Engineering the Key
Murat Tunaboylu on how AI is accelerating novel drug discovery and could bring about the era of personalised medicine
For most of pharmaceutical history, finding a new drug to treat a specific disease has resembled more a game of chance than a precise endeavour. The number of possible molecular combinations relevant to human biology runs to 1060 — or 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 — yet the industry’s traditional approach has essentially been to pick from that vast space at random. Compounds are synthesised and tested, and the few that show promise are taken forward, but the vast majority fall by the wayside. The process is slow, expensive and remarkably wasteful — on average, it takes around 12 years and over $2 billion in development costs before a new drug reaches patients.
Artificial intelligence (AI) offers a different approach. Rather than blindly trying to find suitable drug candidates for diseases, AI-driven drug discovery attempts to both read the underlying logic of how molecules interact and design new candidates from first principles.
Antiverse, a biotech company founded in the United Kingdom in 2017, is betting big on AI to revolutionise drug discovery and design. Its platform focuses on a class of biological targets — G protein-coupled receptors (GPCRs) and ion channels — that are associated with over 60 diseases but where fewer than five FDA-approved antibody therapies exist. I caught up with their Co-Founder, Murat Tunaboylu, to learn more about how AI is transforming drug discovery from a lottery into a design discipline, where the biggest near-term breakthroughs are likely to come, and why the UK’s regulatory posture may be more forward-thinking than it gets credit for.
What we discussed
The key-and-lock analogy: why drug discovery is really a design problem, not a search problem.
How AI moves from process improvement to genuine enablement — doing things that were previously impossible.
Which diseases are likeliest to be transformed by AI in the near term, and which will take longer.
The data scarcity problem in biologics, and how federated learning and open data mandates could help.
Why the UK’s medical regulatory environment is more innovative than its reputation suggests — not least through its world-leading embrace of process-level approval that could redefine the economics of personalised treatments.
The relationship between AI-native startups and big pharma — and why we might expect M&A activity to accelerate.
How the UK is faring as a home for biotech companies — and what would make it better.
Lessons for for scaling AI-driven drug discovery companies
For policymakers:
Mandate open data for publicly funded research. When government-funded clinical trials generate data — whether or not the trial succeeds — there is a strong public-interest case for requiring that data to be deposited in shared repositories. At present, too much of the scientific record stays locked in proprietary silos, duplicating effort and slowing discovery. A clear open data requirement for public funding could be transformative.
Understand how public funding and startup timelines need to align. Antiverse credits Innovate UK funding as being critical to its survival in its early years. Matching grant mechanisms — where public funding is unlocked by private investment — can be particularly powerful, as they leverage government spending while maintaining market discipline. Expanding such instruments would help more companies survive the valley of death.
Protect SEIS and EIS. The Seed Enterprise Investment Scheme and Enterprise Investment Scheme provide crucial tax relief to early-stage investors. Any moves to curtail or remove them risk directly reducing the flow of risk capital into exactly the kind of ambitious, long-horizon companies that the UK should be trying to cultivate.
Address the seed-to-Series A funding gap. The UK has a well-developed ecosystem for founding companies — accelerators, angels, early grants — and an emerging one for later-stage growth. The gap in the middle, between early traction and institutional Series A investment, remains poorly served.
Support regulatory innovation and be ready to surge capacity. The UK’s Medicines and Healthcare products Regulatory Agency (MHRA) appears to be ahead of the curve on process-level drug approvals — a development that could fundamentally change the economics of personalised medicine. This lead should be protected, signalled clearly to international investors and founders, and actively built upon. If AI leads to a wave of new drugs being put forward for regulatory consideration, regulators will need the capacity to manage that.
For founders:
Go after the unsolved problem, not the incrementally better one. Antiverse deliberately targeted GPCRs and ion channels because they were difficult to solve. The biggest prizes in AI-driven drug discovery will go to those willing to take on biological complexity that existing tools cannot handle.
