Q&A: Generate Biomedicines CTO talks AI advancements

Innovating with AI is tricky for many businesses. Not only do leaders need to know how to use AI advancements to best serve their teams, but they must also keep in mind varying governance and oversight considerations.

For Generate Biomedicines, using AI has been natural from the start. Since its inception, the company's mission has been to render therapeutic biology programmable. In doing so, the goal is to enable AI-driven drug discovery to aid in the development of medicines for a variety of human conditions.

"The company was founded with the idea that proteins … can become programmable because we had founding insights that you can build generative models of proteins from data," said Gevorg Grigoryan, co-founder and CTO of Generate Biomedicines.

So far, Generate's use of AI has been fruitful. The company's Generate Platform applies generative AI and machine learning in a "generate, build, measure and learn" loop to identify and validate custom therapeutics. The company has raised over $800 million in venture capital and has partnered with global pharmaceutical companies and cancer research centers, including Amgen and Novartis. As of January 2026, Generate's AI-designed drug, molecule GB-0895, is in clinical development. The company also raised $400 million in its February 2026 IPO.

In this Q&A, Grigoryan unpacks Generate's use of AI, explaining how AI isn't just another tool to bolt onto your business -- it's a fundamental rethink of how your company operates. He also suggests a critical leadership blind spot: executives who value decision-making over driving impact might find themselves increasingly obsolete in an AI-enabled world.

Editor's note: The following interview was edited for length and clarity.

How are you using advancements in AI to your advantage? Is there anything specific that's been crucial to the development of your platform?

Gevorg Grigoryan: A lot has happened in AI and machine learning that has been helpful, but actually, it's the integration between machine learning and experimentation that is the key to unlocking [protein] programmability. If you step back and ask, Why is programming biology hard?, it's because biology is complex. At any given point, it's only partially understood and partially a black box. It's fundamentally a question of how you program a system you don't fully understand.

The thing that is special about this moment in AI is that AI represents a scalable hypothesis-generation machine. We've always been able to design molecules or design therapeutic hypotheses reasonably well. The problem is that we didn't know how to scale that process, and now we do … That is how you capture the value of AI, which is a paradoxical point because one often hears the sentiment that the reason we build performance models is so we don't have to do experiments. But I see it the other way: the more performant your models are, the more experiments you want to do. That model is good at coming up with good ideas, but what it doesn't know is everything, so you still need to test those ideas at scale.

The short version of the story is that we've invested a lot in building up both the wet lab measurement, or validation stack, as well as the machine learning stack. The two going hand in hand is what's key within our space.

How do you determine in what instances AI is a better fit than traditional software?

Grigoryan: It really helped that we were AI-native … the very founding of the company was around the idea of generative biology [and] generative models. It wasn't very hard to convince ourselves to use machine learning or to find people who want to join a team that has that as its vision.

[AI] isn't an add, it's a complete rethink.

What has changed in the last couple of months, and in particular last year, is coding and science and the impact of agents within coding and science. We started noticing earlier last year that agents have become really good at writing code, but, actually, what we care about is the impact of doing so. It's been an incredible enabler, and it got us thinking about how to organize teams of engineers and have them have the best impact.

An even more interesting transformation was around using AI for generating hypotheses outside of molecules … When it became possible for the thing you generate to be not just a molecule, but an idea, it was a change, but it was quite natural within our walls. Now we think of hypotheses more generally … We teach agents to condition molecular generation on ideas and not necessarily detailed specs. It's been a very interesting moment.

I definitely sympathize with companies that have to [integrate AI] as a bolt-on, because the fact is that [AI] isn't an add, it's a complete rethink. This idea of what's a good AI use case -- I don't want to come off as arrogant at all -- but it's almost the wrong question. Because it's not the same kind of technology as, for example, a new kind of toaster. It's not about use cases, it's about given that you can spend GPU tokens and reason, how would you build a company in the first place? It's much more foundational than considering which of our existing processes can be repeated by AI.

Working in biotechnology and the healthcare industry, how do you think about AI safety and human oversight?

Grigoryan: Safety has always been a really important component of how we think about developing drugs. Our stance has been that irrespective of how molecules come to life and [regardless of whatever] process brings them to be, they need to pass the same set of criteria, the same set of metrics and the same safety and efficacy standards as molecules have always passed through.

