Terray Therapeutics has unveiled details of its integrated AI–experimental discovery engine, EMMI, alongside a new generation of Select models designed to prioritize which candidate molecules should advance into synthesis and testing. The platform reflects the company’s ongoing effort to merge high-density experimental workflows with large-scale machine learning in order to accelerate early-stage drug discovery.
EMMI, an acronym for Experimentation Meets Machine Intelligence, brings together Terray’s proprietary ultra-miniaturized assay hardware, a highly automated wet-lab environment, and a full-stack computational suite. According to the company, the system is underpinned by what it describes as an advanced chemistry foundation model capable of supporting generative, predictive and selection-based tasks. In practice, these models enable rapid in silico evaluation of very large molecular libraries, followed by targeted experimental testing that feeds new data back into the models.
While we’re excited by what EMMI can do today, we’re even more optimistic about what the future holds. Because EMMI is used every day across our internal and partnered programs, iterative experimentation and computation continually generate insights that we incorporate directly into the platform. This systematic improvement moves us toward our vision of a closed-loop process driven by advanced reasoning models.
Narbe Mardirossian, Chief Technology Officer
Terray reports that it is advancing a portfolio of immunology-focused drug candidates while also collaborating across other therapeutic areas. The company’s discovery efforts draw on a large internal dataset comprising more than 13 billion measured target–molecule interactions. Within EMMI, generative models propose new structures, predictive models estimate their property profiles, and the new selection models determine which subset should be synthesized and assayed. Terray states that this tri-layer workflow, when benchmarked, reduced cycle time and cost for iterative lead optimization by approximately a factor of three.
For more, see Terray’s recent blog posts on EMMI and specifically on the newest Select models, or below for an abbreviated version.
Terray’s Data Advantage
Terray has amassed the world’s largest precise chemistry dataset powered by the Company’s proprietary ultra-dense microarray, which is generating more than 1 billion new measurements a quarter. With 13 billion target-molecule measurements to date, Terray is able to find novel chemistry where others simply cannot. This data edge has enabled EMMI to find structurally-novel molecules for every program at Terray and then deploy its AI-models to efficiently optimize in these newly-identified areas of chemical space.
EMMI’s Foundation
Terray’s chemistry foundation model, COATI, first released in 2023, is now in its third generation and is trained on over a billion diverse molecules. Conceptually, this is akin to a GPT for molecules and Terray continues to regularly improve this model. COATI3 addresses one of the most basic needs of AI models – expressing molecules in a language that models can understand. Traditional molecular representations that encode structures as vectors are unsuitable as their meaning is lost if they are manipulated, so they do not decode efficiently. In contrast, COATI encodes molecules in a mathematical latent space. This transformation makes molecules accessible by AI models to manipulate and investigate – and, most importantly, provides the language to translate the results back into molecules. COATI3 is the world’s most advanced multi-modal chemistry foundation model and powers EMMI’s Generate, Predict, and Select models.
EMMI Generate
At each cycle of drug discovery, scientists are faced with figuring out which changes in molecular structure will improve a set of properties needed to move towards clinical testing. EMMI guides Terray’s scientists using generative AI models to provide a set of molecules that are expected to have the desired properties. We published our first generative method of latent diffusion which directly leveraged our COATI embedding in 2024. As we have progressed, we have continuously improved this method, including introducing a new model that leverages a reinforcement learning-based generative method that uses a policy gradient algorithm to propose property-optimized molecules that are more synthetically accessible. EMMI’s generative capabilities quickly design millions of promising molecules for any challenge.
EMMI Predict
EMMI then predicts the potency and selectivity for these millions of molecules using a global potency prediction model. Owing to Terray’s focus on potency over pose, this model is much faster and cheaper than public models, which makes routine predictions on millions of molecules possible. The most promising molecules are then processed by ADME and physicochemical models to predict solubility, LogD, permeability, metabolism, clearance, and more.
EMMI Select
The latest modeling capability, EMMI’s selection models, transform the speed and efficiency of small molecule discovery. Select models evaluate the results of the Predict models, measuring and understanding the uncertainty associated with each prediction and then optimizing how many and which molecules to make and test next to advance a program. Tested retrospectively, this means Terray can get to the key molecule 3x faster and cheaper than those using industry standard selection methods. Deploying EMMI’s section models across programs, Terray has seen early results that time and cost savings will be realized prospectively.