All right, well, we have a terrific panel. You've already heard from Hannah about Fiona, Alison, and Pushmeet. We're representing here academics, life science industry, and the IT world. And maybe just a quick intro, I think digital biology or the Large Language Life Models, LLLM, is moving at a pace we've never seen. Just last week, Evo was published in Science. This Wednesday, the Human Cell Atlas Foundation Model will be published in Nature. We've had multiple human methylome models published in the last couple of weeks at [? Bolts One ?] yesterday. It's just dizzying. I've never seen anything like it in my life. So we're going to get into this. That's, of course, the kind of life science side, and of course, it's much broader as already introduced by James and Jennifer in terms of the biomedical applications. So maybe we can just start off with where are things now. So, Alison, maybe you can lead us off, your experience.
OK, so I work in imaging, particularly sort of ultrasound imaging. So I've particularly seen, over the last 20 years, how things have dramatically changed. Ultrasound is a great example, where humans find it very difficult to acquire data or interpret data, and it's all about understanding the physics and acoustic patterns. So along came AI, and it's really transformed what you can do with an ultrasound device, because you can now—provided you've got the very large data sets that are required to do this, you can now find things in ultrasound scans that humans find difficult, and you can also guide people in how to use an ultrasound probe. And this is still emerging as an area, but I think it's probably one of the most exciting breakthroughs in ultrasound, because you no longer require to be able to train for a long, long time. You can now put that into the front end of the device to be able to guide you. So that's just as an example.
Yes, thanks. So, obviously, heading up research at a pharmaceutical company, we're really using AI now right across the R&D process to advance new medicines, starting off with identifying new targets, which comes down to understanding diseases better. And that's where you use big data sets, like the Cell Atlas, but also things like human genetics data, population data. The second area is, once you have ideas for targets, how do you develop a medicine? So that would either be a small molecule. So that's chemistry, generative chemistry, and we use technologies like AlphaFold to design molecules, optimizing biologics or ex-RNA gene editing strategies. A very important area is in predictive toxicology, so making drugs safer as they advance, and then in clinical trial design as well. And there's a lot of innovations there using AI.
So I lead the science team, and it's very difficult to talk about what is happening, because it's just so much across the board. I mean, if you look at topics, I can start with biology. I mean, James already mentioned the AlphaFold and AlphaMissense work. I mean, really, what we are trying to do is decipher the genome, really, understand what is encoded in the genome. And as Jennifer was mentioning, if we have that amazing understanding, it will just accelerate the amount of work that we can do by using that amazing understanding about what is the actual semantics of the DNA. Now, if you look across the fields beyond biology, in every field, we are seeing a transformation. James mentioned our work on weather prediction, GraphCast. Just to give you an idea, GraphCast makes predictions about near-term forecasting. It used to take supercomputer simulations for days to come up with a prediction like that, and now, it can be done on a single chip in a minute. That's like multiple orders of magnitude sort of difference. The same thing in materials. There are around 30,000 inorganic crystals which we knew were stable. In our work on graph neural networks for material exploration, we showed that there were 2.2 million inorganic crystals which were considered stable that scientists did not know about, and we shared 400,000 of them last year. And scientists now have discovered that many of the candidate materials that we had listed are candidate materials for high-temperature superconductors, for better battery design materials, and so on. So in every area of science that we are looking at, whether it's biology, structural biology, genomics, protein design, it's just transforming what is out there.
No question. I mean, Pushmeet and I were talking before we got started this morning about how structural biology used to be regarded as the physics of life science, and now, because of AlphaFold and many other models, and of course, its use during the pandemic, it's been transformed. So this gets us to this acceleration of discovery, which is, I think, a theme of this conference and how the science can be changed in almost a way that was inconceivable. So in many ways, this brings up the role of the scientists in AI, and now, we've seen things that started with the lab-in-a-loop concept. It's been fostered by Aviv Regev, and then the AI scientists or co-scientists. Most recently, just last week, James Zou published a preprint from Stanford about the virtual lab, where he had five domain agents that actually conceived nanobodies for SARS-CoV-2 that were very potent, essentially with infrequent human supervision, and that's, of course, going from virtual cell to virtual lab. So the limits here are something we've never seen before, some level of autonomy. So maybe we can start off by saying this is something that we've not ever actually experienced, and it challenges our thinking about the future of science and the hypothesis generation creativity potentially that we can derive. Who would like to start us on that path?