Data generation is a core competency, not a precondition. In the biologics space, the data needed often doesn’t exist yet. Building the capability to generate high-quality, structured experimental data — and to do so at scale — is as important as the AI platform itself.
Think carefully about international expansion. Capital intensity in biotech makes having a presence in the US almost inevitable for companies in the sector with serious ambitions. But the R&D base doesn’t have to move as well. Antiverse has kept its labs and most of its team in the UK while locating its CEO and business development activities in Boston — a model worth considering for other startups.
Build platform before pipeline, then transition. Antiverse spent its early years proving its platform through partnerships — demonstrating that its AI could design antibodies against targets incumbent pharmaceutical companies couldn’t crack. Now it is beginning to internalise that value by building its own pipeline. The platform-first approach derisked the science; the pipeline is where enterprise value is created.
Build capacity for grant funding. For innovative companies in need of grant funding to invest in research, do not underestimate the reporting burdens and paperwork involved. Startups may be able to bring in outside contractors to support themselves, or have staff dedicated to securing and then managing grant funding.
Full interview
I. Design versus discovery
Antiverse specialises in antibody design against challenging targets. Before we get into the science underpinning that, can you explain what it means in plain English?
This is a question we get asked a lot, and the best analogy we’ve found is that of locks and keys. Think of locks as somehow associated with a disease — where unlocking them with the right key represents the therapeutic intervention. Up until now, drug discovery has essentially been key discovery. It’s as if you are going through billions and billions of different keys and trying them one by one, maybe two at a time, with the hope that some of them will unlock the right mechanism. But for molecular combinations relevant to human biology, there are almost ten to the power of 60 possible combinations. So even if you are trying these keys randomly, you would only ever be able to sample a tiny fraction of the full key universe.
Our approach is therefore to focus more on design than discovery. We analyse the lock, and we look at other lock-and-key pairs to study how they fit together. After capturing that information, we design new keys for extremely challenging locks — the ones traditional pharmaceutical companies have been struggling with for the past twenty years or more.
How does AI make that possible? Is this essentially just about accelerating processes that already existed, or does it enable genuinely new scientific techniques to be deployed?
The first applications of AI are usually process improvement — you get better speed, quality, accuracy. Our approach was different: we said, let’s go after problems that weren’t solved before with existing technologies. And that’s where AI really enters the picture. The complexity of how these keys and locks fit together is incomprehensibly large. Our hypothesis was that if you can generate a big enough dataset of these lock-and-key pairs, you would be able to study them closely enough to infer relationships about them from within that data.
Such tools didn’t fully exist before. Creating the datasets we needed required things like next-generation sequencing, high-throughput screening, and significant automation in the wet lab. And then you also need a tool capable of making sense of that data — something that matches the complexity of the problem. That’s where AI and machine learning can come in. We can now not only generate large datasets, but we can also make sense of how things fit together in a multidimensional space. And then there’s another step: once you’ve captured those relationships, generative models can create new things based on what they’ve learned.
That’s where Antiverse found its niche. A pharmaceutical company will come to us saying: “we’ve been trying to crack this target for twenty years.” We go ahead and design molecules that bind to it. About half of what we do is in the wet lab, and the other half is in the AI.
II. Blockbusters to bespoke
Let’s say the drug discovery industry doubled or even quintupled its reliance on AI. What problems might we actually start to solve?
AI is being used across the full drug discovery pipeline — from identifying a target associated with a disease, to discovering a molecule that modulates it, through preclinical development and into the clinic. There are already many wins, and many areas are still open for significant gains to be made.
In the long run, I think there are a couple of transformative trends coming down the line. One is the shift away from blockbuster drugs — the ones that generate billions in revenue but have only modest clinical efficacy across a large population — towards smaller population groups and more bespoke treatments. AI can enable better decision-making in clinical development, so both the testing process becomes faster and the attrition rate — how many molecules successfully reach later stages — improves.
But what I think will be truly transformative is something more fundamental. In our space — antibody design against GPCRs — there are hundreds of receptors, with a few hundred associated with over 60 diseases. And yet there are perhaps five or fewer FDA-approved antibodies targeting them. It is an extraordinarily challenging area, but enabling technologies here will go far.
