AI x BIO 2026 | Keynote lecture by Pushmeet Kohli, Google DeepMind, UK

Thanks everyone. It's great to be back in some sense. Although this is my first visit to the campus. I don't know if many of you might know, I lived for 10 years in Cambridge, right next door in Great Shelford. So I know the area really well and it's really great to see the plans for the whole sort of how the center is going to expand. It's very exciting and it's one of the best parts of the world to actually do science and research, and I'm really happy how things are shaping up. As was mentioned, I lead the science program at DeepMind.

So I don't know how much of DeepMind you know about, but I'll still mention an introduction. So, DeepMind started with this mission of building AI responsibly to benefit humanity, and it was founded in London in 2010 by Demis, Shane, and Mustafa. And over the last 16 years, it has gone through a different trajectory and has seen the development of AI from what was a very research-focused program to something that is actually changing the world in a very significant way. To describe DeepMind, I generally describe it in three phases. The first phase of DeepMind was proving that thesis that you can have machines which can learn from experience and exhibit intelligence that goes beyond human knowledge. And it might appear quite a non-controversial statement to make at this point in time, but it was not so easy to explain when the organization was started. And the first key milestone in that first phase was that legendary match in Seoul, in South Korea, where the DeepMind system AlphaGo played Lee Sedol, the world Go champion. And I don't know how many of you have seen the documentary or that game. Hands up. So quite a few, but not everyone. I think there was this moment in the match where AlphaGo played move 37, and everyone in the audience and the commentators felt, oh, what a silly mistake this agent is making. It clearly shows that AI systems are not able to exhibit the kind of intuition that is needed to play this game, and that's only human. And only as the game progressed, suddenly they realized that that move 37 was not a mistake. It was an amazingly insightful move that the agent had played, and it impacted how game theory was written because it was a move that was considered something that you would never do, right? And so that proof point of these systems being capable of discovering knowledge that changes and that expands humanity's knowledge and understanding about a particular concept, I think that was a key moment in that journey.

Now, having done that, the second part was now how do you leverage this? And so I've always been personally excited about understanding the world and having technology have impact in a beneficial way in the world. So that's when basically I joined DeepMind. In fact, Demis had asked me to join DeepMind much earlier, and I had said no. I said, 'Well, you are interested in... I really like the team and the culture and the philosophy, but I'm not into games. I play them, but I don't want to solve games. I'm more into practical things.' But after the Go match, the phase two of DeepMind was all about now that we have proven this thesis, we now need to use it for amazing impact, especially in science. So that's when I joined DeepMind and we started the science program, and then the key milestone of phase two was AlphaFold 2, which many of you might have used. And then now we are living through phase three of DeepMind, which is now that we have shown that you can not only have these special-case models that can exhibit superhuman performance in games and also in beneficial applications in understanding biology, can we make a general model that can exhibit insight and intuition and improve in many different tasks, right? Which is what we are now developing with Gemini.

So when we were thinking about going for AI for science, the key question was how do we select problems because science is so broad and we were a relatively small team of around six to seven researchers. And like science, you can do all of science, right? You can take problems in biology, in physics, in chemistry, in mathematics. And so we needed an algorithm to figure out what are the key things that we are going to target. And the algorithm that we came up on had these requirements. The first requirement was look at a problem and make sure that everyone in the community felt that this problem is going to be transformative, that if this problem is solved it will have a transformative impact. There should be almost a consensus in the community that this is a transformative problem. The second expectation is that there should also be some sort of a consensus that this problem is inevitable, the solution of it is inevitable in the sense that we don't want to work on something which is completely impossible. We don't want to work on time travel or something, right, and waste our time. The third condition was that everyone in the community, there should also be somewhat of a consensus that this problem which is going to be transformative, it's going to be inevitable, like whether it happens in 10 years or 20 years or 100 years, it will be solved one day, it is not going to happen in the next 5 to 10 years. Okay. So why that third condition? The third condition was we did not want to work on low-hanging fruit. If there was a problem that a professor in Cambridge or Oxford or MIT or Zurich or Barcelona was working on, we did not want to touch that problem. We wanted to work on problems that require the unique features that DeepMind has, which is multidisciplinary teams which are really going after a problem for a long, long period of time. Right? So our team size, we wanted to make sure that the problems are hard enough for us to fully leverage all the key things that we have in the team and approach them in a mission-focused way.

