Founder Stories: Andrew Hopkins, Founder and CEO of Exscientia: Using AI in Drug Discovery

You're listening to Idea Collider, a show that explores the world of asymmetric learning. On this show, I will sit down with pharmaceutical experts and business leaders to discuss how to embrace uncertainty and the different learning style that follows. I'm your host, Mike Rhea. Let's get into the show. Yeah, so Andrew, thank you so much for joining me on Idea Collider. First of all, could you let everyone know who you are and your journey to your current role?

Hi, Mike. Pleasure to see you again. And it's wonderful to be here. And, you know, thank you, listeners, for tuning in today. My name is Andrew Hopkins. I'm the founder and CEO of Accenture. Accenture is now a public company listed on NASDAQ back in October '21, one of Britain's biggest biotechs now. And the name itself means from knowledge, and that's Latin. Of course, if you're based in Oxford, it's useful to have a Latin name. But where it comes from, actually, is the philosophy of the company. What we are as a company have been driven by a long time about the frustrations that actually discovering drugs is such a rare feat. Actually, how can we democratize? How can we actually make that a much more successful endeavor, where we can go from a world where drugs are a sort of scarce commodity— we only invent 40 or 50 new ones a year, you know, each of them costing hundreds of millions, billions of dollars, compared to hundreds of movies actually that are generated almost the same amount of money each, you know. And to then go to a world where potentially drugs could be in abundance, because if you think of a medical need that's out there, actually that's what we see as a huge need for that. And I've been driven to do this sort of work, uh, easily for the past 25 years now, actually, since my start in my sort of DPhil, uh, in Oxford, uh, with Professor Dave Stewart, looking about how to design novel HIV drugs that overcome some resistance we were starting to see emerging in HIV even as early as the mid-'90s. And that then really set me on a course of thinking about how we can use information to create better medicines. And I was fortunate enough to work in that field, to really get involved in drug discovery from early days. And then also in my career then to move to a company like Pfizer around 1998. And Pfizer was a, was a great time, actually. They really had carte blanche, you know. I was recruited by John Overton, who, uh, you might— became head of Campbell, uh, actually now works Accenture, which is, which is a nice sort of, uh, uh, closing of a circle. Uh, but actually hired almost without a job description. Come work here and, uh, see if we could do interesting stuff. And we did, actually. And we came up with— we started to— had an interesting career where we, um, first job was almost like an internal consultant looking at all the new projects that were emerging. So it actually gave me a chance to look across a whole portfolio and start to see wherever problems were being. And that actually, that sort of work, that sort of insight led to concepts like the druggable genome, which became a sort of very important sort of concept in medicinal chemistry and drug discovery and thinking about how we would exploit all the emerging genome information for the next 20 years almost. And then when doing so, then I started to still have this idea I had for my PhD, which How could even sort of a single person or, you know, design a medicine by themselves? Is this huge effort normally? But how could we get to a world where we could really bring all the information to bear, you know, in how we design medicines? And the first protocol we did there was to think about how we could use the information to find new indications.

And we—

it's called drug repurposing. At the time I called it indications discovery and particularly was— had real resonance with Pfizer. Because you might remember around '98 Viagra was launched. And my first week on the job, I met Andy Bell, who was one of the inventors of Viagra, who also became one of the early sort of founders of Accenture. And, and of course, you know, the history of that was originally developed as cardiovascular medicine. And then we started to think about, of course, other opportunities for medicine, so to say, as the clinical teams in Pfizer were smart enough to realize there's commercial opportunity there. But then we asked the question, could we have discovered that rationally? Did it have to rely on serendipity? How could you engineer serendipity? And that really got me thinking into developing large information systems. How could we exploit the wealth of data that was already emerging in databases like PubMed, the emerging genome data, obviously, at the turn of the century? And we then realized that we could think about creating machine learning systems and knowledge-based sort of ontology systems where you could almost think of it like a graph where we could graph out every possible hypothesis. You know, it's like 20,000 now known genes against the, you know, 6,000, 7,000 diseases known to man, you know, and how many that number is actually we can debate later on. But you can start to define that as a universe and how then we can mine for each of those hypotheses. And that became a very powerful system.. And what I also learned, the power of interdisciplinary teams. You know, we managed to pull together a team, about a dozen of us, you know, computer scientists, clinician, bioinformaticians. And within that then, like multiple iterations of the software from writing the very first version in Perl to they'll build in a, an Oracle system, uh, by the end of the 12 months, uh, with nice sort of web interfaces, et cetera. But it saw me then that the, the asymmetry you could have. Where this small team of sort of a dozen people mining the information could generate hypotheses that could lead to sort of clinical trial proposals. And all of a sudden, then, its output was equivalent to a major site. And that's when I really saw in my career then the asymmetry you could have by combining computer science and mining data to create new medicines. And once we'd done that, we saw that was a Not quite a solved problem, but we saw the potential. And the next challenge then became, which actually then took me the next 15-plus years, was to then think about how you could use that same philosophy, that same approach of mining data and applying machine learning, not just to find new indications for drugs, but actually invent new drugs. Can we create new medicines using this way? And around 2006, we published a paper in Nature Biotech. Which was about the global map in a pharmacological space. And what was interesting about this paper is that we integrated, for the first time, huge amounts of internal data inside of a company, all the different pharmacology data, which normally would be residing within various data tools. You run a project, here's the data, drug moves forward, you collect that data historically, but weren't really getting value from it. But actually bringing all that data together and then integrating with the work John Overton was doing by then, which in a company called Informatica, where he was collecting together all of the data in JMedChem, all of the pharmacology and medicinal chemistry data, and turning that from data that was in a PDF actually into a machine-readable format, connecting chemical structures to targets to sequences and integrating that internal and external data. And you can imagine a company like Pfizer, huge amount of data. You know, they'd integrated with Wyeth, integrated with Pharmacia. And so it is wealth of data to be integrated. And actually, from that sort of large-scale data integration, For the first time, we felt we could survey that pharmacological space. And that then started to become the basis then of realizing we had to form foundations then for building large machine learning models. And then we built something like a few hundred Bayesian models. And that allowed us then to really understand even the polypharmacology of drugs and how every drug was far more promiscuous than we may have realized previously. And as far as I'm aware, that was one of the very first sort of machine learning papers in sort of the pharmacological space that was published.