When you look at the later stages — clinical testing — there are multiple AI-first approaches emerging. Concepts like ‘organ-on-a-chip’ or ‘lab-on-a-chip’ allow you to test at much higher throughput at a lower cost base. Further into the future, I believe we could see approval granted not to an individual molecule, but to a process. What that means is that you wouldn’t need to go to clinical trials a hundred times for a hundred different molecules. You get the approval once, and everything that comes out of that process carries a validated safety and efficacy profile. It would allow treatment of much smaller patient groups in a way that finally makes the economics work. I believe we will eventually see truly personalised medicine — a pill not just for a disease, but for a specific individual.
Which diseases do you think are most likely to be transformed in the near term — say in the next three to five years — thanks to drug discovery techniques that only AI can enable?
I would order the likely gains based on biological complexity. Where we understand the biology better, we’ll see faster progress. So this would include things like certain cancers and metabolic disorders. Then, as you move towards psychiatric conditions, progress will be slower because we understand them much less.
This isn’t purely a problem of process or governance or technology — it is fundamentally a problem of biology. We still barely understand how all the genes fit together, how proteins interact, how cells integrate, or how disease actually emerges. We are poking around and trying to find our way, but the fundamental understanding is still at quite a primitive level. That being said, I do think AI will give us the tools necessary to probe into that biology more effectively, and that could create a virtuous cycle — faster biological understanding generating better models, and facilitating faster discovery.
III. Science in silos
You’ve spoken before about the challenge of data access in AI-driven pharma. How does that actually manifest? For instance, what data are you struggling to get, and what could you do more of if you had access to it?
AI companies are fundamentally data companies. You need good quality data that is representative of the problem you are trying to solve. Sometimes that data doesn’t exist yet — the assays or tooling to generate it haven’t been developed, or it’s prohibitively expensive to produce. In our case, where we’re working on antibodies and challenging targets, the data problem is acute, and because the field previously used a discovery approach, where you would find a few molecules and that was great, the assays were built differently and the data was collected differently.
Some of this data may exist in silos inside big pharmaceutical companies — generated over the past twenty years — but without any conscious effort to clean it, collate it, and put it into a usable database, it is unusable. Even in a quite modern pharma environment, plenty of useful data is just living in somebody’s laptop, and even here, half of it might be in an Excel sheet with the other half in a PowerPoint.
In the small molecule world it’s a bit easier — those molecules are better understood, creating data is cheaper, and automation is more mature. That’s why you see more AI companies there. When you move to biologics and antibodies, however, you are in a low-data regime: where data is much more difficult to find and expensive to produce. But things are improving. New tools are coming to maturity, and the cost of next-generation sequencing is falling on a near-logarithmic curve.
Are there any structural solutions — policy levers, industry initiatives — that you think could help?
Data is extremely valuable, and that is essentially an intellectual property issue. People are understandably not willing to share what they’ve invested heavily in producing. There have been some attempts to create shared databases where you can only benefit from the data if you contribute to it, and these could encourage more sharing.
There’s also an approach called federated learning, which I think will be quite important going forwards. Rather than sharing the data itself, you run your machine learning model on the data at the source. Whatever you take back is the improved model, not the raw data. People are more comfortable with doing that.
Another concept I’ve encountered in the United States that I think could be a big driver of change is that any government-funded research should be required to share its data publicly by contributing it to a database for others to use. Imagine clinical trials funded by the government, generating vast amounts of data — whether successful or not — and that data becoming available to the whole community. In the current situation, you often have a hot target with dozens of companies essentially duplicating the same work. Open data mandates could change that.
IV. Approver approval
Anything relating to novel healthcare technology is naturally going to be highly regulated. What is your sense of how the UK’s regulatory environment is faring? Is it keeping pace, or holding things back?
