DeepMind's New AI Just Cracked The Code Of Life
In this in‑depth interview, DeepMind researcher Pushmeet Kohli explains AlphaGenome, a genome‑scale machine‑learning model that predicts the molecular effects of single‑nucleotide variants across the human and mouse genomes. He describes the model’s novel architecture—base‑pair resolution with a one‑million‑base‑pair context window—enabling unprecedented precision and the ability to capture long‑range 3D genomic interactions. AlphaGenome is trained via transfer learning on both human and mouse data to learn generalizable biological principles. The model’s predictions are compiled into the AlphaGenome Atlas, a petabyte‑scale pre‑computed dictionary of variant impacts, and the AVI (AlphaGenome Variant Impact) score that ranks variant significance. Kohli highlights the model’s dominance on 25 of 26 benchmark tasks, its open‑source release for non‑commercial academic use, and the rigorous validation process that distinguishes causal mechanistic insight from mere correlation. The conversation also covers future directions—context‑aware predictions, expanded training data, linking molecular phenotypes to organism‑level traits, and applications in synthetic biology, disease treatment, and even virtual cell simulation. Throughout, Kohli reflects on the transformative potential of decoding the genome, likening it to the AlphaFold breakthrough and envisioning a future where such foundational models unlock breakthroughs across medicine, climate, materials, and beyond.
Chapters
if you can completely sort of really understand the workings of a cell, that will have amazing consequences for the fight against diseases like cancer.
I feel like I've been dropped into a science fiction movie. Like, we could only dream of having technology like this when I was a child. Now, I also have a question, and I was thinking that maybe only you can answer this question.
What Alpha Genome Atlas does it basically precomputes all of this. Alpha Genome Atlas is 30 times the size of AlphaFold database.
Wow.
So the scale at which Alpha Genome can produce this, the detailed information about the effect of variation, is unprecedented.
Good to see you, Pushmeet. Look, this is really challenging for me because this is very much outside my field. field of expertise, if I have any at all. As always, I come here today as a student to learn from you. What is AlphaGenome?
AlphaGenome is a machine learning model which makes predictions about variation in the genome. So our genome, not just the human genome, but genome is basically the recipe of life. What it is, it's essentially a code. In the case of humans, it's a code which has 3 billion characters. And if one of those characters changes, what would be the effect on the actual cell properties? What will be the effect on the organism? These are the things that AlphaGenome is trying to sort of uncover.
And if you're the person on the street, what could AlphaGenome change eventually about medicine or their life?
Everyone sort of is affected by what happens at the genomics level. The genome is the recipe of life. It has both the ingredient section, which says which proteins are expressed, and the sort of mixing session, which sort of talks about gene regulation. When do these proteins get expressed? Where do they get expressed? And in what quantity? And all of that information is encoded in this recipe of life. Every single person on the planet has a recipe. I have a recipe. You have a recipe. The person on the street has a recipe. All of our recipes, while they are similar, There are some changes, but what do these changes actually mean? What are the implications of these changes? That's what we have been trying to sort of understand for many, many years, if not decades. With the Human Genome Project, for the first time, we were able to read the entire recipe book, at least for one human sort of genome. But now we are trying to figure out how do we decipher it, or understand what actually happens if you make a change in one particular character? What are the consequences of it? Will it cause a specific disease? Will it basically be connected to high susceptibility for some kind of problematic issue, like maybe having a high body mass index, or having more susceptibility to certain cancers. All of these things can be decoded in some sense by better understanding of the variations that happen in the genome.
That is incredible. And I really wanted to ask this because my problem, every now and then I try to read medical papers. It's not my field of expertise, but I'm a student. I just try to learn. And my problem is that whenever I read about, I don't know, any kind of medicine, my problem is that every person is different.
Yeah.
And I go to the barber and the barber says that he never saw 2 heads that are the same. And I read a paper on muscle hypertrophy, like bodybuilding paper, how to build muscle. And they have a bunch of people carefully selected, same age, Same height, same build.
Yeah.
They get the same diet. They do the same training. And one guy gets incredibly muscular. They do the measurements.
Yeah.
Some others get a bit of muscle. One poor guy didn't get any muscle at all. And there was one who I think lost muscle. Like, it's crazy how, how different we are. I'm thinking that, that is it possible that, that we could finally be one step closer to understanding how how everyone is a bit different and what that means.