Published.

And that then became the basis of thinking about if we were going to design drugs by AI. You could think of it as a set of foundational layers that need to be set. You actually first need the data. Don't forget, the mid to early first decade of this century, we didn't have the data. It was still in libraries and PDFs and journals. Now, we have amazing resources like EBI and OpenTarget and everything in our HHS building. But in those days, before Campbell came about, there was very little data in which to build those models. And then on top, once you have that data and there's a machine, then you need to think about which algorithms to apply and which machine learning models to build. And then of course, there's another level of AI on top of that, particularly if you want to think about design algorithms and decision-making now applied. So we always saw this as almost like a 3-layer cake. And actually, the past 15 years was almost about building each of those layers as we moved forward. Then once we saw the potential of that, in 2007, I was lucky enough to be offered a professorship at the University of Dundee in Scotland, which is a fantastic university. Professor Mike Ferguson was one of the key people in bringing me to Dundee. He set up a drug discovery unit at the time, which was one of the very first professional drug discovery units inside a university. He was really focused on particularly neglected diseases. And I saw this then as a chance to really then strike out and think about if we were going to then think about how we're going to create, uh, an artificial intelligence machine learning approach drug design. This wasn't a short-term project. This was actually a major research endeavor. And becoming an academic would provide me with the freedom to go and build this. And so that's why I choose to become an academic. In 2007. I was about 35 at the time, and it allowed me then that sort of freedom to then go off and start to then build, put together a brilliant, small but brilliant team. People like Jeremy Besnard was my student, became employee number 1 of Accenture, basically. His work, he's the first author paper in the Nature paper we published on the algorithms. Richard Bickerton became another key co-founder who then was also then building a lot of the information systems behind the scenes that we were using. But actually, with a small team, we then managed to sort of demonstrate, you know, at least in an academic setting, that an algorithm actually could be used to create new IP, could actually be used to, to generate new chemicals and solve design problems against multiple objectives. That paper then we published back in 2012, really showing that particularly a combination of Bayesian machine learning and evolutionary design approaches could create IP, and now allow you to sort of finely tune the creation of new designs towards a whole spectrum of properties and pharmacologies you wish to have. And if you're an academic, you know, it's great, you published a Nature paper, you know, it's the move on to the next grant. Uh, but that wasn't my, my mission. You know, my mission really was almost to see academia as that sabbatical, you know, to take off and build what you needed to build. But if you want to turn an invention into an innovation, then you really need to engage with the real world. And that's why we formed Accenture, because our mission has always been how do we actually revolutionize this industry? It costs over $2 billion to bring a drug to market. And we thought if we're going to do this, we need to show it in the real world. And I could talk further about the history of Accenture and how we built it, but I probably— it's a good place to stop in terms of my introduction. And that's why it's been an absolute pleasure to get Accenture to the position it is now as one of Britain's leading biotech. And also, I think actually one of the real pioneers of this whole new field now of tech-enabled pharma, of how now we're thinking about using AI to design next generation of medicines.

So I'm going to take you back a little bit in there because there's a lot of highlights in there which include you thinking orthogonally about the way that the industry works and doesn't work. So, you know, it sounds like you've had some of the opportunities to do that, but you must have brought some of that mindset to all of those points of connection with Pfizer and so forth. Um, what were you seeing along the way? Was it a lack of interest in thinking about it differently? Was it lack of talent in thinking about it differently? Was it, you know, what, what did you see that others weren't seeing at that point?