We haven’t progressed far enough in our own internal pipeline to comment on that directly. But through our partnerships, and from what I’ve been observing in general, there’s actually a remarkable story developing. More and more, we’re seeing how the UK, through the MHRA [Medicines and Healthcare products Regulatory Agency], appears to be increasingly open to a certain kind of regulatory innovation than even regulators in the US are.
In fact, there was a recent article in The Economist which focused on a company called EveryONE Medicines, based in Boston. They’re working on antisense oligonucleotides, which are molecules that can regulate protein expression. Misfolded or dysregulated proteins cause all sorts of diseases. This company was founded by the mother of a girl called Mila Makovec, who died of a serious rare disease condition. After encountering barriers in the US, she turned to the MHRA in the UK. And they are seeking something remarkable: a process-level approval. Rather than getting approval for a specific drug, they’re seeking approval for the process — so everything that comes out of it carries a validated profile. Their first clinical trial, I believe for 10 patients, is about to start. If this succeeds, it will be a first for the UK, and a first for the world.
That could be a stepping stone towards what I would call a software-as-a-drug or AI-as-a-drug track. Just imagine: you design something, and in a matter of weeks or months it gets to a human patient who has no other option — whereas the current process for getting a drug approved might typically cost $2 billion and take 12 years to complete, if it’s successful in the first place. Any mistake here would impact people’s lives in a very significant way, so regulators are rightly cautious, but this is a big change, and the UK seems to be leading it.
That’s encouraging to hear! Following on from that, another question I had anyway would be, if AI does produce a wave of promising new drug candidates, do you think the regulatory system will have the capacity to deal with them all?
The number of drugs being approved is already increasing every year. I believe it was around 52 last year. Even post-COVID, I’ve heard that the FDA has been performing well in terms of responsiveness and proactivity. And if you think about the 12-year development cycle, companies build a relationship with the FDA throughout that process — so regulators have quite a lot of advance warning and can plan capacity accordingly. If timelines shorten dramatically, then yes, the regulator could become a bottleneck, but I don’t think that’s the problem we need to solve for right now.
V. Building Antiverse
Zooming in on Antiverse itself, can you tell us more about how the company started out?
We began as one of Deep Science Ventures’ (DSV) first cohorts. It was quite unstructured at the time — more of an invitation to around 30 people to quit their jobs and see what happens. DSV has since become much more structured, to the point where it’s essentially a company factory that defines a problem and builds teams to address it.
We incorporated in April 2017 and have been building in the UK since. About two years ago, we recognised that building a really large biotech company would require very significant capital — more than we could access solely through UK channels. The US East Coast made sense for us as the place to go and establish a presence. The time difference isn’t too large, and in Boston you have Harvard and MIT, plus a high concentration of capital, and major pharma groups.
With all that being said, we made a conscious decision to keep almost everyone and our labs in the UK. As we saw it, Britain has the talent, and we benefited from good government support — so there was no need to move the R&D. But for the business development side of the equation, such as fundraising and striking partnerships, having at least one co-founder in the US felt necessary. My Co-Founder Ben Holland stayed in the UK as CTO; whereas as CEO, I took on a more mobile role and now spend a significant amount of time here in the US.
Where is Antiverse heading next? Is this a matter of scaling what you already have, or are there other avenues you think you could head down?
We are transitioning to more of a biotech company. By that I mean we will deploy a significant fraction of our capital to our own internal pipeline. We’ve identified targets associated with specific diseases, and now that we’ve validated the platform — we know it can design molecules against extremely challenging targets — it makes a lot of business sense to own those molecules and take a bit more risk for the chance of building a much larger company. This allows us to have a lot more control over our own destiny.
We will continue doing platform partnerships with a small number of select pharma companies — perhaps one over the next two years — but our main value driver will be moving our internal programmes to later stages.
How do you rate the UK as a place to build a biotech company? Have your perceptions changed at all since you started out?
When we started in 2017, I think the UK was undoubtedly the best place in Europe to build a biotech company. I still believe it’s the leader in Europe — Germany does have an ecosystem developing, France has made progress too, but there aren’t many places on the continent that can match what the UK offers.