Well, I think it comes back to we're moving so fast, we've got to be careful. It's great to hear about all the excitement, but the notion of what a scientist is, is changing, and we need to also stand back and reflect on that. In our report that we published, which was discussing with many scientists, both AI scientists and across different areas of sciences, this is people's concerns. And it goes back to things like reproducibility, just basics, fundamentals. You had the slide at the beginning of the meeting where people used to get around a table and wanting to see the recipe repeated. Some of the principles haven't changed, and we need to, as scientific communities, demand of ourselves and in different research communities, that we really understand when we publish results that they are sound. That is partly thinking about openness. Openness, I use deliberately, because you can't make everything open. If you work in healthcare, your data is often sensitive, but that doesn't stop people maybe thinking of how they publish in different ways. Some communities are doing this, where you now are required to publish. If you can't publish your data, you have to be very rigorous in how you've evaluated what you do. So this is a responsibility for scientists to rethink in the age of AI how we do present our work and also how we do training of scientists. And not everyone is an AI scientist and maybe understands the methodology, so how do we train basic scientists to be able to trust the outputs? And another thing is the outputs, and AI makes errors in different ways to a human. So we've also got to understand the implications of that as well. So we've got some wonderful early examples, but I think these are the sorts of issues we as a scientific community, and in these areas, we need to really understand a lot better.
Fiona, maybe you could also take us through a bit on the teams at Novartis. Are they into this, or are they—
I think this is a very important point to bring people with us. That's one of our priorities. So we have more of the concept of human in the loop rather than the lab in the loop, and how can we really enable everyone to embrace new approaches? And there's a number of ways of doing that. I think making everyone's job much easier by using AI tools, making them accessible, so that people immediately see it as a benefit rather than a threat, that's one of the things. But also, I've changed some of the hierarchy of how we organize our labs. Historically, the wet lab scientists generated data that then the data scientist had to sort of analyze and feed back, and they would be center of the authorship on a paper, shall we say. But now, I've actually put the data scientists as the lab heads in some cases. So they're really the super users of computational means, and then they're sort of dictating the experimental procedures and the experiments to the wet lab scientists who then carry out those experiments. So that changes the way people think about data and use it, and it's the fastest way of really sort of driving change in the organization.
I think there are two sort of elements to it. One is AI systems as tools, and they are very powerful tools. They can make some things—they can make predictions which are quite unique. And as was mentioned by Alison, it's really important to understand when they are making progress and when they are correct and when they are wrong. And I think the real breakthrough of AlphaFold was not just that it was very accurate in predicting the structure of proteins. It was also that it was very accurate in saying, 'I did not make the right prediction,' when it did not make the right prediction. So the uncertainty, the predictive uncertainty of AlphaFold was extremely well calibrated, and that gives confidence to people who are using these sophisticated tools that they know when the tool can be trusted and when the tool cannot be trusted. Now, this is not the case for all the machine learning models out there. In the case of LLMs, LLMs hallucinate, and one of the grand challenges is that you don't know when they have hallucinated and when they have come up with something remarkably intuitive and profound. So one element of this is, as you are using these powerful tools, how do you build the scientific community and how scientists can understand how to leverage these tools to the most effective manner. Then there comes a point where you're using the AI systems in a more autonomous fashion, and this is what we are seeing in our fusion projects and our astronomy projects, where you have these controllers now which are running fusion reactors, and they are stabilizing plasma configurations in a way that has not been designed by hand. It has been learned. So people who are working in those systems, they are extremely sort of cognizant of the safety implications. They want to understand what is the out-of-distribution generalization behavior of these models. So validation, again, is extremely crucial. When you give autonomy to such a system, you need to understand what are the limitations, how do you sandbox it, how do you properly test it, how do you exhibit—like, how do you test all the different ways that systems can go wrong? So I think, both from a tooling perspective as well as from an autonomy perspective, there are interesting challenges that we are actively sort of seeing the whole community and us focusing on.