That's a really interesting question. And in fact, uh, Pfizer at that time, so the first decade or so of this, and particularly in Sandwich where you work in, was, um, an incredible place. A lot of intellectual discussion was going on about the nature of the fever or medicinal chemistry, etc. A lot of new ideas, things like, you know, ligand efficiency, druggable genome, network pharmacology. These ideas are all fermented in that sort of, uh, the discussions in the coffee rooms and how we were thinking of a project. So in fact, we did see a lot of sort of, um, open discussion inside the company, people wanting to think about how you turn medicinal chemistry from an art into a science. How do we formalize it, you know? And I think there was a That was a great place to learn the trade, so to say. I think the difference I brought, I came at it as an outsider. I'd— an insider and an outsider. I'd been an insider because I was already thinking about drug discovery from my DPhil days in Oxford, designing drugs and being involved in that sense. But at the same time, early in my career, actually, I used to work in the steel industry. I got a scholarship from British Steel. I used to spend my summers since I was 16 working in the labs in Port Talbot. Spent a year there after my— degree in chemistry in Manchester whilst I was looking for a fantastic PhD to do, which turned out to be working on HIV at Oxford. But actually, I learned a lot in that time. And what I learned was how an industry gets transformed, how technology and economic factors are some of the real drivers behind innovation. And I worked in a place where, in the '80s, there used to be nearly 30,000 people working in this huge integrated steel plant. Plant that was almost several miles long. And then sort of new technology, new working practices, you know, by the end of the decade, it was down to about sort of 3,000, 4,000 people working there, you know, but the output had gone up nearly tenfold. So what you saw then was how economics, new technology, the market pressures, you know, can drive innovation. Obviously, you know, plenty of sort of social issues can come with that as well. But actually, when I went to the pharma industry, I came with a very different sort of experience of life and what industry and pressure has. And when I first went to Pfizer, I remember I was offered a lectureship or a job at Pfizer, and I thought, this is like getting tenure, because in those days— very different today— but two decades ago, a job in big pharma was like a job for life. And it felt like, as I said, getting tenure. Great place to do research. You had a very different sort of attitude. And I don't think the winds of creative destruction, so to say, were felt as strongly inside Big Pharma at that time as it is today, and potentially how I saw it. So I always had this outsider's view that the world is going to change. Is it going to be forced upon us? Are we going to be agents in driving that change? And of course, we then started to move to an entirely data-driven approach. This obviously started with the wealth of data that was being generated from the Human Genome Project and all of the wealth that genomics started to bring. And then you started to then think about how do you use data. That obviously then takes you to machine learning and data science. Then think about that intersect. And I was really struck by a quote from the management guru Peter Drucker, who said pharmaceutical industry is an information industry. And what he meant by that was that pill you buy, it might only cost a few cents to make potentially, but the reason it's charged what it is because of the value it brings. And the value is not the actual chemical synthesis of the few molecules, et cetera, of that 500 megs of chemistry that's inside our pill. It's all of the intellectual property. It's all of a huge amount of research and development of not just that pill, maybe all of the other pills that failed. That has led to that invention. And that's where the real value is. That's what he meant by that. But I then took it a step further and think, if you think of it as an information industry, we then think about it as an information technology industry. We then think about it as a data industry. And what then are the tools, the sciences and approaches then that are going to take us then to follow that logic through?

Interesting. I love that. I mean, you know, A, that you separate innovation and invention. I mean, that's been a theme of ours for a long time, right? Is that industry tends to treat innovation as if it's the whole thing. You get, no, no, there's a, there's a very big difference. And there's a quote that we've built on Drucker's, which, you know, essentially the molecule you discover and the molecule you launch are the same thing. You know, the only difference between those two points is what you've learned along the way and what you collected as evidence. But, um, and I'm going to come back— there's a few other things that you said which I love, which, you know, A, the asymmetry. I mean, this whole— the subtitle of this podcast is Asymmetric Learning. You know, what you learn that somebody else doesn't learn. Um, and even the idea of plans Serendipity. I love that idea of, you know, I've written a lot about serendipity not just being about luck, but about observation and being prepared.

Absolutely. Pasteur's sort of quote, I think, is a wonderful quote about being sort of prepared minds, you know, spotting serendipity. And it was actually what drove me when we were thinking about that serendipity concept and how we can engineer it was we discovered that, you know, penicillin, we all remember Alex Fleming discovering it and the value it was. And I put up this quote describing the discovery of penicillin. And then I pressed a button again, and it came up who said it. It was a guy called John Tyndall in something like 1857. Actually, penicillin was discovered and forgotten about 7 or 8 times before Fleming actually then rediscovered it. And then I realized that some of this was already sitting in literature. And imagine if we had rediscovered penicillin a couple of decades earlier. And all the people that, you know, died of various sort of traumatic injuries in World War I, you know, potentially then could have been saved. And the impact obviously would have had. And it really did struck me then about information being lost and how then we could potentially find and mine for it. And that was actually— was something became very powerful in why I was— it became so interested in then about how then we could use technology to then link this information together and hopefully enhance our own creativity.

Yeah, and it's interesting because I think that you link that concept of serendipity and that long process of discovery and development. And I mean, even manufacturing— I mean, that was the issue with penicillin, was even being able to manufacture enough for one person in the beginning is, um, is that this is— it's complex. And I think you used the, the phrase, you know, the industry became a data industry. But, and I think the sources of data are, you know, I'll be interested in your definition of what data is and what it, you know, what it, what the definition of data isn't, because I think you talk about it in way more complex terms than most people have, because, you know, you've got that sort of path dependence, right? The data that we've got is the data we went looking for. But you just described a bunch of data that is, um, it's equivalent to sort of like the cholera epidemic, right? A lot of people wrote things down without a purpose in mind, and like the Genome Project is a lot like that. Um, so you've essentially set out to capture stuff that wasn't built for a purpose, but it was built for a reason and started to put a purpose together next to a few different datasets. Is that, is that a reasonable statement? It is.

