Jacob Berlin, chemist-turned-entrepreneur, is cofounder and CEO of Terray Therapeutics. He’s on a mission to reinvent how small-molecule drugs are discovered — building ultradense microarray chips, generating the world’s largest chemistry dataset, and using AI to explore chemical space at unprecedented scale.
Jacob joined me to talk about why models alone aren’t a moat, and why the future of drug discovery is agentic, fully integrated, and built to deliver real medicines. — Sajith Wickramasekara
* Editor’s note: The conversation has been edited for length and clarity.
The elevator pitch (and science) that launched a family business
Sajith Wickramasekara: Can you start with the founding story of Terray?
Jacob Berlin: Terray is really a journey of hardware innovation, data revolution, and AI transformation.
It started in my academic lab, where my co-founders and I developed novel hardware — an ultra-dense microarray the size of a fingernail that lets you measure billions of interactions between small molecules and targets of interest. With that dataset, we can map dark areas of chemical space and find starting points others can't.
We've connected that to a full-stack AI that generates structures against a desired profile, predicts properties, and selects the optimal ones for testing. We work on a pipeline of wholly owned assets and one co-development with Odyssey focused on immunology, and we're on the cusp of the clinic next year.
Sajith: You're a scientist who spent years in academia. Your brother Eli was in private equity. Was co-founding a company together always the plan?
Jacob: I don't know if it was the master plan. Our entire lives we were great friends and good partners, and we hoped maybe there was something we could do together besides trade baseball cards.
This one arose serendipitously — the opportunity was there, we had the right overlapping skills, and there's no one I trust more to build something with than my brother.
Sajith: What was the elevator pitch that convinced him to join?
Jacob: Over the six years we built the technology in academia, we'd see each other at Thanksgiving and I'd say, “We got this incredible new project in the lab, you got to get in on this.” So he'd ask, “Does it work?” And I'd say, no, but it's an amazing idea.
So he stayed close, and every time we met I'd say, “Well, now this part works.” Then, finally, it really worked and was ready to scale. The sales pitch for life sciences is easy but also hard: this is an incredible opportunity to help the world, improve people's lives, and make a substantial return doing it. That intersection of mission and opportunity is why we do it.
Sajith: You have another co-founder, Kathleen Elison, who helped develop the microarray chip in your lab. Was there a specific moment when it stopped feeling like an academic project and started feeling like the basis of a true platform company?
Jacob: The magic moment (and probably the one that closed the sale with my brother) was when we had our first novel molecule discovery.
We reached a point where we could search millions of molecules. An academic partner brought us a novel target, and we were able to hand back our first validated hit that was able to engage a previously undrugged target. We realized this whole thing works.
The future was scaling it: large datasets, compute, AI, and the experts to tie it all together and deliver medicines. That was the moment we knew it didn't belong in academia anymore.
Why small molecules still demand reinvention
Sajith: Biologics have dominated venture capital for over a decade, yet small molecules still make up most approvals. What's the enduring advantage of small molecules, and why is it worth reinventing how they're discovered?
Jacob: The durable advantage of small molecules is ease of administration, manufacturing, and distribution. You can scale a pill to enormous quantities — think about how cheap Advil is. You can carry pills in your pocket, take it on long journeys, it’s no problem at all.
Biologics have come a long way, but they're still inherently more complex to manufacture, harder to distribute at scale, and difficult to reach other markets with. Small molecules will probably forever be the most common medicines humans take.
Sajith: Why reinvent how they're discovered?
Jacob: Although they're the world's most ubiquitous medicine by far, their discovery process is every bit as hard, or harder, than many other classes of medicine.
Your total success rate before the clinic is something like 1%, for a trillion-dollar-plus potential market and the health of millions to billions of people.
"Small-molecule drugs still follow industry norms: decades to fully develop, billions of dollars to fully develop, 10% success rate from clinical initiation to completion. We're at a moment where the data, AI, and compute tools are ready to radically transform that.”
Sajith: Is the AI opportunity fundamentally different in chemistry versus biologics, given the vastness of chemical space?