So what does mission-focused execution mean? So when someone joins the team, I sort of explained it to them that typically when you are in academia, you are a PI and you have students and postdocs. But when you join DeepMind, everyone is a PI because all the people in the team are selected to be specialists, are people who are really at the top of their field, and then you are essentially building these collaborative projects with a set of PIs. And one of the first things that you... that first challenge that you sort of see when you're working together as a team is figuring out what is the right metric that we should be solving because people have different views. And I mean, yes, every metric is sort of wrong in some way, but we want to make sure that people could agree that it is worth us to get collaborating together on this one thing which really makes a difference. The second important thing was we really wanted to focus on good engineering as well as good ideas. The idea sometimes basically in research what happens is you have to sort of figure out novelty, novelty in terms of the techniques, and novel ideas are of course interesting and useful and so on, but sometimes when you really want to solve the problem, it's not just about novel ideas, it's about bringing all the good ideas which might not be novel working well in a combined system. So we really wanted to make sure that we are focusing on end-to-end building the system rather than just coming up with novel ideas. The third element was understanding that if you're trying to build intelligence, if you're trying to build intuition, that does not happen in a vacuum. You don't basically sit in a room and without any knowledge or experience you get knowledge. Knowledge comes from experience, and that experience is in the form of what data that you have, whether it has been collected through experiments or whether you have collected that data through simulations. And so data is the key raw material for AI, and I think we really wanted to make a science of how do we leverage the right data, and this is now true even for LLMs where people have looked at scaling laws and that has really taken the field much further. And then finally, validation. You need mechanisms to be able to say how are you making progress, that once you have solved the problem, can you really quantify that the problem has been solved. And this is where things like the CASP assessment were super critical because they gave us a mechanism to do a blind evaluation of true progress in the field. And with this philosophy, over the last nine years, we have now taken on grand challenges across a number of different domains, from biology to energy and materials, maths and computer science, geospatial intelligence, and agents for science.

So I'll focus on biology, and I think the key thing about biology, biology was one of the first projects that we took on, and for us in some sense, both Demis and I feel that in some sense biology is like the perfect application for AI. It's so sophisticated, it's so complicated, and in some sense it tests the human mind's ability to make sense of it. So I think if humans needed help in understanding a problem, then biology is the key problem for us to understand. So over the last many years, we have been looking at various problems in biology. The first being understanding proteins, the building blocks of life, and we started with predicting 3D structure of proteins, which resulted in AlphaFold, predicting their function, how do you understand what these proteins are doing, the role that they play, and then can we also design new proteins to solve different functional elements, right? So to design new binders or design new functional enzymes and so on. The second key grand challenge was understanding the recipe of life, understanding the genome. Like as was mentioned, we have been able to sequence the human genome, but we understand this language of life. We can read that language of life, but do we really understand it? Right? And can we understand the semantics of the genome? So this is the other grand challenge that we started trying to predict the impact of missense and non-coding variants across the human genome and other genomes for that matter. And finally, moving further, really building a virtual cell. Can we really simulate biology so that we can understand what are the processes that are happening in the cell? Can we predict all the kinase-substrate interactions? Can we understand the grammar of cell state, like if you introduce this transcription factor, what happens to the cell state and so on? Can you predict everything that is happening in a cell?

So I'll start with AlphaFold 2 because this is the most famous element. And what does it do? It basically goes from a sequence describing a protein, and this block that you see, there was a lot of engineering that went into developing this architecture, and it finally resulted in the 3D structure. I think the important element that, I mean, it was at that point in time not at all clear how much progress we would be making on this problem. If you are looking at this problem in 2016, or end of 2016 or early 2017 when we were starting the project, it was not at all clear that you will hit the 90 GDT mark on CASP. Now in hindsight it looks like yes, you can make progress, but even after AlphaFold 1, which came out in 2018, the team had to completely abandon the architecture for AlphaFold 1 and build a completely new architecture to go after AlphaFold 2 because we saw how difficult it was to make progress on even like 5 GDT improvement was incredibly hard. And of course the impact of AlphaFold was very clear because we were doing this work, especially for AlphaFold 2, at the height of the pandemic, and when sequences from SARS-CoV-2 came out and we were seeing the world over people trying to understand all the proteins. I mean, the spike protein everyone knew about, but the accessory proteins, people did not know the structure, and the fact that we could find the structures of those accessory proteins like ORF8 and share them with the scientists really showed the potential impact of this technology.

