AlphaGenome, a 1 Mb DNA deep‑learning model, predicts base‑pair‑resolution genome tracks (RNA‑seq, splicing, chromatin) and scores variant effects, achieving state‑of‑the‑art performance across modalities.
0:00Welcome to Base by Base, the papercast that brings genomics to you wherever you are. Thanks for listening, and don't forget to follow and rate us in your podcast app. It's great to be here. So I want to start with a question that on the surface seems simple, but it's actually, well, it's incredibly complex.
0:18How well do we really understand our own genetic instruction manual? That's the big question, isn't it? And the answer is, it really depends on which page of the manual you're looking at. Right. Because we've had the sequence, the string of 3000000000 A's, C's, and T's for decades now.
0:36We could read the letters. We can read the letters, absolutely. But I think a lot of people don't realize that we're still just trying to figure out the grammar. That's a really good way of putting it.
0:44We're very good at the coding regions. you know, the genes that make proteins. That's maybe one to 2% of the... only 2%. Which leaves this other 98%, this huge vast space that for so long was just dismissed as... Junk DNA. Junk DNA.
1:01And we now know that's a massive misnomer. That's where the control panel is. That's exactly where it is. And here's the kicker. It's also where the vast majority of genetic variants linked to common diseases are hiding.
1:14And that's created a huge problem for scientists, this technological bind. It's a classic trade-off, a real dilemma. You could use tools that were like, a microscope. Super high resolution. incredible resolution.
1:27You could see a single base pair change, but you completely lost the big picture. You were just too zoomed in. Or you could go the other way. You could use a telescope. You could see those long range connections, how a switch way over here affects a gene way over there.
1:42The image. It was blurry, lower resolution. It's like trying to proofread a book by either putting your nose right up to a single letter, so you can't read the word. We're standing so far back, you can see the page layout, but the text is just a fuzzy line.
1:56That is the perfect analogy. And clinically, this is everything. A typo mutation, its impact depends on context. You need to know the word it's in, yes, but also whether it changes the meaning of a chapter 50 pages away, and until now, technology forced you to choose.
2:11You want to see the letter, clearly, or understand the chapter. You couldn't have both. You couldn't have both. Well, what if you didn't have to choose? What if a single AI could read a 1000000 letters of DNA at once, spot a single typo in that massive string, and then tell you exactly how it rewires the cell's 3D structure, changes how genes are spliced and flips them on or off?
2:33It sounds a bit like science fiction. But, I mean, that is the threshold we are crossing today. So today we are celebrating the work of the team at Google Deep Mind, led by Higa Absek, Natasha Ladasheva and their colleagues. Their paper is titled Advancing Regulatory Variant Effect Prediction with Alpha Genome.
2:49And this is incredibly fresh. It was just published in nature 2 days ago, January 28, 2026. This isn't just a small step forward. It feels like a genuine rethink of how we model the genome. So let's set the stage bit.
3:03We're in the world of what are called sequence to function models. Can you give us a quick primer on what they're trying to do? Sure. The concept is pretty straightforward. You feed a deep learning model, a raw string of DNA sequence, and you ask it to predict.
3:16Well, biological features. We call them genome tracks. So things like, Where do proteins bind? Is the DNA opener closed? Exactly how much RNA is being made from this region, things like that. But as you said, previous models just kept hitting a wall.
3:33The trade-off. The trade-off, you had the microscope models, things like splice AI or BP net. They were fantastic at spotting tiny patterns, these little motifs, but they were very nearsighted. They could only look at, say, 10,000 base pairs at a time.
3:45So they'd miss any of those distant control switches, the enhancers. completely missed them. Then you had the telescope models, like informer or Borzoi. These could see huge chunks of DNA, 200,000 base pairs or more, huge context, but the output was binned.
4:02Like a low resolution photo with big blocky pixels. That's it, exactly. You'd get a prediction for a block of, say, 128 base pairs, but you couldn't tell which specific letter inside that block was actually causing the effect.
4:14And for a doctor trying to diagnose a patient, that ambiguity is a huge problem. It's a nightmare. A single letter change can mess up splicing. It can change histone modifications. It can refold the genome.
4:28If your model is blurry, you might completely miss the Y. You miss the mechanism. So this is where alpha genome comes in. The model that claims to solve this tradeoff. So let's look under the hood. What are we talking about here?