Yeah, exactly. I think so, this understanding or deciphering this code of life, what are the consequences? Sometimes those consequences might not mean anything, right? Like sometimes the changes are like they are not that material. Sometimes they are very material sort of consequences, which are very well understood, right? For example, sickle cell anaemia. There is one particular sort of mutation which causes sort of sickle cell anemia, and we understand what happens there. But most of the things that sort of you care about these days are sort of polygenic in the sense that there are multiple sort of mutations that are acting on, that are causing that particular incident or that particular phenomena. And in some sense, what AlphaGenome tries to do is tries to figure out like what happens at, especially at a single sort of mutation variation level at the molecular sort of phenotypes. Now, how do those molecular phenotypes then affect the property of the whole organism? Or in the case of humans, what the human sort of feels like the susceptibility to specific disease, that's— something that will still need to be sort of connected and research will need to be sort of done on that.
My understanding is that before we dive deeper, AlphaGenome is free for all of us forever for non-commercial use.
Yes, exactly. So as we have sort of done for many of our scientific models, whether it's AlphaFold 1, AlphaFold 2, AlphaGenome genome, we have made it accessible to the entire scientific community for academic sort of use, for them to sort of use it for their own sort of research, fine-tune it, sort of, and the model weights are available.
Are these true? And please explain. The genome is the instruction manual for cells.
Absolutely. I think the genome, as I was saying, it's like a recipe book. It has both the coding part of the genome, which is like the ingredients sort of section of the recipe book, which says, what are the proteins that will need to be synthesized? So the human proteome has 20,000 proteins. So our genome sort of encodes these 20,000 sort of proteins, and there could be variations in those 20,000 proteins. And then there's the non-coding part, which is which is sort of, which looks at how are these proteins sort of mixed, how are these proteins sort of expressed, where are they expressed, in what quantity. And that part is basically the dark part of the genome, which is much less well understood because it sort of talks about the regulation of genes and protein expression. And so I think what Alpha Genome is trying to do is trying to sort of sort of decipher and inform this whole sort of space.
Now, I also have a question, and I was thinking that maybe only you can answer this question. So I'm really curious about this. This is trained on human and mouse genomes?
Yes.
And what I would love to know is that you train it on the human genome, it has some amount of knowledge and quality of results, then you train it on the mouse genome too. After that, does it get better at the human part?
Yeah. And so this, I think, Caroline, I mean, you have sort of seen this in many different sort of places. In machine learning, neural networks try to fit to the data. And if you just give it the human data, the neural network has this susceptibility that it can memorize or sort of think about the mechanisms that only work in the case of the human genome. But by training it, by forcing it, forcing the model to explain the same effects in the context of the mouse genome, we are asking the model to think broader and think about general sort of principles that can explain both these different types of variations. And that really helps in not only model generalization, but also making the model sort of learn about sort of what really matters and what's really happening at the physiological level, sort of in those molecular phenotypes.
I've seen it in another Google DeepMind paper that was the SEMA project, which played a bunch of video games. And I really found it stunning that you play a video game you have some score on it, and after playing a bunch of other video games, you get better at this game. I think it seems like maybe the same phenomenon, but that sounds like intelligence to me, right? You apply the knowledge from other things to this thing.
Yeah, and I think that's transfer learning in action, where you are able to extract concepts from things which are related, but are different problems, and it helps you become better, right? So if you sort of solve your maths problems and then solve your physics problems, but just getting better on maths helps you get better on physics.
My understanding is that the big advance here was that previous models were either too zoomed in and short-sighted, or they looked from far away but at a low resolution. Now, this one can do both.
Yes.
True or false?
This is true. So when we started on this journey of going after this extremely ambitious challenge of deciphering the genome, there were a lot of technical challenges. Our genome, as I just mentioned, is 3 billion sort of base pairs. ATCG, right, sort of a very long sort of sequence. And if you want to understand what is the effect of a particular sort of variation, you need to look at a broad sort of an expanded neighbourhood around that variant, because it can affect sort of things which are quite far off because of 3D interactions and so on. It's not that a particular sort of base pair only affects what happens in its neighborhood. Like, the neighborhood could be very large. So earlier sort of variations of the model, so we had a model called Informer, which preceded AlphaGenome, where both the context window in which we were looking at the neighborhood was sort of smaller, as well as the resolution at which we were trying to analyze the genome was much coarser. Right. We were not looking at what's happening at the base pair level. We were sort of looking at sort of a large number of sort of base pairs taken together. And we were not looking at 1 million base pair window. We were looking at a smaller sort of window. What AlphaGenome tries to do and is the first model to achieve that is to operate at the base pair resolution and have this very large window. And that allows it to not only make sort of predictions at a lot of precision, which require a lot of precision, but also sort of make predictions which involve reasoning about a much larger set of things that are happening in the neighborhood of that variant.