And so over the last five or six years since AlphaFold 2 has come out, what has really surprised me is not just the breakthroughs that it has resulted in, but how much it has been used. If in 2017 when we were starting the project you would have told me that there are 3.3 million users of the AlphaFold database, I would not have believed you. I mean, sometimes people get down about the state of the world and all the wars that are happening and so on. But if you want one positive story about the world, the story is that there are 3.3 million people on this planet who are studying protein structures. I mean, that is an amazing thing, right? That 3.3 million people are using the AlphaFold database across the world. We don't see it that way, right? Like we normally don't encounter people talking about what is the structural conformation of this particular protein on a bus or a train, but there are 3.3 million people like us who are interested in science. So that is the most rewarding thing that I can think of. Of course, the 40,000 citations and the Nobel Prize and so on are rewarding, but I think that usage number is what really, as they say, floats my boat. And of course, one thing to remember is that Demis and John shared the Nobel, but the broader team that was behind it, right? It was a real collective team effort of a lot of people who came together to make this project work. And this would not have been possible without the data from the PDB, the assessments from the CASP competition, and without EMBL-EBI to advise us on how do we share this. So partnerships, data validation, and that whole engineering team spirit is extremely critical to making these grand challenges and moonshots a success.

So what were the lessons from AlphaFold? I think there were two key lessons. When you're thinking about these machine learning and AI models, you really have to look at not only the strengths of the model but also the weaknesses. Even if AlphaFold had 95 GDT but its uncertainty was completely broken, like you would get a result and you would have no idea whether that is a valid result or not, people would have thought twice before using it. The fact that AlphaFold, when it is making mistakes, can tell you that it's uncertain about certain predictions, I think was extremely crucial. So that was one lesson. The other lesson is that the true power of an AI and machine learning system is its ability to generalize. Otherwise, it's basically just memorizing stuff and that's not very impressive. Sometimes that can be useful, but I think the true power is in generalization. And one such example of generalization is that when we shared AlphaFold 2, people told us that when they looked at the predicted uncertainty of AlphaFold 2, it was the state-of-the-art predictor for disorder in protein structures. And this is something that we had never trained AlphaFold to do, and it just emerged from its training that it could generalize to this new problem without having been trained on it and being the state-of-the-art on that problem. So true intelligence is not in its ability to memorize, but its ability to generalize and explain concepts.

And so over the last many years, the story did not stop with AlphaFold 2. We carried on the work with AlphaFold 3, extended the structural characterization work beyond protein structures to more general biomolecules, RNA, DNA, how do complexes, the structures of complexes and so on, and then spun off a lot of this work with Isomorphic Labs, which is now taking this work to accelerate drug discovery across a number of disease areas, and they are growing as you might have heard in the news. So I think it's a great story about the whole UK and European biological progress. We have of course been working on further expanding the abilities, so working on de novo protein design, leveraging some of the lessons from AlphaFold 2 and in fact some of the models as well with AlphaFold 3. And this is again showing quite impressive results, not only for finding binders with much higher success rates, but also finding binders which have much higher binding affinity, right? And they can work with sometimes nanomolar or picomolar concentrations. So all of this work just shows that AI is basically helping us do a lot of the work in silico and move much faster, and then of course validation is needed to really find the final solutions.

Now coming to the genetics part and genomics part, we have been working on both AlphaMissense, which Jun actually led, basically taking the AlphaFold model and its inspiration and using it for predicting missense variants, and then following up with AlphaGenome, which does variant effect prediction, like splicing, gene expression, and so these are key developments on the genomic side of things. AlphaGenome, just to give you a key snapshot, if you have more questions, the team, Jun and Josh, a number of people are here from the team, so you can ask them key details of the question. But I think the key characteristics of the model is that it is a truly long-context model with one megabase pair context length, and it works at the base pair resolution. And I think these two key elements are unprecedented in terms of its ability to make predictions about these multimodal phenotypes like transcription and splicing and so on. And if you look at its performance across the best-known models across a large number of both sequence tasks as well as variant effect tasks, you see a general uplift across the board across RNA expression, splicing, DNA accessibility, histone modifications. You really see that it's not just a model that has been fine-tuned on one single task. It shows generalized improvement across a whole spectrum of different things, which is what you would expect from a model that is able to transfer and generalize across the biological modalities.