I've always been interested in this idea that individual sort of data, small datasets by itself, you know, often are created to answer specific questions. But there's another level of value that comes when the data starts to be integrated. And then not just one type of data, but then a multimodal approach where whatever relationships you can learn about of our wealth of information connected together in the literature, how that links then potentially to connections we see in genome data. And bringing all these together then has a huge potential for knowledge discovery. Within it. And that's what I think the advantage we have in this generation that previously we didn't have. It isn't just about the, the isolated piece of information. And if you're well-read enough in a library, maybe to make those connections. But actually now it's how do we build systems that make that? And there's a couple of things actually that really drove me to think like this. One of them was, I think it's, I think it's in the poem The Rock by T.S. Eliot, where he says, Where's the information we lost in data? Where's the, the, the wisdom we've lost in knowledge? And it's about that data, information, knowledge, wisdom sort of hierarchy that becomes really important. And that's ultimately what we're trying to build. How do you build sort of true insights from just the data? And I think sometimes we do get a bit stuck on just the data and trends within it. And that's why when we were building Accenture, we thought about it's beyond, beyond just predictions. It's about then how you then start to then think about design design as a creative endeavor, which hopefully is more on the side of knowledge and wisdom and data information as we drive towards it. That idea then about the levels of information and insight is an important way, understanding how you distinguish between data, information, knowledge, and hopefully wisdom sometime. That's been a key driver in how I've been actually thinking about why firstly, these integrated, this wealth of data systems is so important. But actually the data systems by themselves don't give you the answers. You still need to think about creatively about what questions you want to ask. One of the things I encourage my staff about at Accenture, asking the right question is as important as thinking about the right answer. And that's vital. And I think one of the key things that I think distinguish us and how we thought about building Accenture was to ask these questions, you know, Could we automate creativity? Could we automate how we want to do drug design? How would we apply AI, you know, to the whole processes? And I think sometimes by challenging your assumptions and drilling down to about why do we do things in a certain way is as important. I think asking the right question is as important as building all the technology to try to find the right answer.

Yeah, I like that. And any Drew to mind, um, Creativity Inc., the Ed Catmull book about Pixar versus Disney. In your description of the sort of steel plant going from 30,000 down to 3,000 while being more productive, same thing happened, right? The technology didn't exist to make the first Pixar movies when they started to set out to make the first Pixar movies. And it sounds like the technology didn't exist when you set out to do what you're doing. Um, how do you see that role of asking questions and finding answers in, in technology or in data.

Oh no, that's— you're absolutely right, Mike. We— the technology didn't exist when we started to build a question, build Accenture. The key thing was asking the right question. So when we— so that's what we started off with. We started off with the concept of if the pharmaceutical industry is an information industry, how could we then start to— and if we thought that one of the key sort of economic challenges is how do we, you know, design better drugs faster? How then are we going to do that? So that led us to then thinking about, okay, we want to take a machine learning AI approach. Why do we want to do that? Because I think one of the challenges we face now is the complexity of decision-making and the wealth of data that one needs to integrate. So, you know, if you look at sort of the human creativity that a drug hunter, a drug designer has, you have to integrate a huge amount of data into their projects, into their design. As you know, the positioning of one atom sometimes is the difference between that being a billion-dollar drug potentially and being a lifesaving medicine and being also one component fails in the clinic. Because that's the difference between metabolism, whichever selectivity enzymes it'll be binding to. So how then fundamentally do you precision engineer drugs to that level? And that's where we wanted to take a much more precise approach using AI. So that's been a key sort of driver within that. And when we were doing it, we were then sort of our first sort of starting point in how we're going to think about these algorithms was actually starting with the human. So the way we approached it wasn't a necessarily computer-first approach, you know, start off with a deep learning algorithm and move forward. We actually started off by thinking about what is the human process of creativity. Actually, for the first year or so of my research, it was all about what is the psychology of creativity. There's different aspects one can consider. One of them is coming back to your serendipity argument. What's that creative point when you read one paper, read something else, and then make that connection? That connectivity between data, that's important. Sometimes you have very pedantic friends who might say, "Don't make that connection," because they want their data to be very solid. And other times, maybe your more free-thinking, creative friends would see those long-distance connections. And maybe 90% of their idea is not very good, but 10% might be absolute gems. So thinking about connectivity between data became important. And then the other thing became really important is thinking about that sort of act of creativity that the drug designer goes through in designing a drug. How are we going to mimic that? And we actually led in our thinking to think that sort of Darwinian evolution and evolutionary computing was actually a natural way to think about exploring the space. It actually also fits with some of the interesting analogies of human creativity. It's about generating lots of potential options, lots of ideas, generating then new models to then select what you think are the best options. We were building models, obviously, from all this wealth of pharmacology data. It doesn't matter if your models are Bayesian, recursive methods, using reinforcement learning. It doesn't matter if you use them in multitask learning. It's all about then creating that sort of selection pressure within it. And in many ways, if you're an artist, your selection pressure is your concept of beauty, whatever you see. And here, our selection pressure was the data from the models that we were— the data that was feeding into the models we were using. So therefore, we realized that concept of combining machine learning and evolutionary computing actually could give us that sort of mimicry of that creative process. And when we first started to run the algorithms, we saw that we were starting to get amazing results, which led to that Nature paper. And then that gave us a confidence then to think this could be a methodology. These days, of course, we use a wide range of different methodologies and a wide range of different algorithms inside the company to do that. But what led us first was starting off with that question, then starting off by sort of we wanted to break it down between the mimicry of who does it good right now. Humans do it good. Humans are very good at creativity and design. How might we sort of mimic that in an algorithm? And that was the sort of the first set of logical path we followed actually to create the technology.