Jacob: They're related but very different problems.
Antibodies are a good setting for AI to begin because they're essentially a sequence of letters (amino acids) with a constrained, linear build. The combinatorial space is vast but you can probe it computationally.
For small molecules, the search space is ultimately infinite — you can go a different way off any given molecule and add what you want. Plus, biology can make and test large numbers of antibodies quite rapidly, so you can build a model, collect a large dataset, retrain, and use reinforcement learning efficiently.
Small molecules, prior to our work, meant: here's a prediction, now call the chemist to make ten or twenty things. The modern world of AI doesn't really work if you don't have large, precise, iterative data.
That's what we've built at Terray — an experimental platform that measures billions of interactions, paired with automation to follow up with thousands or tens of thousands of measurements.
"The small molecule challenge is how do you develop technology and workflows to feed these models data at scale? You have to have disruptive hardware to get to the datasets, whereas in biologics, you can use biology to get to them."
Building the stack in sequence: Hardware, then data, and AI
Sajith: You've described Terray's evolution as a hardware revolution, then a data revolution, then a compute revolution. How much more dense is your hardware than prior generations?
Jacob: The closest technology before us would be inkjet-printed microarrays. The miniaturization factor is something around 10,000-fold.
“What fits on our ultra-dense microarray, which is the size of a fingernail, would have previously fit on about half a tennis court.”
It's an incredible compression in size, acceleration in speed, and reduction in cost. It unlocks a wholly new chemistry dataset that didn't exist before.
Sajith: And what does that data unlock?
Jacob: It lets you map dark areas of chemical space that no one's really been in before and find novel starting points to really hard problems. From there, you kick into the AI piece: generating structures against a desired set of properties, predicting the full property package for those molecules, and selecting the optimal ones for testing.
We started with hardware innovation, used it to generate this data revolution to find where others can't start, and then built the AI revolution on top to leverage all those novel starting points.
Sajith: If you didn't have that proprietary hardware, would it have been cost prohibitive to generate this much data?
Jacob: Without our hardware, you're stuck in a brute-force commodity workflow where everyone is competing on the same general framework.
Our dataset is 40X larger than the entire public chemistry dataset, and we measure a billion new measurements every quarter. To try and brute force that scale with existing technologies, you'd be bankrupt long before you got there.
The true moat in AI drug discovery
Sajith: You've built a full-stack AI platform, called EMMI. Others in the industry seem more focused on just the computational models. Why does the integration of wet lab and computation matter?
Jacob: The whole reason it's called EMMI (Experimentation Meets Machine Intelligence) is that we strongly feel the durable moat is at the intersection of experimentation and AI.
If you build models but don't use them yourself, it's very hard to separate real progress from "I performed better on this benchmark by 0.04." You really need the tools in the hands of your scientists. Everyone at Terray is using them every day.
“You can't learn fast enough if you're only distributing your models to others.”
Sajith: Can you explain the role of selection models in EMMI and in drug discovery generally?
Jacob: Selection is fundamental because it's where AI meets reality — deciding what you actually make and test.
For small molecule discovery, maybe someday we’ll get to one-shot design. But for now, that’s not how it works. As you move through development, you continue to add properties over multiple cycles: potency, selectivity, avoiding off-target interactions, ADME. At each cycle, AI gives you far more options than you can physically test.
"I call it ‘AI abundance.’ You get to the end with 10,000 really interesting molecules, and a chemistry team that can make 50. How do you pick? Solving that optimally is really important."
EMMI Select computes the underlying uncertainty of the predictions and builds you an optimal set of things to test where the risks are anti-correlated. If molecule one doesn't work, molecule twenty probably does.
This radically compresses discovery times and saves months and millions of dollars.
Sajith: Terray deliberately avoids highly published or well-mapped chemical spaces. What's driving that strategy?
Jacob: It's certainly quicker to follow known chemical matter and make variants of it. Many companies can build models that are efficient in those settings, plus you have the rise of fast-following ecosystems, like in China.
If you're working in those spaces, the moat over time is really shrinking. There are going to be 8 or 10 companies in every race.
