Platform
EMMI Experimentation meets machine intelligence.
AI changes the world when it knows the rules and has the data. In small molecule drug discovery, neither exists. The EMMI platform creates the data and discovers the rules.
>90%
Of difficult targets screened yielded unique chemistry starting points
8
AI-driven programs progressed for undruggable targets in the past 4 years
2x
Faster than traditional approaches at evolving a novel chemotype
100%
Decisions guided by full-stack AI for small molecules
A platform built for working in the dark areas of chemical space, to solve problems others cannot, to shave years off development and turn novel molecules into real medicines.
EMMI Workflow
We’ve built a chemistry-first, AI-native platform for small molecule discovery that turns impossible problems into inevitable solutions.
01
Find unique starting points for even the hardest challenges.
By navigating billions of possibilities in places where others can’t venture, we see patterns emerge and find hits, particularly for targets with no known starting points.02
Generative AI designs optimized molecules.
Our 3rd-generation foundation model lets us computationally speak the language of chemistry. This allows our advanced generative AI models to iteratively explore synthetically accessible molecules in the dark areas of chemical space broader, deeper, and faster.03
Predictive and Selection Models ensure we make only the best.
By combining Terray's industry-leading universal potency model with a broad suite of ADME models, we characterize every molecule in silico and use a hedged strategy to pick the right molecules to make, ensuring efficient program progression.04
Automation closes the loop.
Our custom software layer connects AI design to lab automation, making it easy to test the right molecules in the right experiment, feeding iterative data back into our models to drive the next round of design.How EMMI Works
AI driving discovery in the hands of every scientist, every day.
Over 16 billion unique measurements, growing by 1 billion every quarter.
We measure millions of molecules simultaneously on our patented, ultra-dense microarray. Our proprietary target-molecule interaction dataset is 50x larger than all public chemistry data — and growing.AI can only design what it can understand. That's why we built COATI, our foundation model trained on 1 billion molecules. It encodes the full complexity of chemical structures into a mathematical space that our AI can reason within, manipulate, and decode back into real molecules. It’s what makes it possible for us to explore new areas of chemical space with precision.
COATI is multi-modal, contrasting molecular representations across strings, graphs, and 3D structures — the same contrastive learning approach behind breakthroughs like DALL-E, applied to chemistry. It's not a general-purpose tool adapted for chemistry. It's chemistry, from the ground up.Every drug starts as an idea — a virtual molecule that doesn't yet exist. Generating the right candidates efficiently is where most AI drug discovery falls short.
Terray's generative models are trained on our proprietary datasets and purpose-built to propose molecules with specific properties from the start. Our latest reinforcement learning-based approach doesn't just suggest promising structures — it prioritizes ones that can actually be synthesized, without sacrificing performance.Every molecule must be characterized in silico — for the full suite of properties that determine whether a drug can succeed in clinical testing. From potency to selectivity to ADME, we have built proprietary models for every endpoint, and our universal potency model, TerraBind, is up to 20% more accurate and 26x faster than comparable models.
TerraBind works in two stages: first, scoring every generated molecule with an ultra-fast sequence-only potency model that runs 20,000x cheaper than public models, then applying a deeper structural analysis to the most promising candidates. Our focus on “potency over pose” results in precise potency predictions at a fraction of the cost and time of other industry models.Choosing which molecules to synthesize and test is one of the most consequential decisions in drug discovery. Our models surface thousands to millions of promising candidates in every iteration. Narrowing those down to the optimal handful requires something most AI platforms don't offer: an honest accounting of uncertainty.
EMMI Select doesn't just rank molecules by predicted performance. It quantifies the confidence behind every prediction, then uses that uncertainty as an input to select the best molecules to make and test in the lab. The result is a selection method that outperforms conventional approaches by 3x in benchmarking, saving significant time and cost.Our automated experimentation closes the loop between experimentation and AI — continuously, at scale, and without the bottlenecks that slow traditional discovery down.
From our ultra-dense microarray, to compound synthesis, to assays, to model retraining, each piece comes together, interwoven and synchronous, to feed the next design cycle. The loop closes faster here, and every time it does, our models get smarter.Pipeline
EMMI delivers results
EMMI is used every day for our internal and partnered pipelines. We learn what works, and we constantly refine.