Yeah, how interesting. Now, Brian Christian, who was on this podcast, wrote Algorithms to Live By and some of the other stuff on AI. It was really interesting as he began to look at, you know, how do, how do machines learn? And the sort of the fact that it's not monolithic, right? That you have different ways that the machines can learn, whether it's coaching or whether it's, you know, making mistakes and, you know, learning from, as you said. So interested maybe just in your view on, you know, people talk about AI drug discovery as if it's all the same thing with slightly smarter people or slightly, you know, not so smart people. Clearly this isn't one box of companies in the AI discovery space. How do you address being called an AI discovery group?

We absolutely don't mind being called an AI drug discovery group because that's what we do. Absolutely. Now moving into the clinic, now really thinking about how AI is used not just to design the drugs, but also select the patients, select the targets. You see that end-to-end process developing. But you're right. I mean, we're on the verge now of a whole range of different computational approaches being used by a range of different companies where it's using AI to, you know, analyze a huge amount of image data or cell data, genomic data, where it's using AI to then help sort of, as we said earlier, repurpose those drugs, discover new uses. That's a perfectly valid use of that approach. Other people, of course, using more stronger computational approaches, physics-based methods, which some people sort of mix with AI. We absolutely believe the future is combining both, combining machine learning and physics-based methods. They both have advantages, and that's where we are today. In fact, We, we have a significant number of people developing both methodologies now and integrating them together because many ways the physics-based method give you new data actually creates another dataset from which you can then build models, you know, that can, that can power the AI. So it's an interesting sort of, I think, an incredibly fascinating world now and a biotech sector starting to develop And also what's interesting, I think, about these companies is that, you know, they have an ambition. They are frustrated by the cost development. They want to bring new ways of thinking in. And you've got some great founders of these companies and you've got some great people working in them. Each one has got, you know, different strengths and approaches. But the way Accenture approaches it is that we started off and I think we've really were the first in this, in using AI to actually create molecules, in how we use it to actually optimize the compounds. We're always interested in the creation of new drugs. And as I said earlier, that was my motivation. We sort of done drug repurposing using computational approaches, how to create new medicines. And that's important for us because if we think about the requirements of the patient, a drug really is a precision-engineered piece of technology. As we said earlier, that information that goes into that pill you have is a precision-engineered approach where one atom difference is the difference between it being a great medicine and not being a great medicine. And so therefore, bringing that wealth of information then into that design process was what drove us. And then as the company developed, particularly as a strategy of doing partnerships and working in the real world, and that was an important piece we found, the difference between the invention any innovation is taking it to real-world projects. That's why we set up Accenture, because the best way— you can only really sort of test your algorithms out in the wild on real projects to see what real results. So it's no point just doing it in a closed environment, in sort of a safe sandbox. You've got to be out there, you've got to show real value. So therefore, being exposed to real-world drug discovery projects, testing algorithms in that space became absolutely vital to validating the technology. And that's where it's important then to really show that as we then got exposed to more, we started to see more problems. And then the more we do, the more we learn. You discover this kind of problem, this kind of modality now needs to be added as a new module to the algorithms, to the whole platform. And then you started to see greater problems you need to solve. You know, it's— you then need to move upstream to thinking about how we can use AI for target identification. And some of the work now I see going on inside the company combining deep learning and knowledge graphs and vector embedding, I see it's going back almost 20 years to that original hypothesis generation, but now on a far more powerful scale. And we start to see evidence that deep learning approaches can identify those hypotheses actually years before they appear in the literature. And that's an incredible sort of piece of work. And then, of course, moving downstream into the clinic.. And the work we're doing now with a network of hospitals building biobanks of patient samples. And then you want to use AI to have analyzed at a single cell level this complex mixture of cells you have in actual cancer patient tissue material, cancer cells, immune cells, stroma. You want to be able to easily distinguish each individual cell, identify what it is, and how are drugs affecting it. And that's again where AI approaches come to. So what we find now And I think everyone in many ways, it's not one AI algorithm, it's actually a whole suite of different algorithms. And in many ways, the focus of some companies often depends on maybe which part of a process they're focusing on. Where Accenture is going is that we're interested in the whole process. We're interested in the end-to-end approach of idea, ideation, all the way to idea, IND, all the way hopefully then to the FDA. And the reason we are interested like that, because it struck us early on, is that if you have a new technology, you know, you don't just become an add-on to an existing process. In many ways, that existing process we have in pharma itself is like the fossilization of older technologies. And what we find then is that the new approaches now of real data-led, AI-first approaches allows you actually to think about rewriting, re-engineering that process in an AI-first process from scratch. That's what's really exciting about this. If you think about the impact of technology on the banking industry, we didn't go and invent robots in high street banks that sit there counting out your money. You reinvent the whole thing. So now it's an app on your phone where you can transfer money across the world in seconds. And so therefore, the technology allows you to rethink the processes of the industry. And potentially then reinvent it. And that's, we think, has been vital to all of this.