Sure. I mean, the idea that was emphasized by [INAUDIBLE] and you, as well, Alison, about this trust issue—oftentimes, even the experience with chatbots is you tend to accept. They assert, and you think this may be a truth, in fact, when it certainly isn't. And also, I just think one case in point in genomics, a space that I work in quite a bit, because as James mentioned, one of the greatest challenges in human genomics today is that variants of unknown significance, VUS, like UFOs, we don't know what they are. And in fact, the jump from AlphaMissense to get to such a high number of variants that we understand more—not that it's absolutely right, but the fact that it's an order of magnitude increase in essentially decoding or understanding what variants of potential pathogenic significance. Now, one of the things we want to get into is this data shortage. So in talking with John Jumper last night, I think he agreed that Helen Berman and the Protein Data Bank, if it wasn't for them, there wouldn't have been AlphaFold, and we have limited resources of data sets. You mentioned, Pushmeet, the GraphCast, which the weather—we got lots of data on weather. So the problem we have now is limited great data sets. So we have this path of discovery, but we're missing the inputs. So maybe you could help us. How are we going to get these inputs?
So I think this point is extremely valid, right? If we are thinking about developing intelligence, intelligence does not arise from a vacuum. It arises from some raw material, and that raw material is experience. Now, that experience could be in the form of data that has been collected by scientists in a lab, or that is experience that the model has achieved by interacting with a simulation. So we see both these elements working with high-fidelity simulations, which can give the model a lot of intuition about what is going to happen. Like, in the case of fusion and many of our physics models, we do use a lot of simulated data. And then, of course, there is experimental data, like the PDB and the work that we have done in AlphaMissense, which is also extremely sort of crucial. There is the third element, which is synthetic data, which is where you are trying to augment these data sources by making predictions and distillation as a concept, where you try to make predictions about unlabeled data and then use those labels to train a model in the hope that some of the right predictions will give you intuitions, and the wrong predictions will be thrown out. But that sometimes can be wishful thinking, and it's really important that, when you do these things, you have a very pure and firm test set which keeps you on the ground and keeps you grounded on the fact that you are actually making progress.
Yeah, so I'd like to mention two examples here, because I think you've hit upon a very important topic in the context of drug development. The first one I mentioned earlier was the predictive safety. So if we generate a compound or a biologic, can we predict whether it's going to be safe before we take it into humans? And of course, all the pharmaceutical companies have years and years of historical data that they've done, testing in rodents, for example. So what we've done, and this comes by way of histology slides, so glass slides in huge cabinets of drawers. It's like the Amazon warehouse, so very difficult data to search that. So, essentially, what you need is to digitize all of those glass slides and then load up that data. And by doing that, you can then develop predictive algorithms for certain types of toxicity, so certain organ tox, what would produce cardio tox or renal tox. Now, we've decided that type of data is not something we should just keep internally. I mean, this is best shared across the pharma industry to lift us all up. So we've partnered with a company called Deciphex. They're doing all of the scanning, but they're going to develop the algorithms. But they could also partner with other pharma companies, and then everyone will benefit from it. And I think it's one of the wonderful things about AlphaFold is everyone having access. So the more we can share data, the better. The other area, which I know you're excited about, is population data, population health data. So in the UK, recently, we have Our Future Health, which is an initiative to get NHS data. Again, one of the benefits of the UK is having the NHS with all of this actually consistently stored, and nomenclature is important. How do you actually describe diseases? What do you call them? So again, this is a public-private partnership with universities, the NHS, and pharma companies contributing. And I think that produces an incredible amount of data to mine using AI.