I saw that it is so dominant, it matched or beat the strongest comparison model in 25 out of 26 variant effect benchmarks. I think only at AlphaFold did I hear this kind of dominance. That is absolutely incredible. Now, what I'd like to ask you, when you saw that it first started working, what did it feel like?
Yeah, so I think one thing that we were— we have been working on this problem for a long time, as with many of our scientific challenges.
It is—
these are incredibly hard problems and require innovation in many different things like data and sort of coming up with a better architecture, coming up with sort of new ideas on training and so forth. When we started seeing these results, we knew that we were in a different phase. And part of it was expected because we had done all the hard work of enabling training to happen at that resolution, which had never been sort of done before. But to get it right and to actually show that the original hypothesis that we had gone in with, that if you increase the resolution, if you increase the context window, if you regularize the training properly, then neural networks will sort of show the right results. And so that was really rewarding to see for the whole team.
Was there a phase of suspicion? Like, it can't be right, you know, maybe we tested on the training set or something?
Yeah, so I think this is where we have our team sort of— we are very, very cautious scientists. So there is a lot of work that the team does on sort of on metrics and evaluation. And so when the results came in, there is a genuine sort of expectation that you're going to check it again to make sure that everything is working properly. And these are sort of valid results. But we had a lot of confidence in our metrics that we had set them correctly because it was not something that we have been doing for the last sort of 3 months. And it has been a project for the last almost 4 or 5 years or more.
What is Alpha Genome Atlas?
So Alpha Genome Atlas is essentially a sort of a dictionary of what happens when you have variation in the genome. So now I told you about Alpha Genome. What Alpha Genome does is it looks at your genome, and you can ask it the question, if a change this base pair or this character in your genome from A to a T or T to a C, what would be the consequence? And that's very helpful. Suppose you wanted to understand the effects of all variations. Now, all variations, there are a lot of variations because, like, there are 3 billion base pairs and all of them can be changed 3 sort of times, right? A could be changed into a T, or a C or a G, right? So there are 9 billion possible single nucleotide variations that you can study. And for every single one of them, there's a lot of data that you can sort of, that you can extract as to what will be its effect at the molecular level, at the effect on splicing, the effect on sort of gene expression, and so forth. What Alpha Genome Atlas does, it basically precomputes all of this. It basically looks at what happens for all these variations and makes this massive dataset, a petabyte-sized sort of dataset, and makes it available to the scientific community.
That's unbelievable. It feels like this AlphaFold moment where the inference of time of folding became so quick that at some point Demis just said that, you know, let's fold all the proteins like Atlas. sounds like that for the genome.
Exactly. So, I mean, for AlphaFold, we had— we made the predictions for structures of 250 million proteins, almost all the proteins that were known at that sort of time. But even the AlphaFold database, which has now been used by more than 4 million users across 190 countries, Alpha, in terms of its actual size, Alpha Genome Atlas is 30 times the size of AlphaFold database.
Wow.
So the scale at which Alpha Genome can produce this, the detailed information about the effect of variation is unprecedented.
All this from 1 megabase of DNA sequence?
Yeah. So with a 1 million sort of base pair window in context?
For mutations that affect gene activity, there is a fairly strong correlation between the predicted and the observed effect size. I saw in the paper that rho is about 0.5. But that's stunning. That's incredible. And my understanding is also that this doesn't just predict whether a mutation matters, but often gets the magnitude of the effect roughly correct too. Is that fair to say?
Yes, absolutely. And I think what we did with the specialty Alpha Genome Atlas is we came up with a new way to sort of help scientists sort of look at the data and explore this massive amount of data. So the Alpha Genome Variant Impact Score, or the AVI score is sort of carefully calibrated to say if the score is high, that means that there is— that particular variant is a variant that you should really think about and analyze sort of carefully. And if the AVI score is low, there's high confidence that it will not have a dramatic impact. on various different phenotypes.
