Lindsay Edwards knows what it takes to build AI systems from the ground up. From pioneering the first data science group at GSK to leading AI teams at AstraZeneca, Lindsay is now CTO of Relation, where he’s helping redefine the future of data-driven drug discovery.
Lindsay joined me to talk about why AI needs a reality check, closing the gap between hype and impact, and why this is the best time to be a scientist. — Sajith Wickramasekara
* Editor’s note: The conversation has been edited for length and clarity.
From chart-topping musician to machine learning pioneer
Sajith Wickramasekara: You’ve produced multiple UK top ten singles and now you’re the CTO of Relation Therapeutics. How did you go from music to drug discovery and machine learning?
Lindsay Edwards: Five years ago, I had no idea I was going to be where I am now.
I was a professional musician and a record producer in the 90s, and worked with everyone from Sting to Mariah Carey, Whitney Houston, and my own band. Then it got quiet, which is pretty normal in a music career.
I'd left school at 17, so I wasn't sure what I was going to do, but basically discovered that I liked science. I got a scholarship at Oxford and no one was more surprised than me.
I loved research and from then on, was a career academic. I got into machine learning, which at the time was weird and not cool to be in.
Sajith: Is that what led you to industry?
Lindsay: It was a mix of two things. First, I discovered that raising a young family around London on an academic salary is a fool's errand. I was going to move back to Australia, but then got a call from someone at GSK who asked, why don't you come work for us? So I did.
At GSK, it was the very beginning of people starting to think about applying machine learning and data science. I set up the first data science group at GSK.
One of the first things I built was a visualization layer over the top of those data. The scientists could type in a gene and see gene expression in all the different patient populations. Really simple, but incredibly useful. It just went from there.
“When you have to build everything from scratch, you have no data infrastructure. We had to unify all of the data across all these different studies.”
Sajith: What led you to Relation?
Lindsay: In the last ten years, there’s been increasing focus on how we can use machine learning. I still think we're barely in the foothills in terms of understanding what might be possible.
I'd built some useful things at GSK, and a good team. But it was time to move on. I went to AstraZeneca and my team worked a little bit on the COVID vaccine, which is incredible. But quite quickly, I got this feeling that I hadn't really scratched the itch.
Sajith: You wanted to build from the ground up?
Lindsay: Being part of a machine learning startup was something I really wanted to do. I met Charlie Roberts, one of the founders of Relation, in front of a taco truck at NeurIPS. He started talking to me about Relation, which was about three people at the time. And I just thought, this is it.
Sajith: You've mentioned before that both science and music require creativity, inspiration, and insight. Can you elaborate?
Lindsay: One of the big misconceptions about science is that it's this dry theoretical thing, but it's incredibly creative. You have to look at data, think about problems, and come up with ideas. It's definitely not this dry thing that I think a lot of non-scientists think it is.
“I think one of the big misconceptions about science is that it's this dry theoretical thing, but it's incredibly creative.”
Sajith: What have been the biggest surprises going from academia to big pharma to startups?
Lindsay: I’m definitely a creature of industry more than academia. At the time, academia tended to reward picking an ion channel and working on that for 30 years. Other people have had great careers doing that, but I'm just too intrigued by shiny things.
My lab did tissue engineering, machine learning, and all sorts of computational modeling. It was cool and fun, but I think it was too early. It just didn't connect [with granting bodies].
When I went to GSK, I really liked the commitment to being sure about what you're looking at. In academia, you're just trying to publish. Whereas in drug discovery, the final arbiter is a $30 million phase three clinical study.
I also liked that there was a lot of room to be entrepreneurial inside GSK.
“In these big companies, if you're willing to ask for six months to try to make something work, often people will say, ‘Yeah, why not?’”
Why AI needs a reality check
Sajith: You've been quite vocal about avoiding the term AI, calling it vague. Why is this specificity so important?
Lindsay: I use the term much more now. In the last couple of years, that term has become more appropriate for the technology that we have.
“My main objection is when someone says, ‘we're using AI for drug discovery.’ That’s like saying, ‘I'm using chemistry for drug discovery.’”
The phrase lacks any specificity that allows you to have a mature, interesting conversation about what you're actually doing — and is so vague as to be unprosecutable.
I want to hear the details, like what exactly is the problem, how are we going to solve it, and crucially, how will we know if it's working? These things are really hard.
We spent a lot of time at Relation trying to figure all this out. It's incredibly difficult because we're trying to model things that are difficult to test in the lab. It's not straightforward to figure out whether things are working as well as you think they are.
Sajith: Today in machine learning and drug discovery, what's real and what's not?