The more honest comparison to make is with the US — where companies attract more funding and more collaboration, and have more backing. Part of that is generational — the US has been through the biotech cycle several times already, and so you have a deep pool of investors who genuinely understand the space and are looking for their next opportunity.
In the UK, we see a lot of biotech startups funded by angels or groups that made their money in real estate or fintech. They might be intelligent investors who are willing to take risk on new modalities, but perhaps without the same domain expertise to push companies forward at the right pace and in the right direction. Changing that will take generations — and it’s not something a government can fix by decree.
What policy interventions have made a real difference for Antiverse, and what else would you change if you could?
Funding from Innovate UK was, at a critical moment, extremely important for us. Without it, we might have had to shut down, or at best would have been lagging significantly behind other companies.
One thing funders should appreciate more is just how important timelines are for startups. From the moment you apply to the moment the money arrives, six months could have passed. If you miss a funding cycle, it can stretch to 12 or even 18 months. Not many of us raise enough capital to run comfortably for two years while waiting. If you are relying on grants for your company’s survival, that uncertainty can be quite unnerving. We were fortunate in that we had the ability to always treat government funding as a top-up rather than our primary lifeline, meaning we could afford to wait. But I’ve heard others describe the reporting burden as quite significant — and some companies hire dedicated staff just for grant management.
The matching grant mechanism is something I think is genuinely well-designed. You raise private capital, the government matches a portion of it. That creates a productive alignment of incentives.
A lot of Antiverse’s work involves solving problems that bigger pharmaceutical companies couldn’t crack themselves. How do you think that relationship between AI-native startups and incumbent pharma plays out over the next few years?
The large pharma companies are definitely embracing AI, and putting significant resources behind it. There’s just a huge number of things you can do with it. But their risk profile is quite different from a startup’s. Groups within big pharma tend to look for things that are achievable and make an immediate impact. Startups fill a different gap: they’re more willing to go after the moonshots, the things with maybe a 1% chance of success that require 10 years of sustained commitment. That doesn’t naturally fit a large pharma’s strategic planning horizon.
It is actually a smart play for them to watch many companies develop, wait until one becomes the category leader, and then acquire it. Maybe that acquisition costs half a billion — which might be a few percent of their annual R&D budget. So I don’t expect that dynamic to change anytime soon. What I do expect is that we will see a lot more M&A [mergers and acquisitions]. Until very recently there were almost no AI-driven biologics or antibodies in pharma pipelines. Then that jumped to around 10%, and many are now getting to clinical stages. I can totally see a future where the majority of the pipeline is touched by AI in some way — either improved or created as a novel molecule.
Is there anything else the Government could do that would move the needle for biotech startups?
The SEIS and EIS [Seed Enterprise Investment Schemes and Enterprise Investment Schemes] tax relief schemes were a significant reason why Antiverse — and probably many other companies — got funded in the first place. I know there are ongoing conversations about possibly changing or curtailing these, and, in my view, if that were to happen, it would have a very direct and negative impact on how early-stage companies get funded in this country.
More broadly, there is a persistent funding gap between the early stage, where you get companies founded on a story, and the Series A stage, where institutional investors want numbers. In that gap — between a compelling pitch and a data-driven investment case — companies often fail. Governments, accelerator programmes, and family offices currently fill some of that gap, but not nearly enough of it. That’s where most of us fail, and that’s where more attention is needed.
We ask all our guests the same closing question: what’s one interesting thing you’ve read or listened to recently that you’d like to share with our readers?
Well, I mentioned it earlier, but The Economist article about Mila Makovec is worth highlighting. Obviously the personal story was heartbreaking, and the science fascinating, but what struck me most was the regulatory discussion it encompassed. It offers a glimpse of what truly personalised medicine could look like in practice — where individual cases can drive systemic change that benefits many others. That regulatory flexibility — that willingness to adapt frameworks — is what the sector needs if we’re serious about innovation.