Yeah, I'm so glad you mentioned it, because, first of all, we rely on the UK Biobank with 15-year follow-up, 550,000 participants. No data set we have of people is compared to that, and now, I was astonished at how quickly Our Future Health—over two million people, participants of the five million in just a short time—
And it just makes all of us in the US a joke, because you're going so fast, and you have—unlike the NIH, they're unwilling to have public-private partnership. So you're doing that, and it makes a big difference. So we're relying, again, on the UK to have data sets assembled. Before going back over to Alison, I just want you to comment. As you mentioned in our discussions, people think of drug discovery AI acceleration just in the molecule side, but there are so many different ways. Could you comment, for example, on just in the design and enrollment of clinical trials and other nodes of—
Yeah, so we're using AI to help us run clinical trials. This is actually the slowest part of R&D. So designing the molecules, we are getting pretty good at. But once you have the candidate molecule, that can still take 10 years to actually get a marketed drug, and it's all the slowness of running the clinical trials and generating data. So one of the key things is how do you enroll patients, and using huge databases that have all of the demographic data of people around the world. Let's say, let's take the US. So where do people live who have certain types of diseases? What is their age? What drugs are they taking? What hospitals are they associated with? By having all of this data, you can write the protocol for a clinical trial, and you can enter different parameters. Like, I would like women age 30 to 50. How does that affect the enrollment rate? I don't want them to be taking this drug. So we talked about breast cancer, whether they would be taking estrogen pathway modifiers. How would that affect the enrollment rate? And by optimizing then all of the parameters, you can massively accelerate the recruitment of a trial. So something that might take you three years to just recruit the patients to run the trial, you could now recruit those patients in, say, six months, and this is a huge acceleration of getting drugs through to the patients and to market.
Yeah, no, it's phenomenal. Alison?
Just partly listening to what's also said, we've talked a lot about partnerships, and when we talk about the culture change that's going on, this is dramatic in academia, where, traditionally, people would have their own—they'd work on their own in their small lab. They would be highly competitive with everybody else in the world. Along comes a tool, the family of tools of AI, and you now want to be able to have the data to be able to answer questions. You've suddenly got to partner with everyone, and this is proving to be quite difficult in certain areas. And this notion that competition drives innovation—competition, you've got to balance that with how do you manage partnerships as well. And also, in academia, how do you manage rewards now? The classic reward system in academic science would be to publish papers, and now, you have many, many—20, 50 people can be on a paper, because they've all contributed to a particular idea. But how do you make it attractive for young scientists to enter a field, be noticed, and their careers to progress when you work in this interdisciplinary area that's team science? And I think this is something else. We need to have new ways to reward. It's not new to suggest it. We've had the idea interdisciplinary research is always a challenge, but I do think we need to think about the new ways that we can recognize people.
Absolutely. I also want to take the chance to zoom in a bit on the ultrasound work you're doing, because a lot of—it was brought up in the inspirational video that started this morning about reducing inequities and the idea that you could obtain images through things like a smartphone anywhere in the world, even by people who have no training in capturing images. Could you comment about that?
Yes, because I have some experience of doing that, of taking the very low-cost probes. They're sort of under $1,000, and the idea there is that you will be simplifying what a person does. So sweep across the pregnant mom's abdomen, for example, and then let the AI decide what's there and what might be the issue. And these sorts of—we've both developed some early technology to do this. We've also, as part of the research, gone out and asked in Kenya and also in India, would they find it acceptable to use tools of this kind? And the answer is they are like everyone else. They want to know that tools work, but otherwise, they would very much like to use them. But it's a very different way to do healthcare. You've got a different—you mentioned human in the loop. The phrase I use is worker in the loop, because whoever the person is holding the probe, you've got to empower them, and that means that you might have different capabilities, different explanations come for what's found. Are you, for example, using it for triage to refer in, or is the decision being used for actual decision making? So it's really an early emerging area, but it's a fascinating area as well. But the notion that everything needs to be safe, but understanding that worker is really key to see adoption as well.