4:41Okay, so the 1st thing that just jumps off the page is the input size. Alpha Genome takes in one megabase of DNA. 1000000 letters. One million. That's a massive jump from the 10 or 200,000 we were just talking about.
4:53It's a huge leap. And that number wasn't just picked out of a hat. A 1000000 base window covers something like 99% of all known, validated enhancer gene interactions. So you're basically capturing the entire relevant neighborhood for almost any gene.
5:07You're capturing the whole regulatory landscape. But here's the real magic. They didn't have to sacrifice resolution to do it. So big context. And microscopic precision. It processes that huge input and gives you predictions at single-based pair resolution.
5:23No more bins. It's crystal clear. And it does this for nearly 6000 different genome tracks in humans all at the same time. I want to geek out for a 2nd on the architecture because this is usually the hard part, right?
5:36Getting a neural network to see both the big picture and the tiny details. It's incredibly difficult. They used a hybrid approach. The backbone is a unit, which is famous in medical imaging for being great at preserving local fine details.
5:51That's your microscope. But they combine that with a transformer tower. And transformers are what's behind all the big language models everyone's talking about. Exactly the same technology. And transformers are built to understand long-range dependencies or attention.
6:04It's what allows it to understand how a base pair at the beginning of the sequence relates to one at the very end, that's your telescope. The paper also mentions something called sequence parallelism. That sounds important.
6:15Oh, it's critical. I mean, you have to understand, processing a 1000000 letters that this precision would just choke a normal computer. It's too much. So they split the sequence across 8 interconnected TPU devices.
6:27Google's custom AI chips. Right. Think of it like a factory assembly line. Each chip works on a piece of the DNA, but they're all talking to each other, sharing context. It's a mini supercomputer built just to read one stretch of DNA.
6:40Which sounds incredibly powerful, but also slow and expensive. Not something you could use in a hospital for every patient. You've hit on a really key point. And that's why they used a process called distillation.
6:52Okay, walk me through that. So they have this massive, powerful, but slow model. Let's call it the teacher. Then they train a much smaller, nimbler student model, but the student isn't learning from the raw data.
7:06It's learning to copy the teacher's answers. Ah, so you get the wisdom of the giant model, but in a much faster, more efficient package. That's it. The final student model. The one people can actually use can analyze a variant in less than a second.
7:21That's the difference between a cool research idea and a tool that could actually be used for diagnostics. Right, the tech is impressive. But does it work? Let's get to the biology. Does it actually outperform what came before?
7:34The results are pretty staggering. They tested it on 26 different standard benchmarks for variant effect predictions. Okay. Alpha Genome matched or beat the state of the art in 25 of those 26 tests. 25 out of 26.
7:49That's not just an improvement. That's a landslide. It's dominance, really. And you can see it when you look at really tricky biological processes. Let's take splicing, for example. Splicing is how the cell cuts and pastes the RNA message.
8:02It's notoriously hard to predict. It is. Most older tools could just guess where the cuts happen, the splice sites. Alpagino predicts the actual connections, the splice junctions, and how often each version is used.
8:13The paper had a really cool example with the DLG one gene. Yes, a fantastic real world validation. In tibial artery tissue, there's a tiny 4 base pair deletion in this gene, a tiny change. Alpha genome correctly predicted that this would cause Exxon skipping.
8:30Meaning the sales machinery just literally jumps over an entire section of the genes instructions. completely skips a chapter. And that totally changes the final protein. It's a very specific complex outcome from a tiny change and the model nailed it.
8:46It also did really well on gene expression, right? Predicting how much of a gene gets turned on. Yes, and specifically with EQTLs, the big challenge is getting the sign right, is the variant turning the gene up or down.
8:56A lot of models just say it changes, which isn't that helpful. And alpha genome can tell the difference. They showed it with a gene called APO 4 in colon tissue. It correctly predicted not just that a variant would change expression, but that it would lower it, and with much higher confidence than previous models like Borzoi or Informer.
9:13But the part that really felt like a leap into the future for me was its ability to predict 3D structure, how the DNA actually folds up. This is where it all comes together. The model predicts these contact maps, which show which distant parts of the DNA are physically touching inside the nucleus, and it actually did better than specialized tools that were built only to do that one job.
9:35Which brings us to what felt like the star of the show in this paper. The tell one onca gene. This whole section read like a great detective story. It is. It's a perfect case study. So towel one is a gene that can drive a very aggressive cancer, TCL, acute lymphoblastic leukemia, TL.