Lindsay: Structural biology is real. ML for chemistry, a lot of that's real. We've made a big investment into DNA foundation models. That's real.
Sajith: What’s the use case there?
Lindsay: We use it for variant gene mapping. Although it's actually opening up the problem space a bit more. We’re finding that a variant maps to multiple genes in the locus and it's context-dependent. So, these models allow you to at least try and tackle some of that.
Sajith: What is overrated right now?
Lindsay: Virtual cells.
“When somebody says “virtual cell”, I imagine laypeople think of a computational model that synthetically captures all of the complexity of a cell. We are light years away from that.”
What people actually mean, at the moment, is predicting gene expression — and that's not a virtual cell.
The other problem is that the way machine learning scientists think about things is often really disconnected from the way that biologists need us to think about things.
As a machine learning scientist, you can take a gene expression dataset, train on part of the data, replace some data, and watch your metric go up. You go show a biologist, and they say, “I don't care. I can do that experiment and it's taken you three weeks to create the model.”
Actually, what the biologists want us to do is to predict on cells that the model hasn't seen before. That is much harder because cells can behave very differently, often for reasons that we don't fully understand. The more complicated models are very bad out-of-distribution, AlphaFold included.
The formal term is a “convex hull.” If you imagine your data with a rubber band around it, it functions well within the rubber band, because it's got so many parameters. As soon as you go outside, they're terrible. If you train a model on this cell and a scientist asks, “Can you predict on this other cell?” We say, no.
“There’s still a disconnect between the way that we do things as a machine learning community and what's actually needed from us.”
Sajith: How do you bridge that? How do you get machine learning scientists and biologists talking enough to make sure that doesn't happen?
Lindsay: As a community, we've tried things that seem intuitively reasonable, but haven't worked. I don't think the answer is to teach biologists how to code. I’m sure the answer is to teach machine learning folks about biology either.
But if you get people to coexist around a problem, and you've got smart autodidacts, they'll tend to learn as they go. The best recipe is to form interdisciplinary groups of people, give them a clearly defined objective, and then give them the time and space to figure out how to do it, rather than being too prescriptive.
The biggest gaps in AI today
Sajith: How do you think about the gap between having powerful AI models that are off-the-shelf and making them useful for scientific work?
Lindsay: The promise of foundation models is real and it's not new. If you look at when vision models first broke through in 2014, we were talking about using general representations and fine-tuning them, but that took years. Contrastive loss was the first method that allowed you to train a more multipurpose embedding that could be useful for other things. Until then, supervised learning still won out. I think we're still there.
Whenever we train DNA foundation models, the ones that are trained on a supervised task outperform things like the Nucleotide Transformer or Evo. It's about finding the right pretext task. In theory, you could have a good general-purpose DNA foundation model that you can fine-tune.
When you train a model, you can use a supervised task, with labeled data. Or, you can train on an unsupervised task, without labels. LLMs are trained on an unsupervised task where you mask a word and predict the next one — that’s called a pretext task.
In DNA modeling, you could do the same pretext task. You take a stretch of DNA, blank out a few letters, and ask the model to predict them. A couple of models have been trained that way. You could apply the same idea to gene expression data: hide some of the data and ask the model to predict it.
The advantage is that you don't need labels — you don't need humans identifying cell types and so on. But the downside is that the pretext task has to be good enough for the model to learn something useful.
In language, it clearly works. With DNA, the model learns about codons and other conserved sequences. But beyond that, how will it learn what’s a regulatory element?
That’s the current gap. There's a lot of research to be done to figure out whether there’s a pretext task. Will models need to remain supervised? Are there better supervised tasks that to train on?
Why biologists need to be more like physicists
Sajith: You've talked about biology's data bottleneck before: that our measurements are messy and consistent, and hard for machines to learn from. You've even said that biologists need to be more like physicists. How does that help solve the data bottleneck?
Lindsay: In biology, we've basically been measuring things in arbitrary units, because measuring things in absolute units is really hard.
With gene expression, we’re looking at a gene’s expression relative to a control that’s been randomized in an experiment, to make sure batch effects are marginalized out.
Even when we go to the doctor, measurements are compared to an internal control or a normal range. We have this culture where numbers don't necessarily mean anything in isolation; they only mean something when compared to something else. That’s been fine throughout the history of biology, but it's a death knell when you're trying to train models.
Often, the effects that we're interested in are confounded with the batch effects that we normally get rid of with randomization. And so the models just learn things like batch effects.
“One of the first lessons in machine learning is that models will learn what’s easiest. It's very easy to convince yourself that they’ve learned what you wanted them to learn. More often than not, they’ve learned the easiest way to solve the task that you gave them.”