It's really interesting. I was thinking of parallels. I had a chat off podcast with this Formula 1 designer about their process of designing cars. He said, well, there's this— you said you probably don't know this, but there's a huge divide in our industry, the engineers and the designers. There are some people who believe you can do it all in silico, and some people like us but think you can't. I wrote a piece about this, and it was, it was asking how many prototypes they build between races, you know. So they go to Monaco and they go somewhere else, and then they build 1,000 prototypes between races. He said, you know, you've got to take into account weather, you've got to take into account track conditions, whether you're first versus sixth on the grid. He said, but all those prototypes are useless because they're all in silico until you put the car on the track, and then you learn something that you couldn't possibly have known until you put the car on the track. He said, and our philosophy is The track's right, the computer's not. But some companies don't believe that. They think the driver's doing something wrong or the mechanics have done something wrong. But he said, you have to believe that observation is going to trump prediction in this case. So it sounds like that's your philosophy as well in that approach to learning.

No, you're absolutely right. So I would almost say active learning approaches, which is around how do you want to learn as quickly as possible? Is it more important than deep learning? And I want to build upon that because I think what you're describing actually is some interesting analogies with how Accenture operates, actually. So for example, a lot of our design systems we call Centaur. And that derives from the concept of Centaur chess, which actually is a combination of using algorithms and humans together, almost in a freestyle. And when Kasparov lost against Deep Blue back in the late '90s, he didn't go off and say, this was the end of chess. He became interested in the idea of what happens if you merge human talent and the algorithms together and with it. And they went off— I think it was in Leon in Spain. Every year, they had a competition. There's some really interesting observations they saw then when they started to combine machines and humans playing together. Firstly, they found a master, which is several levels below grandmaster, using the right set of processes, how they use the algorithms, could outperform a grandmaster, even if the grandmaster is also using an algorithm. And you could start to level the playing field of skill. The other things they started to find was the games became sort of flawless in some degree, technically flawless. People stopped making mistakes. And that's an important element in a company. And so we, we thought these were really interesting analogies of how we saw, particularly we want to introduce— because even when we start off, first we start off with design algorithms. But as you know, the whole process of drug discovery is incredibly complex. We haven't sort of fully automated that yet. Our philosophy is to encode and automate as much as possible, and we always look in how we, we move up that sort of knowledge chain. But of course, you bring out expert— and that's why it was important to have expert drug hunters like Andy Bell when he joined the company, um, you know, bought two drugs to market in his career, you know, Sildenafil and Vifent. And so that was, that's combining that insight. And what we want to know is like, Andy, how did you think like that? You know, how do we, how do we put that knowledge into a, into a black box, so to say, into an algorithm? And that combination then of understanding human expertise, understanding how it's relation. So therefore, you know, as Kasparov says, it isn't just humans plus machines. Human plus machine plus the right process together together is what gives that productivity. And that's been a key driver in how we think about tech at Accenture. And there's also another element to it I think is important for us to think about. So, I mean, one of the paradoxes we have in pharma is that we have more data than ever. We have more knowledge than ever. We have incredibly sophisticated, experienced drug hunters working in our industry. We have incredible technology now that we never had. Now we can do single-cell sequencing. Whilst, you know, 20 years ago it cost us nearly a billion dollars to do one human genome. And yet in that time, you know, cost of creating a drug has continued to go up. It's about at least $2 billion now, latest figures. Uh, the number of drugs the FDA approves every year is consistent. It's about 40 to 50, you know, in a good year. And how is that? How we got this paradox of effectively decreasing a product which people call e-room as you know, in the literature. And I think part of that actually is the paradox of we've actually made the process more complex, you know, compared to the drug discovery project today to when Paul Janssen was working. You know, Paul Janssen, I think, has his name on nearly 80 drugs. Incredible, you know, a long order more than even more successful drug discoverers today, you know. And yet, if we probably look at the, the quality of data The quality of insights on any one of our projects today is probably far greater than maybe Paul's projects back in the '60s. And the level of regulatory hurdles we have to overcome, et cetera. So in some senses, you could argue we've been running into standstill. All our new technology has actually made the process more complex, much more complicated. And in fact, the thing that hasn't changed in all that time is the cognitive bandwidth of the human mind. Now, we expect the poor drug designer to absorb all this new information we're generating, and actually then incorporate that into the design or an optimization of a new medicine. And I think that's where the challenges come. And that's the only thing that hasn't changed in the industry. Huge amounts of new data, huge amounts of technology, and yet, our own cognitive bandwidth, I think, has been a rate-limiting step. And if you follow that logic, that's why I think AI is going to be the game changer that I believe it will be.