So to me, this sounds like one of those root node problems where you solve this problem and then the applications can grow out of it in so many directions. I check AlphaFold every now and then, which was one of these root node problems. And in the first couple of years, I could actually look through the papers that used it and I tried to understand what they do. Now, there is so much that's built on top of AlphaFold that it's completely hopeless for any of us to understand all of it. Does it look like AlphaGenome is going to be the same?
Yeah, I think AlphaGenome, with AlphaGenome, we are trying to solve that root node problem of deciphering the genome, right? And as you sort of mentioned, it is really truly a root node problem in terms of its implications. So not only does it have implications for understanding sort of diseases better, it also has implications for how would you treat diseases? What are the sort of key things that you need to sort of understand about biology? And what are the places where you can sort of control expression, or what variations sort of control expression of certain sort of proteins which are relevant in that disease pathway? But it's not just applicable for understanding diseases and human health. It has implications that go beyond in broader sort of biology. So in fact, sort of now we are seeing a lot of interest in use of generative AI in synthetic biology. And again, when you are sort of synthesizing sort of new organisms or new sort of proteins, like what are the properties that we should be sort of thinking of, right? Like you would have seen DNA language models like EVO sort of come up and people using them to sample different genomes to solve different types of problems. And again, AlphaGenome gives you sort of a lens by which you can study the properties of those genomes. And so the implications of this model and its future sort of variations, which we are working on, I think will be sort of in the same way, very broad.
I feel like I've been dropped into a science fiction movie. Like, we could only dream of having technology like this when I was a child. I think that's unbelievable. How do we know that it learned mechanisms, not just correlations? Usually that's the problem with many of these papers, that the difficult part is actually proving the connection between the two. It's almost impossible in many cases.
Yeah. So I think that's a very good question. I think that has been one of the challenges when you are looking in computational biology, that you don't want just to learn all the statistical correlations. You want to have the model, you want the model to be able to sort of understand the actual sort of causal sort of links that happen in, that make the predictions possible. And if you think about it, like in AlphaGenome, how that is sort of encoded, one is that is in terms of the properties that it predicts. It predicts properties at the molecular sort of level. So you can really see when it makes a prediction about a particular variation, about, say, splice sites or splicing or gene expression. So those are very sort of— those are predictions which can be explained biophysically, right, rather than sort of correlationally that, oh, this variation happened, basically causes that particular sort of disease. But this alpha genome is not saying that. It's basically actually telling you what is happening at the cellular sort of level. And then in terms of validating, what we have seen is actually when you take these predictions and actually try to validate them in the lab, we see that those predictions make sense. And you can make novel sort of findings of variation and importance of certain variants.
Essentially, because the whole thing is based in physics, this is what essentially helps in terms of verification?
Yeah, I think the tasks are more physically required to reason about the physics. Now, how the model is doing it, that's something that we are still sort of looking at and interpreting. But the The tasks that it is being asked to solve are grounded in physics.
Now you have this incredibly sophisticated learning algorithm in AlphaGenome, and then in the paper on top of it, you have this tiny, super simple layer of LASSO regularization. And I found that really amusing. I imagine this almost like a Formula 1 engine. you invented, and it has a little bicycle bell on top.
Yeah, I think this sort of shows the power of the representation, right? What you want models like AlphaGenome to exhibit is sort of a fundamental understanding of what they are seeing or how they are perceiving the DNA sequence. So in some sense, most of the work should be done by the model in terms of the feature representation. And then the sort of the final sort of predictor on top should take a very simple—
Almost trivial.
Yeah, almost trivial form, because most of the work has happened at the representation sort of layer.
What's the next step?
So there are so many things that we are working on, which is like with AlphaGenome, the story does not end here. Like there is a lot that we need to do to improve AlphaGenome to make it even better in terms of its understanding of the genome. If you look at genomics and biology, biology sort of has a lot of different contexts. For example, different cells behave differently, right? So making sort of AlphaGenome more sort of aware of the context in which it needs to make these predictions is one thing that we are sort of trying to work on. Expanding the amount of data that AlphaGenome can be trained on so that it can sort of become much more accurate about different kinds of properties. Connecting these molecular phenotypes that AlphaGenome is able to sort of produce and showing their connections and impact on properties at the organism, sort of at the human sort of level, right? The propensity to disease and so on. How do we sort of facilitate that work? All these things are sort of directions that the team is now looking at.