9:55And scientists already knew there were mutations in the non-coding regions that were involved. They knew there were mutations way upstream of the gene, but they couldn't piece together the mechanism. Why does a typo way over here cause a cancer gene to turn on way over there.
10:08So they fed the data into Alpha Gino. What did it spit out? It predicted an entire chain reaction. It saw the single letter change, and from that, it predicted, one, an increase in an active chemical mark called H3K 27 AC.
10:21Two, it predicted the DNA would become more accessible, basically unwinding to be read. And the final step. The final step, it predicted a huge spike in the Tao one genes expression. But it even went one level deeper.
10:34It figured out how. What was the mechanism? It saw that the mutation created something called a neo-enhancer. So the mutation literally built a new power switch from scratch. Exactly. And the model even identified the protein that flips the switch.
10:48It showed the mutation created a brand new binding site for a transcription factor called MYB. That is just an incredible level of detail. It didn't just say, this mutation is bad. It said this letter change creates a landing pad for the MYB protein, which opens up the DNA, which activates the region, which turns on the cancer gene.
11:08It connected all the dots from a single letter to a cancer driving mechanism through a 3D loop in space. That's the power of having both resolution and context. So let's call back. The implications here seem enormous.
11:20For our listeners, clinicians, researchers, what does a tool like this actually mean for their work? I think the most immediate impact will be on variants of uncertain significance, the V US? The dreaded VUS, the lab report that basically says, we found a change in your DNA, but we have no idea if it matters.
11:37It's so frustrating for patients, and often those variants are in these genomic deserts miles away from any gene. In the past, you'd probably just ignore it. Now, alpha genum can look at that desert and say, hold on, this variant is breaking a long range connection to a crucial gene.
11:54It eliminates the dark matter of the genome. It really does. And then for researchers, there's in silico mutagenesis. I am. Doing experiments on the computer instead of in the lab. Right. Instead of a slow, expensive CRISPR screen, you can just ask the model, what happens if I change this G to a T?
12:11And you can do that for every single base in a gene regulatory region. It's a way to map out the function of the genome at a scale that was impossible before. Which seems perfect for follow-up on things like Genome White Association Studies, or G-Was.
12:25It's the perfect tool for it. GWAS gives you a neighborhood where the culprit lives. Alpha genome can go door to door and point to the exact address, the specific variant that's causing the problem. Now we have to be thorough.
12:35No model is perfect. What are the limitations here? Of course. And the authors are very upfront about them. First, the training data is still mostly focused on protein coating genes. So it's predictions for, say, non-coding RNAs might not be as strong.
12:51And it's not predicting the disease itself, just the molecular consequence. That's a key distinction. It tells you, this gene's expression will go up. It doesn't tell you, this patient will get heart disease.
13:02You still need that biological and clinical expertise to connect the dots. And for now, it's only available for human and mouse. So, not quite ready for saving the rhinos just yet. Not yet. But the architecture is flexible, so expanding it is definitely possible.
13:17So if you had to boil this entire breakthrough down to one take-home message, what would it be? I'd say the take home is that the compromise is over. We no longer have to choose between resolution and context between the forest and the trees.
13:30We now have a unified framework that models the deep grammar of the genome splicing expression folding all at once. It really feels like we're finally starting to read the instruction manual the way it was written.
13:41I think we are. And if I can leave you with one final thought. Please. If a model can so accurately predict how a single letter change will rewire the cell. What does that mean for the future of designing DNA?
13:55Go on. We're moving beyond just reading the code. If we can predict the outcome of a change this well or on the cusp of being able to write corrections. Could we design a therapeutic gene switch that we know with high confidence will only turn on in liver cells?
14:10We could test it a 1000000 times in Silicon before ever trying it in a lab. From debugging the genome to actively programming it. That's a fascinating and powerful frontier. It is. That's the next horizon.
14:20This episode was based on an open access article under the CCBY 4.0 license. You can find a direct link to the paper and the license in our episode description. If you enjoyed this, follow or subscribe in your podcast app and leave a 5 star rating.
14:35If you'd like to support our work, use the donation link in the description. Now stay with us for an original track created especially for this episode and inspired by the article you've just heard about.
14:45Thanks for listening, and join us next time as we explore more science base by base.