I think the next set of questions actually, and Jack Scannell, who came up with the Room's Law, I mean, he's one of our advisors. He's been on this podcast too, talking about this issue of productivity. And a lot of the problem that I see is in the development space, right? You talk about the rate of productivity. Drugs, but it's still essentially a human process. And it's also largely defensive, right? So it is, you know, that Catmull in Creativity Inc. talks about feeding the beast. You know, Disney was largely feeding the beast. You have 200 people in new product planning, you've got, you know, you've got 5,000 marketers and salespeople. Those people need to be paid, they need to have something to do, they need to be given. So the costs go up, and then the CRO costs go up, and everyone's got the same drug going to the same clinical trial sites, so the cost per patient goes up. So of course we're feeding megalith, not just the monolith, but, uh, but that, that challenge. So the development process, I mean, that's where we spend all of our time focusing, is still largely human prior prejudice pyramid of, I'm going to say, terrible decisions that really haven't looked very hard at decision science in that world. Um, so I guess the key question, the first question for you is, when an AI-discovered drug hits hits the pipeline of a pharmaceutical company or your own pipeline? What is it better at? Is it expected to go faster? Is it expected to have higher probability of success? Is it, you know, what's, what is an AI drug look like? What's its phenotype when it hits a drug pipeline?

That's a great question. Um, in fact, I'm going to break that down, Mike, because there's several ways to, to think about that. We, you know, our goal is actually to produce better drugs faster. What do we mean by that? We want to create a world where we can start to think about changing economics where if we can bring new innovations forward quicker, quicker in discovery and hopefully quicker into development— and I'll explain how we figured about that shortly— then actually, uh, that's a win for everybody. It's a win for a patient, it's a win for science, it's a win for the, the, the stakeholders. But also then, better. What do we mean by better? If we can try to bring as many models and as many potential, uh, observations where it's computational or, um, real world, as you mentioned very soon, um, then hopefully then we can get to a position where, um, we are creating hopefully better medicines. And better is not just, uh, the potential safety profile or the potential quality of that medicine in terms of its pharmaceutical properties, but also it's better in terms of can we target the right medicine to the right patients, and are the right patients receiving it so there's a much better signal within that particular cohort. So there's various elements now when I sort of break down of how we think about that. So firstly, we see our fundamental goal is to— how do you fundamentally change those economics? You know, what we call a strategy shift in the curve. So, um, what curve I'm talking about is a sort of economic investment and then sort of sales and return as you go forward. So imagine sort of a negative sine wave, you know, uh, but you can imagine, like, uh, take a dollar bill. It's quite strange I'm describing on a podcast, a graph. But so imagine, imagine the dollar sign, the S with the line through it, and tilt it 90 degrees, you know, so you're starting off at zero almost. And your y-axis going up is sort of profit and bringing money in. And your y-axis going down from zero is cost and spending money, spending money in R&D, for example. And of course, your x-axis going across is time.. And that, that sort of S sitting on its side is like we need to spend a lot of money, you know, tens of millions potentially in discovery, hundreds of millions then going deeper into a red in development. And then we start to hopefully get approval by regulatory agencies, hopefully start with drug selling, and then we start making money back. And hopefully then we get into profit, start making money, hits its peak sales, then maybe we decline into competition then we go off pattern as generic competition, and then maybe end up, who knows, close to zero again. So if we think about that curve, you know, and that curve is something which represents maybe at least, as we said, potentially $2 billion, you know, in the red, how then can we shift our curve to make our timeline shorter and to make the cost in development sooner? And fundamentally, we believe technology and technology-enhanced decision-making can help us do that. So that's, that's really the the base philosophy of Accenture. That's what we're trying to do now. If we think about then moving it into development and, you know, we now have our first AI-designed drugs in the clinic, hopefully in the next few months you'll be hearing about more from Accenture moving into the clinic as well. And what we're also thinking about now is how we can be as innovative in the clinic as we have been in discovery. And we were very fortunate this year to, uh, for, uh, Mike Crams to join us as our Chief Quantitative Medicine Officer. The quantitative piece is an important, uh, uh, addition to his title. So before joining us, Mike was head of J&J's Quantitative Science Department, and he had 500 people working for him. And he joined our team in Vienna where we have our Precision Medicine platform based, and that's really important for us. So, and, uh, I knew Mike 20 years years ago. And I first met Mike when, um, he was running the very first adaptive Bayesian clinical trial. He had a room full of laptops he connected together. And what was interesting about his approach, he was letting the trial make its own decisions. The algorithmic decision about which patients, which to dose, etc. And of course, Mike then became one of the real, uh, pioneers and champions in the whole industry. Of adaptive Bayesian trials and using continuous monitoring, continuous response of how we're going to do it. So basically, we want a world where trials start to make their own decisions, start to run themselves. So what we're interested in and where we see the future now is combining that sort of automated algorithm decision-making with how we run a trial and combining that then with much more precision about how we identify and select patients. And that's where the, um, biobank-based precision medicine approach using actual patient tissues comes into play. So earlier this year, our team in Vienna published what I think is one of the landmark clinical trials in the past 10 years, a trial called EXALT-1. And this was a trial that for the first time showed that an AI algorithm can improve patient outcomes in oncology. So what the trial was is that we developed a system where we are taking patients' samples for cancer. So the first trial was in blood cancer.. And within about 5 days, we were able to take that tissue sample, able then to build a high-content model around it using AI machine learning approaches, then to really do sort of analysis so we could see how screening, say, over 100 different cancer drugs gives that individual patient's tumor sample and see how individual cancer cells were responding. And how cancer cells are being affected relative to immune cells, relative to normal cells. You could really find, is your drug selectively affecting the cancer cells? When we run that, we're able to then run effectively a personalized screen. This for me is true personalized medicine. And then report the top drugs from that screen back to the clinicians who are then able to prescribe it. We run our trial. And in fact, when we looked at the group there where the prospective trial where the physicians recommended with drugs prioritized by the platform, over 25% of the patients were surviving nearly 4 years longer. We had a hazard ratio of about 0.53, an ORR, objective response rate, of about 55%. And I thought this was an amazing result. It did for us give us real clinical validation of the use of these AI models then in improving clinical outcomes, that actually what we can test in a laboratory does give us predictive power, what we can see in, in, in things. So this, this is where we're going. This is where we think the future of sort of AI-based clinical development would be. The combination then of identifying the right patients directly using these AI-driven precision medicine approaches, combined then with how one thinks about continuous monitoring of adaptive-based trial approaches. And we think that's a really interesting and unique combination of things and technologies to bring together for development.