And not only is it something that is performing well, again, as we learned the lesson from AlphaFold 2, the key impact of the model is making it available, making it available more accessibly and more broadly. And so AlphaGenome, we released the API which can be used by researchers, and we also shared the model weights. And then it's really heartening to see researchers building on top of AlphaFold. So the fine-tuning work by the Cole lab and the Kundaje lab, that has been really good, and the fact that now there's a PyTorch port, it really shows how the whole community is building up, building on top of these models. And I would like to now take a moment to share how we are expanding this further. As I mentioned, data is the critical piece. So for this part, I'll probably invite Julia from the Sanger Institute to come and join me. Data is basically the key ingredient for training all of these different models from AlphaFold to AlphaGenome, and we are really happy that we are starting this consortium with a broad partnership with the Sanger Institute, with Google.org, to expand the amount of data that we can gather for genomics research and for really training models and showing the next level of capabilities of these models. So Julia.

Thanks, Julia. Thanks everyone. We are really excited about this partnership, and I think this new data repository will not only expand our research, but I think will expand research across the globe on this very important problem. Now, changing gears, I wanted to talk more broadly on where we are in the AI ecosystem and what does this mean for science. Now, when you think about science, and when we see these models like AlphaFold and AlphaGenome, these are models which are working on specialized problems, problems that the scientific community knows are important and will change how things are done. But if you look at science much more broadly and look at all the activities of science, science involves a lot of different things, right? It involves data extraction and data curation. You want to capture data around the world. There's a lot of the work that we do as researchers and scientists is basically doing literature surveys, understanding what is coming out, doing data analysis, coming up with new ideas, and then finally there is this idea of experimental intuition. This remarkable superhuman ability to predict what would be the result for an experiment without actually doing it. And if you think about the models that we have seen so far, like AlphaFold and AlphaGenome, they span on this spectrum on the experimental intuition part, right? They tell us basically what kind of results we might expect to see and so on. But on these other two parts, I think there is a revolution happening in general-purpose AI models, really expanding our ability to move fast on science. So I'll give you an example on data extraction. What we are seeing is that modern LLMs, so this is a result from our LLM Gemini, they are able to extract and curate structured data sets from thousands of papers, right? So what would have taken a team of researchers multiple years to curate data from a number of papers or from all the papers in a particular field can now be done in a matter of a few hours by the agent working autonomously, and that really expands what researchers can do. There is another revolution that is happening, and you might have heard in the news on how AI is affecting coding, right? So all these large valuation increases of AI startups and so on partly have been driven by the effectiveness of modern LLMs in assisting as coding agents. And what we are seeing is that these agentic coding platforms like Google Antigravity and many others, when you augment them with scientific tools and scientific skills, the ability to reach out and use the AlphaFold database or InterPro or protein sequence similarity search and so on, they really start becoming a programmable research workbench. And so what we did is basically we collected a bunch of these skills across genomics, proteomics, and the scientific literature and made these available to these coding agents. And now these coding agents can be used by researchers to really expand and accelerate their own research paradigms.

But these agents are not just speeding things up, they are actually discovering new things. So if you look at the scientific history of these specialized agents to do discovery, we have been working on this problem for many, many years. We started this work. The first result that we actually released was the AlphaTensor agent. And this agent was much inspired from the original successes like AlphaGo in being able to search over large search spaces. And what AlphaTensor found was it was able to discover new algorithms that had not been seen before. And over the last three years, what we have seen is that these agents have become extremely more powerful in their ability to search, but also much more general in the types of problems that they can solve. So they've become more effective in many different applications. Here's an example.

This is the problem of matrix multiplication, right? This is something that you would sort of have learned in school in linear algebra, and it's a very simple sort of problem. You want to multiply two matrices, like the A matrix and the B matrix, and get the C matrix. And in order to sort of do that, you need eight multiplications, right? These M1 to M8 are the eight multiplication operations that you need to do, which are then combined to get the final sort of result. And this problem is so fundamental. Almost every AI system is based on neural networks these days, and the basic operation of a neural network is matrix multiplication.

