By Terray TherapeuticsLearning Objectives
- Understand why owning the measurement hardware changes what an AI chemistry model can learn
- Understand the partnership-plus-pipeline model these companies run on
- Evaluate vendor-reported model benchmarks with the right amount of caution
What Is Terray Therapeutics?
Terray Therapeutics is an AI-native, chemistry-first drug-discovery company built around EMMI, which stands for Experimentation Meets Machine Intelligence. The distinguishing choice is that Terray builds its own measurement hardware rather than training on public data alone: a proprietary ultra-dense microarray that measures how millions of small molecules interact with a target, feeding a fully automated laboratory. Terray reports that this system has produced more than 13 billion unique binding measurements, which it describes as the largest global database of small-molecule binding data, growing at roughly one billion measurements per quarter.
That data advantage is the point. Most AI chemistry models are limited by the sparse, heterogeneous and often unreproducible binding data available publicly. A company generating its own measurements at that volume is training on a fundamentally different substrate, which is the argument for the whole vertically integrated approach.
⚠️Warning
The platform is EMMI, not tNova. Coverage of Terray's 2024 Gilead collaboration names tNova, because that was the platform then. Terray introduced EMMI in November 2025, and EMMI is the current name. Writing tNova in the present tense is a common error carried forward from the older reporting.
💡Key Concept
Why owning the assay matters: a model of molecular binding is only as good as the measurements behind it. Public binding data is assembled from many labs with differing protocols, so much of the signal a model learns is laboratory artifact. One instrument running one protocol at very high volume removes that noise.
Inside the EMMI Platform
COATI
Terray's chemistry foundation model, built as an invertible latent space for molecules and trained with contrastive learning. The company compares the training approach to how DALL-E is trained, contrasting representations such as SMILES strings, molecular graphs and three-dimensional conformers. COATI3 was trained on over one billion diverse molecules and uses a 768-dimensional representation.
TerraBind
The potency prediction model, run in two tiers. A very fast sequence-only model scores molecules first, and a structure-based multi-modal model then refines the survivors using a representation derived from COATI3 together with a protein language model.
EMMI Select
Chooses which molecules are actually worth synthesising, combining Epistemic Neural Networks with an acquisition function so that model uncertainty informs the decision rather than raw predicted score alone.
Generative Design
Two approaches run side by side: a latent diffusion method using classifier guidance, and a newer reinforcement-learning approach using policy-gradient algorithms, which Terray reports produces more synthetically accessible molecules.
📝Note
The performance figures are Terray's own. In February 2026 the company announced an EMMI prediction model it says delivers 20 percent higher accuracy than leading open-source models while running 26 times faster and reducing inference cost by 96 percent. Those are vendor-reported figures from Terray's own announcement, the comparison set is not named publicly, and they have not been independently reproduced. Its earlier reported efficiency claims for the two-tier TerraBind approach come from the same source.
🎯Tip
Visit Terray Therapeutics: terraytx.ai — note the domain moved from the older terraytx.com address.
Pricing
Terray is a drug-discovery company rather than a product a reader can buy. There is no self-serve access and no published pricing; it earns through pharma collaborations and its own pipeline, which is the standard shape across this category.
- EMMI platform applied to partner targets
- Gilead, Bristol Myers Squibb, Calico
- Collaboration terms, not published
- Shared programs
- Odyssey Therapeutics
- Terms not published
- Terray's own immunology programs
- Wholly owned
- No external access
Strengths
- Owns its data generation — proprietary microarray plus an automated lab
- Measurement scale — more than 13 billion binding measurements reported
- Full stack — hardware, dataset and models designed together
- Serious validation — collaborations with Gilead, Bristol Myers Squibb and Calico
- Investor signal — NVIDIA participated through its NVentures arm
Limitations and Considerations
- Not a usable product — no self-serve access and no published pricing
- Benchmarks are vendor-reported — accuracy and speed claims are Terray's own
- Designs are hypotheses — synthesis, testing and the clinic still decide
- Partnerships are not approvals — a collaboration is not a drug reaching patients
- Binding is not efficacy — the measurement advantage is upstream of the hard clinical questions
Best Use Cases
| Use Case | Why Terray Matters | Caveat |
|---|---|---|
| Understanding vertically integrated AI chemistry | Owns hardware, data and models | Not a tool anyone can buy |
| Tracking data moats in drug discovery | 13 billion proprietary measurements | Scale is self-reported |
| Comparing AI pharma platforms | Direct peer of Genesis Therapeutics | Both are partnership-only |
| Following chemistry foundation models | COATI3 trained on a billion molecules | Benchmarks not independently reproduced |
Key Takeaways
- Terray Therapeutics is an AI drug-discovery company whose EMMI platform stands for Experimentation Meets Machine Intelligence
- Its distinguishing choice is owning the measurement layer, with a proprietary ultra-dense microarray reported to have produced more than 13 billion binding measurements
- EMMI combines the COATI chemistry foundation model, TerraBind for potency prediction, EMMI Select for uncertainty-aware molecule choice, and generative design models
- It runs discovery collaborations with Gilead, Bristol Myers Squibb and Calico, plus co-development with Odyssey Therapeutics, and counts NVIDIA among its investors
- Its accuracy and speed figures are vendor-reported, and like every platform in this category its designs remain hypotheses until synthesis and testing decide
