this lane, because the company published its foundation model rather than describing it. COATI appeared in a peer reviewed journal with a digital object identifier, and the paper specifies the architecture as a pre trained multi modal encoder decoder of druglike chemical space trained by contrastive learning over text and three dimensional molecular representations, enumerates the properties claimed for the embedding, and demonstrates a generative optimisation method on a named model protein against an explicit multi parameter objective of potency, solubility and druglikeness.

That is a disclosure another group could interrogate and attempt to reproduce, which is the standard this axis is meant to reward and which almost nobody in this category meets. The proprietary corpus itself is not released, which is expected and is the reason the model claims can be examined while the data claims cannot.

Better than the lane norm because the principal model is identified, named and published, so a reader knows exactly what the central component is and who built it rather than inferring it. Nothing external is named: no framework, library, third party model or compute partner appears in published material. One inference should be resisted, that NVIDIA's venture arm leading the Series B implies a disclosed technology dependency. An investment is not a supply chain disclosure, and none has been made.

Partnership evidence is strong and clinical evidence does not yet exist. Multi target agreements with Bristol Myers Squibb and with Calico mean two sophisticated organisations examined the platform and committed to it, which is worth more than most self reported metrics in this lane. Against that, the pipeline is preclinical.

The company told the trade press in October 2024 that it expected its first compound in the clinic by 2026, and no investigational new drug application or trial initiation was located in this assessment. The platform metrics are also entirely self reported and structurally hard to check: nobody outside the company can verify five billion measurements or the claim that this is fifty times all public chemistry data, and the comparison depends on which public sources are counted.

Not applicable, and this is the clearest instance of it in the lane. The training corpus is chemical binding measurements generated by the company's own microarray platform, so it contains no human derived material at all, unlike the single cell and multi omics companies whose data originates with donors and patients.

The stewardship question here is commercial confidentiality around partner chemistry rather than personal data, and nothing published describes how partnered target information is segregated from the shared corpus, which is the analogous question worth asking.

Not applicable, and more cleanly here than elsewhere in this lane. There are no covered entity customers, no protected health information, and no clinical programme yet, so no human subject data is being handled under any regime. The qualification recorded against other companies in this category, that the not applicable convention strains once a company enters human trials, does not currently apply. It will if the stated clinical ambition is realised.

No certification, attestation or security page was located. The exposure profile is narrower than for companies handling human data: there is no protected health information anywhere in this platform, so the assets at risk are the proprietary chemistry corpus and partner target information rather than personal data.

That second category is not trivial, since two pharmaceutical partners have entrusted target chemistry to the platform and such arrangements normally carry contractual security obligations, none of which are described publicly.

There is no regulatory status yet, which is the honest description of a preclinical company rather than a criticism of one. No investigational new drug application, clinical trial registration or regulatory milestone was located. What holds this at the middle band rather than lower is that the company has stated a clear public expectation, first compound in the clinic by 2026, which is a checkable commitment rather than an open ended intention. That expectation is now due and unconfirmed as of this assessment. Nothing published describes a regulatory strategy for the internal immunology programmes or which asset is intended to go first.

Higher than the lane norm on a narrow but real basis. The COATI paper states that the model is constructed without human biasing of features, which is a direct claim about a specific and well understood failure mode in cheminformatics, made in a peer reviewed venue where it can be examined rather than in marketing copy. That is bias disclosure in the sense that matters for this domain. Two limits keep it below the top grade.

It is a methods claim rather than a governance framework, and there is no published policy on model monitoring, drift or review. And the coverage question is untouched: five billion measurements is a large corpus and nothing describes which regions of chemical space it represents, so the blind spots of a model trained on it are unknown.

Lower than the clinical stage companies in this lane for a structural reason rather than a disclosure one. Where a company with an asset in trials can point to regulatory clearance, review board oversight and consent as an external accountability apparatus, a preclinical company has not yet engaged any of it, so nothing currently stands behind a model driven decision except the company's own judgement. Nothing published describes internal accountability for a wrong selection, and no person is yet exposed to the output, which is why this sits at the middle band rather than the floor.

Not applicable. No electronic health record touchpoint, clinical workflow surface or provider integration exists. The research analogue is the coupling between the microarray experimental platform and the modelling layer inside tNova, described as an iterative loop in which each cycle of design and experimentation trains the models that guide the next, and that integration is internal rather than a connection to any third party system.

Not applicable in the customer sense. tNova is operated internally and is not licensed, hosted or deployed for buyers, so there is no tenancy, region or residency commitment to evaluate. The platform is stated to power partnered programmes as well as internal ones, which implies data flows to and from two large partners, and nothing published describes how that is structured or where partner derived data resides.

Not applicable in the vendor sense, since nothing is sold and there is no price. The available disclosure is capital and partnering, and it is complete: a seed round, a $60 million Series A led by Madrona in 2022 and a $120 million Series B in October 2024, with lead and participating investors named in each and a cumulative equity figure above $200 million stated. Both pharmaceutical partners are named rather than described as undisclosed, which is not universal in this category. No economics are disclosed for either partnership, so their materiality cannot be assessed.

Narrower than most records in this lane, and deliberately so. The internal pipeline is immunology and no other therapeutic area is claimed for it. Partnered work is described only as challenging targets across a range of diseases, with no indications named, so the apparent breadth of the platform rests on programmes whose subject matter is not public. That is a defensible commercial position and it leaves a reader unable to establish where the technology has actually been applied beyond one therapeutic area.