It's almost the definition of planned serendipity, right, which is that you're going to notice more things as you begin to study that allow you to pivot, which, you know, many traditional pharma companies is a bad thing, right? You know, you know, as I say, mostly Viagra would have been stopped in most places, but seeing an opportunity to go forward is, uh, was the, was the, was the intuitive step, right, that you had to go prove there's a market, you got to go do other things. But, you know, if you, if you're only doing simple proofs of principle, simple proofs of concept, of course you don't see the stuff that would allow you to adapt Um, but I'm very aware that I could talk to you for another few hours.

Me too, Max.

It'd be great enjoying this conversation, getting to the bottom of, uh, of what you're bringing to the, to the table. Um, I'm going to ask you some other questions now about, you know, uh, in the last 3 minutes, I think, of, uh, what do you read for fun? You know, what do you think the next couple of years looks like in terms solutions that you wish you had 5 years ago, 10 years ago?

Oh, that's a great question. I look at some of the things we're developing now inside the company, and I wish I had them when I was— I wish I was smart enough when I was at Pfizer to invent them compared to some of the people we have working here now. The way that I'm now seeing that knowledge and hypotheses and the scientific process now is potentially becoming predictable.

Possible.

In fact, we are predicting, you know, those relationships between targets and disease and when papers may occur, you know, potentially years before it. I think that's an incredible piece of technology that's emerging because that then opens up the idea of like, actually, could we increase the speed of science even, even more by understanding and directing our resources? I think, you know, these AI systems might be better at giving out grants and maybe some grant committees in that sort of Understanding the, you know, it's almost a structure to knowledge and how knowledge is emerging. And that's really exciting. The other thing as well I'm really excited by, you know, particularly inside Accenture, is the work I just described in Exalt One. And this really interesting idea now that the— I think the, you know, your pharma company of the future is this concept of combining this understanding of of patient information. This idea that we can actually use that data, you know, to help identify what's the best drug for you, you know. And actually, that's an incredibly powerful approach. And that idea then of true personalized medicine now being possible within that— we've talked about this for 20 years, and I think now we've seen some of the first clinical trials actually starting to show this. And that's not just going to help us design better drugs more targeted drugs. But you can also see sort of other approaches of expanding that further about how that becomes whole platforms of how we think about treating sort of diseases. What I read for fun, I'm a real bibliophile, I buy, you know, and one day when I retire, I hope to read all the books I buy. But, um, but I, I, um, I'm a big fan of sort of history of ideas, actually, Mike. It's, uh, I think that's really important to understand the providence where ideas come from because it allows you then to, to, you know, a day where everything we think about is important. I mean, Keynes said this quote about, um, you know, people who think they're practical men are often in hock to some old ideas from some old philosophers or old economists, in a way, you know. So many of the things that define our assumptions of how we operate, you know, are set by philosophers and economists and scientists, you know, maybe many generations ago. Understanding where those ideas come from, I think, is vital for us break our assumptions and then go forward to do it. Actually, Peter Watson's book Terra Blute and his other book Ideas are suggestions I would read for anyone who's interested in the history of ideas.

Nice. We'll put those on the, on the show notes. And actually, I think you're going to follow. So Steven Johnson, who wrote the book Where Good Ideas Come From, precedes you in the release here. So I really appreciate that.

If someone wants to find out more about you, Are you easy to find socially? No, I'm completely stealth, actually, Mike. So contact Accenture. Interestingly enough, so I run an AI company, so I'm not on social, not on Facebook, not on LinkedIn, not on Twitter. I think that tells you something about how I think AI and data has been used by people.

Yes, what a great way to start. I will certainly include links to Accenture in the show notes. But Andrew, I knew this would be fun, and it's been even more fun than I thought. Thought it would be. So thank you for your humor, your depth, and your knowledge.

It's been a real pleasure. Mike, thanks for your time, and thanks to the audience for listening.

That's it for this week's episode of Idea Collider. To continue the conversation, visit our website at ideapharma.com. Follow us in your favorite podcast app so you never miss an episode. Don't forget to rate and review us on Apple Podcasts. Until next time, I'm Mike Rhea, wishing you great success.