We work on targets and discovery challenges where our chemical equity is totally novel and unique. If you solve it, you're essentially alone for quite a while — and you have a durable moat for your programs and your value.
Why the platform isn’t the finish line
Sajith: Your first candidate is expected to enter the clinic next year. Was building your own pipeline always part of the plan?
Jacob: 100%. It's been a mission to bring new cures to people's health, all throughout my academic career and now through Terray. Ultimately, it's the maximum-value realization for transformative technology like ours — technology that lets you crack problems that have been unsolved forever. It's always been part of the plan.
Sajith: Do you think it's possible today to just be a platform company, or does every platform ultimately have to become a product company?
Jacob: It's certainly possible to be a platform company. I just think the platform-to-product arc, at the intersection of AI and experimentation, has such an incredible value unlock for solving impossible problems, delivering first-in-class medicines worth billions of dollars, and changing millions of lives — it's just the preferred path. Why wouldn't you?
Why we’ll always need humans in the loop
Sajith: Are there areas where AI drug discovery models still fail in ways that only an experienced medicinal chemist would detect?
Jacob: When it doesn't work, it's pretty clear — you put it in the assay and it comes back no instead of yes.
The more interesting area is the generation of seed datasets. When you get to a property where the model isn't as performant — owing to the complexity of the measurement or the scale of data available — figuring out the right seed data to start the iterative cycle is absolutely an expertise-driven step.
Experienced chemists play a key role in the selection process too: when the optimal basket of molecules is suggested, there's a human training and selection step where they say, “You picked these eight structures, but these two, I don't think so.” The model recomputes, comes back with a better basket, and learns for next time.
Sajith: Have you been surprised by any blind spots that medicinal chemists have that models have proven them wrong on?
Jacob: Our most advanced program has a really uncommon motif in it. When it was first suggested, everybody was like, “It's not going to work.” Because we're built at the intersection, our teams are built to try the tools. They said: “This has scored so highly and consistently, we have to see it.”
It turned out to be an incredible breakthrough for the program.
Sajith: There's a growing camp that argues the physical lab is becoming a liability, that everything will happen in silico. What's your case for the other side?
Jacob: I just don't know how we get all the way there.
I was in a conversation recently with someone who simulates the building of lasers — they only build it once or twice because they know whether a circuit will work and don’t need much physical testing. If you work in de novo discovery, you don't know the rules in the beginning. For the foreseeable future, the better path is large, precise, iterative data — the Terray story — versus hoping you can make it only once. The integration with the wet lab is essential to solve these problems.
The future of AI drug discovery is agentic — and actually practical
Sajith: What do you expect to see in AI drug discovery 12 months from now?
Jacob: Agentic orchestration and workflows will become common. The models will become more performant in the predictive space. There will be a huge emphasis on synthetically practical generative structures.
Generative AI arrived in small molecules with enormous promise but would suggest molecules that took forever to actually make. We've built a leading module, and I’m sure others are hard at work too, where you can constrain the synthetic space.
“In 12 months, I think every generative module will be synthetically constrained, timelines will compress radically, predictive tools will be better, and all of it will be agentically orchestrated.”
Sajith: Many of the frontier model companies in tech are now moving into healthcare and life sciences. What's your view on the role of LLMs in drug discovery?
Jacob: I think they're an incredible opportunity for orchestration.
You probably cannot just ask a frontier LLM what molecule solves this problem — these models are not actually very performant in the small-molecule chemistry space. But if you give them access to expert tools like ours, they can be incredible orchestrators of those tools, building optimal workflows and closing the loop.
It's the same way that Claude will use a calculator to do arithmetic because that's the most efficient tool for the job. These frontier models are incredibly powerful with the right tools in their hands.
Sajith: Do you have any advice for new biotech founders? Is there anything you wish you knew when you started?
Jacob: Whatever you pick to start on is actually your business. If you say, ‘We're going to start on this target just to show the platform works, and then we'll pick the real things to work on’ — it never happens. You're actually just working on whatever you started working on.
Also, the next step of your journey is always the hardest — and the most fun.