But it's not just neural networks. A lot of modern computing is based on matrix multiplication. Image processing, even sort of how you do encryption, everything can basically be seen as multiplying matrices and so on. Now you would say this fundamental problem, this very simple but fundamental problem in computer science, computer scientists would have sort of made sense of it, like how to do it best, right? Like this would be completely understood, and you would be mistaken, because this standard algorithm... there was a German mathematician, Strassen, in 1969, who was trying to prove that the fastest way to multiply matrices in terms of the number of multiplications was eight, right, for this 2x2 matrix. But he found that, in fact, surprisingly, it was seven.

And he came up with this sort of algorithm. And this really sort of confused the whole community, because then it brought up this question: well, what is the best way you can multiply matrices? And these are such basic, fundamental operations in computer science.

And this problem stayed open for more than 50 years until we asked one of our agents to find something which is better than this Strassen algorithm on the right, and it did find something. It found this algorithm. Now, I don't know about you, but I certainly would not have been able to come up with this algorithm. And this shows the power, the superhuman ability of these models to sort of reason and search in spaces that we wouldn't be able to sort of search. And that non-human ability is, I think, something that we need to leverage in problems which are really difficult and where we are seeing the limits of human reasoning. And now we have more general models which can just search in the space of programs. So you give them a task and they start writing code to be able to find a better solution for this task.

And this is not just scientific work. We are actually applying it in commercial applications, from data center optimization, hardware optimization like designing new chips, and even sort of training these machine learning models. And what you're seeing is that they can produce very significant sort of impact. So imagine all the trillions of dollars that are being spent on data center development, and this single agent was able to save 7% of the overall data center compute. Can you imagine the actual magnitude of that amount by just one agent running?

And now we are seeing this applied to science itself. We have these agents which, when given a problem, are able to do massive search in the space of literature and come up with interesting ideas. Once you're given a problem, it takes the research goal, generates ideas, reviews those ideas, refines those ideas, and finally gets the overall result. And the main thing with these agents is that as you give them more time, they are able to get better, they're able to define much more interesting ideas, and so forth.

And people have been using it across. So we have been sharing this with scientists, and they have been using it for a number of different applications, from drug repurposing for AML, understanding targets for liver fibrosis, understanding gene transfer mechanisms in bacteria for understanding AMR. And these are not just experiments. These are resulting in scientific publications which have now been peer-reviewed and submitted.

And so this is an example from one of the scientists, Dr. Joseph Andes from Imperial, on his reaction when he saw the result. So he and his team were working on this problem for the last 10 years. They submitted a paper to Cell. This was confidential. Nobody knew about it. And then we asked him to use Co-Scientist, and he gave the same problem to Co-Scientist, and Co-Scientist basically produced almost the same result in a matter of days that the team had found in 10 years. So this was the reaction.

So I mean, you can see the potential of these sort of tools for science, and these are not just prototypes and so on. We are actually making these tools available now across the board. So if you go to this link, we share three types of tools: Literature Insights, which can extract data from across the scientific literature and allows you to sort of do data curation; Hypothesis Generation, which is based on Co-Scientist, which Dr. Andes was talking about; and Computational Discovery, which is similar to the AlphaEvolve demo that I showed you about, which can discover new algorithms and in fact can even train new machine learning models for your data analysis tasks.

So this brings us to the natural question which I get asked: well, you are automating many different elements of science, so what is the future for science and what is the future for scientists? And I think I am really optimistic about the future of science and the future of scientists, because what these models and agents do is they solve the 'how' problem, in the sense that they all require the problem to be specified and then they can help us in solving the problem. But in all our experience, when we are sharing these mathematical agents with mathematicians or with biologists, the key challenge has been finding the right problem and articulating the right problem well. And I think that will be the role of scientists in the future, because that is something that is deeply human: understanding what is it that the scientific community wants to understand. What is the problem that will really impact society at large?

This is something that we think scientists will need to focus on. And of course the validation part as well, right? Really sort of capturing how do you assess that you have understood what you were looking for. These things will become modern-day tools for really accelerating the process of science. But the key scientific questions will need to come from the scientists.

And then the second question I wanted to leave you with is: how do you now facilitate beneficial applications? The ability to do science will be broadened by access to these tools. This is like a superpower. Many people will be able to do research with much more speed and much more effectiveness. And how do we make sure that all that increased effectiveness and speed goes towards the more positive uses rather than the bad uses that can also be enabled with that power? And that is, I think, something that we as a scientific community need to think very carefully about. How do we build the right mechanisms and the right protocols to make that happen?