Sajith: And physics has this solved?
Lindsay: Well, they're more used to measuring things in absolute values.
AlphaFold is a good example where you've got two measurements that are naturally aggregatable across labs, and across time and space. If you measured a sequence, whether it’s in London or Kendall Square, it's going to be approximately the same. Then, the measurements of structure are in angstrom distances. So if you take all of the structures and sequences from different labs, you can combine that data and train models on it.
Gene expression is a really good example of data where that's just not true. The challenge is having laboratory systems that can measure it so consistently that it’s a reproducible absolute value.
Why too much compute is a problem
Sajith: What's the most critical infrastructure problem that's holding back progress in machine learning?
Lindsay: It doesn't feel like there's a technological bottleneck in machine learning right now. The available compute is just vast. Think about companies that have raised half a billion dollars, but somehow lose their way. They have too much money and lose focus.
Something similar is happening in machine learning. When you have ten thousand GPUs to throw at a problem, you don't need the same discipline you would if you had a smaller number.
We recently published a paper on a DNA foundation model with just 19 million parameters, which is tiny by today's standards. It’s also topping the charts on performance, mostly because we're tokenizing slightly differently.
Sajith: Having constraints forces creativity and focus.
Lindsay: It does. DeepSeek is a good example of that.
Why legacy data is worth less than you think
Sajith: You’ve said that companies shouldn’t sink resources into curating old data, but instead have policies to ensure all new data is machine learning-ready. What does that mean in practice?
Lindsay: If you’re a massive company, there's a temptation to look at your data estate and assume it must be worth something. Then you spend $2 million bringing in consultants to run some normalization on it — only to find that it's not useful.
With biology in particular, there's so many things that we don't see or measure. And until you know to measure it, you're not going to measure it by accident.
“Rather than spending $2 million trying to figure out how to drag marginal value out of old data, just spend that money making sure that every single clinical study you run from now is consistent. Within three years, you'd have an industry-defining dataset.”
Recursion really showed how quickly you could go from a standing start to an industry-defining dataset. They started releasing it into the public domain, and it was like, let's do that.
The future of machine learning in biology
Sajith: How do you see ML changing over the next few years? How do you see how we understand and treat disease evolving over the next 2-3 years?
Lindsay: We are genuinely at a point where things will start to have a real impact. People have been talking about it having an impact for years. But the truth is, it hasn't.
The gap between having a model that “works” and having something that's actually useful is far bigger than people realize. Members of the AlphaFold team have even said they were surprised more hadn't done with the base model.
It’s one thing to build a model that does something academically interesting. It’s another to make it useful. That’s the gap.
Still, there’s a lot of promise. DNA foundation models look good. Protein folding is obviously starting to take off. Antibody design and chemical design are really exciting now.
The big gap for cell models is still the data. There's a lot of observational expression data out there, but if you want to understand a system as complicated as a cell, you need to poke it a lot.
Sajith: What has to happen for you to say that machine learning has made an impact on drug discovery?
Lindsay: We need to see approved drugs that wouldn't exist otherwise. I don't think we've seen that yet. We've had an effect around the edges and we can speed things up.
It takes a long time to make a medicine, so there’s that natural lag as well. It’ll come, but it’s going to take a while. We need to lower the temperature of the conversation a little bit.
There are a lot of really excited, smart people in the machine learning community that want to get their teeth into meaty problems. We struggle to give them those problems at the moment.
Again, AlphaFold is a really good example. The fact that CASP existed made it possible. The more we identify foundational biology challenges with really good datasets, the more we’ll see that kind of progress.
Also, drug discovery isn't the only reason to do machine learning in biology. There's a lot of important work that might not seem transformative, but can end up driving progress in drug discovery as a side effect.
Sajith: What are you thinking of?
Lindsay: DNA foundation models are a good example. In the synthetic biology community, scientists are building simple biological systems from the ground up, but there isn’t always an immediate application. Organisms like yeast are simpler than human cells, but we have a reasonably good understanding of the gene regulatory network. There’s value in solving a simpler system.
Sajith: What's your advice for the next generation of scientists working at the frontier of machine learning and biology?
Lindsay: Be excited. Do stuff that you love. I've been blessed to have had two careers where I've been excited about my work. It's simple advice, but it's fundamental.
“I couldn't have survived working on things that fundamentally didn't make me happy.”
There's also a difference between the lofty aspirations of curing cancer or other diseases, which is great — but you have to love it too.
We're blessed being at a time when there's some really interesting problems with fascinating technology available. There’s a huge white space that you can occupy and try stuff out.
Sajith: Best time ever to be a scientist?
Lindsay: It's got to be up there.