There's so much going on.
Yeah.
I don't wanna get greedy, but I just wanted to hear about the next steps. Like, what are you building on top of this? That's awesome. You use Gemini with your son to explore science questions like how airplanes generate thrust. Has Gemini ever explained something to your son better than you could? So where you said, oh man, that's way better than what I would have said.
Yeah. So I think it's not only the ability of these models in explaining concepts, but by using different analogies and so on, but also the ability of these models to actually show the theory in action. So I was a few sort of weeks back, I was trying to sort of explain to my son like what happens in a wind tunnel when sort of engineers are designing sort of, say, airplane sort of parts. And what are the effects and how do they do these simulations? And we asked Gemini to explain it. And in fact, it created a simulation. And where sort of you could actually play around with different parameters and see how the airflow sort of changes. Now, it was not completely accurate, but it was a very good demonstration of what is possible.
Right.
creating these visual artifacts that help someone understand the concept.
We are living in an incredible time because I also remember when I first wrote my Navier-Stokes physics simulation for fluids. I remember when I first looked at the theory and I thought, I'm never going to understand this. This is never happening. And it took months and months of studying and trying. And now I understand it. very little about it, but at least a basic simulator I can write. And you just, here you just enter a prompt and you just got a simulation, by the way, on the side.
Yeah.
Because the intelligence is so incredible that it can do this sort of thing, like just make up a wind tunnel test. And you could probably running real time, you can see the streamlines as It's unbelievable. And it's only going to get better.
Yeah.
Now, I think we had about a 1 million x AI speedup in the last 10 years. If you factor in everything, it's not just hardware. It's if you factor in, I don't know, transformers, how much better the implementations got, the engineering, the everything. Let's say we got a 1 million x in the last 10 years. If in the next 10 years we have another 1 millionx, what could we do?
There are so many open scientific challenges that are beyond our current capabilities. Simulating even a cell, creating a virtual cell. Imagine the consequences of that, right? Like we would, if you can completely sort of really understand the workings of a cell, That will have amazing consequences for the fight against diseases like cancer or synthetic biology and so forth. Now imagine not— and that's, we are talking about a cell, right? Imagine sort of simulating organism. That will be sort of— that is the power that can come from that understanding and that capability. But there's so many things about the world that we still are scratching the surface on, sort of understanding long-term climate, for instance. What are going to be the consequences of certain sort of actions on long-term sort of climate, sort of coming up with new material types that can help us sort of store energy, sort of transform energy in different forms in a much more efficient manner. understanding biology from first principles to being able to truly design novel function for tackling some of the challenges that we are seeing, both in human health as well as sort of in agriculture and many other things that the world is facing. There are so many sort of problems that can be unlocked with that progress.
That was fantastic. Usually I ask this question people, and I see that it kind of breaks their brain too, because 1 millionx, like 10 years ago, like 2016, we had no idea what we could do today. And it's so exciting to think that, okay, if you add another 10 years, what would we be capable of? And everyone is thinking so small, but this one, this one was really good.
Nature is extremely sophisticated, right? Sort of, we are, and really, I think there's a lot to look forward to in terms of understanding the world. I mean, understanding— we are talking about understanding cells, understanding our own brain, like, is another sort of challenge. What is actually happening in sort of materials? Why are they exhibiting certain properties? Even sort of properties that we have seen Forget about new types of properties, but properties that we have already seen, like superconductivity. We don't have a complete theory about superconductivity, right? We have one sort of narrow theory of superconductivity. Imagine the new properties that we are yet to sort of explore or sort of really crisply understand why these phenomena happen and the ability to be able to design those objects.
Fantastic. Pushmeet, Every time we meet, I learn so much. And you know what the best part is? I will go home and I cut these videos together myself. So I'm going to watch it like 5 or 6 more times. And I guarantee you that every listen, there's going to be something new that I learned. So this was amazing. Thank you so much.
Thank you.
I use LaMDA to reproduce AI research papers, often in minutes. It's also great to train your own models or fine-tune an existing one. Run inference, or text to image or video? Easy peasy. Running a DeepSeek chatbot or agent? Super fast, super reliable. Lambda gives you powerful NVIDIA GPUs to run your own experiments. I test ideas from the papers I cover, and moments later, results. Love it! Seriously, try it out now at lambda.ai/papers.