When folks ask, how do we know that AI-generated molecules are safe or non-immunogenic -- which are questions we used to be asked a lot in the beginning, before we were in clinic -- my answer was and continues to be, we have to scrutinize them in exactly the same way as we've always done.

They still have to meet the same criteria, so we test for immunogenicity in the same way we do in vivo studies and animal studies, in the same way as we used to in the past … Trust is the important point. Just because I describe a process that's good doesn't mean it'll be trusted. [Trust] can only be earned through seeing over and over that things are safe and efficacious.

There's [also] enough complexity where the human in the loop is, at the moment, a requirement. The very fact is that things are sufficiently complex that we still need quite a few humans in the loop. In some ways, we're trying to reduce that, not as a way to reduce human involvement, but as a way to empower humans. The fewer handoff points we can have -- for example, teaching our hardware to speak in a language where it can be directly driven by some type of an API -- we want to do that as much as possible. Not to remove the human scientist, but simply because of the humans. The most enjoyable part of doing science isn't setting an instrument up and then waiting until it runs. It's thinking through the next idea and the experiment.

Closely related to AI governance is data governance. What have been some of the challenges around data quality and curation, and how have you been addressing them?

Grigoryan: I would separate things into two categories. There's the early pre-clinical stage, where we do experiments, and the data are not subject to as much regulatory scrutiny as clinical trials. In clinical trials, there's more infrastructure and governance regarding who gets to see data and making sure things are appropriately blinded.

In the [pre-clinical stage], which is the majority of data we generate, we have, since the beginning, thought of data as a collective asset. This was a bit of an adjustment for some of our scientists. The idea is that when you make a measurement and deposit the data set, it's not your data set or your measurement. It belongs to the system, and it's for everybody to be able to analyze, critique and make observations. We have a singular database where every piece of data lands. Data have to be there. And that database, with some rare exceptions, is basically accessible to everybody.

This moment reveals an axis that was otherwise hidden in the past: Do you value the decision or do you value the impact?

That's why when agents and scientific agents came on the scene, it was a very natural adjustment because then "everybody" included agents. An agent at Generate can think over every measurement, download it, do a different type of analysis and draw conclusions. That was very important.

The exception I mentioned was around data generated within our partnerships. This is where we have various obligations for our partners, where there are some things we can't do with certain data. Almost always, though, we can use them for platform purposes. We've been insistent on retaining rights to the data for improving our in-house capabilities.

What do leaders need to understand about innovating with AI if they want to be successful at it?

Grigoryan: This moment reveals an axis that was otherwise hidden in the past: Do you value the decision or do you value the impact? People who value impact are having an absolute blast in this moment because they got a tool that's indistinguishable from magic to be able to have more of that impact.

Folks who value driving the decision are having a hard time, because that sort of expertise is becoming increasingly commoditized. I think the right instinct is: Can we standardize [and] quantify decisions as much as possible, and really argue about the objectives and not necessarily each decision?

If you've been used to [thinking], "The reason I'm valuable is because I know a lot of stuff and I make the right call," this is a hard moment for folks like that -- whether it's engineers or C-suite leaders.

It doesn't mean one group is right or wrong, but understanding that there exist different phenotypes and they're having a different moment in the face of AI enablement is very important for leading.

What excites you about AI when you're looking into 2027 and beyond? Is there anything on your radar that you want to incorporate into your company in the future?

Grigoryan: The thing I'm excited about, and I wonder if this is on the horizon, is long-term memory and more natural learning, which I'm sure everybody's asking for. It seems like perhaps the one missing ingredient, because sometimes these systems are so good that we forget the models are static. They're not actually changing, and the only thing that's changing is the prompt.

If you want a true partner and a co-scientist or a group of co-scientists, it would be really nice to have them naturally learn on the fly like we do. It's not yet directly happening. We're emulating it with prompts and harnesses and context that's compressed over time. But it does still feel like an emulation. And I know that many folks are working on that. I think there are versions of that as this applies to, for example, molecular generative models that we're working on -- and certainly where our own AI research or machine learning research is really driving the frontier is on generative models and how to make them have higher fidelity, higher confidence and higher experimental rates -- that sort of iteration with experimental feedback to feed into the models to get better performance, because it's becoming increasingly scalable to generate ideas and to test them.

Olivia Wisbey is a site editor for Informa TechTarget's AI & Emerging Tech group. She has experience covering AI, machine learning and other emerging technologies.