Now other communities like nuclear and so on, they have been thinking about it in a different way. The biology community has taken a very informed view when it has looked at other powerful technologies like CRISPR and so on. But this technology, the key issue with this technology is how many people will have access to it, right? And so it will introduce its own sets of unique challenges, and we are working on some elements of that as well. But this is something that this community will really need to think about as we move towards really understanding the carbon substrate more broadly and thinking about how do we leverage the whole... I think we have been in the silicon age, but I think we are moving to the carbon age with this new expanded understanding of biology.

So my concluding thought is we are at the start of an incredible new era of scientific progress by leveraging these tools, and this era will bring amazing opportunities for us to solve many of the challenges that we are thinking about, like the key diseases that we want to tackle like cancer, the key global challenges like biosecurity, energy security, climate change. All of these challenges, we will have a tool that can help us make progress on these things, but at the same time there will be risks, and so we really need to think about what are the ways that we should bring these tools and this technology across the world. So I'll stop with that. Thank you.

Yeah. So I think this is something that people don't realize, right? That if you wanted to talk to a specialist in the world, you would have very few people that you will have access to, right? So people who know genomics and people who are interested in, I don't know, some manga, for instance — these are completely different sort of things, right? Manga, a Japanese sort of creative art, and genomics — you will have very few people at the intersection of those two interest groups.

Now with these agents you have something which is able to have amazing knowledge of manga, a very specialized understanding of coffee making, and a very specialized understanding of genomics and protein structure prediction.

So why am I bringing these topics up is because multidisciplinary research is going to become an incredibly easier thing to do. And so if I'm giving advice to people, it is: learn new concepts, learn new vocabulary, because you will be able to speak with these agents in different vocabulary in a way that you would never have been able to do. So these tools will give you the ability to explore concepts that typically you would never have had the chance to explore together.

And so understanding different fields, having an understanding of biology, going deep into various different disciplines at the same time and connecting them together — I think that will be the key opportunity of leveraging these tools effectively. So my advice to anyone who is now coming into the field, who is learning, is that go deep into problems, because you will have these tools that you can basically converse with and find useful, rather than you being the only person on the planet who is a specialist on manga and genomics.

Right. The second thing is basically to use these tools, because you want to understand their strengths but also their weaknesses. Only by truly playing around with these tools do you know that, because they are not perfect, they are imperfect. They make mistakes, they hallucinate, and the ability to distinguish what is a hallucination from a true insight — I think that is the skill that scientists will need to have, and that is the skill that people will need to have. Not just for science, for even coding and many other applications, you will need to have that understanding of distinguishing between hallucinations and true insights.

And what was the second question?

Yes. So we shared this with — of course we published this work. It appeared in Nature, and it really created a lot of interest in the mathematics community. By the way, the mathematics community — I can tell you some elements of them were very skeptical about AI when we started this work, like around six or seven years back. And I have to say they are the community which have also taken this up in the most expanded way now. So the use of AI across mathematics is becoming extremely ubiquitous, and they as a community have taken up these agents much more broadly and much more quickly than I would have imagined.

Can you repeat the question?

Yes. Yes. So of course, I think there is this element of training these models and curating the right datasets and making sure that the right sort of information is fed into the model, and I think we will need those people as well. But I think the majority of the work will be in basically coming up with the right specification, coming up with the right sort of problem to give.

Another sort of example: a few years back we developed this agent called AlphaProof, and what AlphaProof was able to do was for the first time come up with proofs of problems that appeared in the International Math Olympiad. Now these are very hard problems. They are for school-level students, but these are not typical school-level students. These are really, really mathematical prodigies that take part in the Math Olympiad. Like if you give me some of the harder problems, I'll not be able to do them.

And when we applied the agent on solving these problems, we had to translate the problem into this formal language called Lean. And we realized that most of the time the problem in achieving the right solution was in the misspecification of the problem rather than in the solution of the agent. Like mathematicians, when we were framing the problem in the more precise way, we were making mistakes, and so the agent would solve the problem, but it was not the problem that was asked. So I think coming up with good specifications of what is the right biological problem or the chemistry problem or the math problem — that will be the majority of the work that will be needed. And of course there will be work in gathering the right data and the validation piece on making sure that the final results are valid and we can extract things that are understood by human scientists.