Nine Billion Letters, One Petabyte: DeepMind Maps Every Possible Change to Human DNA, September 9, 2026

0:00Welcome to the DX Today podcast. Your daily deep dive into the AI ecosystem. I'm Chris and joining me is always is Laura. Chris, I have been waiting all morning to talk about this one. Because yesterday, Google deep mind quietly published a resource that I think is going to reshape how an entire scientific field does its daily work. You said quietly, and that is the part that caught my attention too. Because I saw the announcement scroll pass and I very nearly did not stop on it. Tell me what it is. On September 8th of 2026, deep mind released what they are calling the Alpha Genome Atlas. And the simplest way to describe it is a pre-computed map of what happens when you change any single letter of human DNA. Any single letter. Give me the actual number. Because when I skimmed the coverage, I saw a figure that I honestly assumed was a typo or a badly rounded press release exaggeration. Nine billion. They ran the model on every possible single letter substitution in the human genome. All nine billion of them. And then publish the entire set of predictions

1:01as a public reference that anyone can query. OK, slow down and build that up for me. Because I think most people hear the word genome and picture a long string of letters without any real sense of what changing one of them does. Your genome is roughly three billion letters long. And at every one of those positions, the letter sitting there could instead be one of the three other options, which is exactly how you arrive at nine billion possible substitutions. So to be precise about this, it is not a catalog of mutations that anyone has actually observed in a real patient. It is a catalog of every mutation that could theoretically exist in anybody. That is exactly right. And that distinction is the entire strategic point of the project. Instead of waiting to see a variant in a patient and then scrambling to interpret it, the interpretation is already sitting there, pre-computed and waiting for you. How large is that physically? Because nine billion of anything with a pile of predictions attached to each one sounds like it would blow through most institutional storage budgets before lunch on the first day.

2:02One petabyte to give you a sense of scale, that is roughly 30 times the size of the Alpha Fold protein structure database, which was already by a wide margin, the largest public resource this team had ever shipped. The Alpha Fold comparison is the one I want to pull on. Because that database genuinely did change how structural biology gets done day to day. Are we saying this is the same playbook run on human genetics? That is precisely the playbook. Alpha Fold predicted the structures of more than 200 million proteins, put them online for free. And within a couple of years, millions of researchers were treating it as basic infrastructure rather than as a novelty. And the defining property of infrastructures that you stop noticing it exists. Nobody writes a grand application asking for permission to look up a protein structure anymore. They just look it up and get on with the actual experiment. That is exactly the behavior change deep mind is trying to trigger here. And the genuinely open question is whether human gene regulation is a problem that yields to that same approach. Because it is a much messier target than protein folding.

3:05Let's slow down on the technology itself before we talk about the database built on top of it. What actually is Alpha Genome, the underlying model, and when did it show up? Alpha Genome is a deep learning model. Deep mind released in June of 2025 and its defining feature is that it reads about 1 million base pairs of surrounding DNA at once while still making predictions at single letter resolution. 1 million letters of context is an enormous window. Why does the model need to see that much surrounding sequence just to evaluate a change at one specific position in the genome? Because the letters that control a gene are very often nowhere near the gene itself. Regulatory switches called enhancers can sit hundreds of thousands of letters away and still determine how loudly that gene gets expressed in a given tissue. This is the part of genetics that I think the general public completely misses. We talk about genes as though they are the whole story when they are actually a startlingly small fraction of what is written down. About 2%, only 2% of the human genome

4:08actually codes for proteins. And the other 98% is regulatory machinery that we have historically been extremely bad at reading with any confidence at all. And that 98% is exactly where most of the disease associated signals from large population studies end up landing, which as I understand it is the single most frustrating fact in modern genomics. You have got it precisely. Genome-wide association studies keep pointing at regions out in the non-coding genome, and then everyone stares at the hit for years with no mechanism available to explain why it matters. So the promise being made here is mechanism. Not merely that this variant is statistically associated with the condition, but that this variant appears to break this specific regulatory process inside this specific cell type. Yes, and the atlas goes considerably further than one prediction per variant. Each variant carries thousands of predicted molecular effects, spread across hundreds of different human and mouse cell types and tissues, which is a staggering amount of detail. Thousands of numbers per variant is scientifically rich and practically unusable for a working scientist

5:10under time pressure. How is anybody supposed to act on a spreadsheet that is thousands of columns wide? That is exactly why they also built something called the Alpha Genome variant impact score, which distills all of those thousands of predictions down into one single number intended to help you rank which variants deserve your attention first. One number. I have a very strong instinct that one number is going to get badly misused within roughly a week of anyone outside the field, discovering that the website exists and is free. You are not remotely alone in that worry. One researcher quoted in the coverage said the score has a clear and legitimate use, but that it is also probably going to be very easily misinterpreted by people outside the field. That is a wonderfully polite way of saying somebody is going to pace their own variant into a public website, see an alarming number come back, and walk into a doctor's office, holding a print out. Which is exactly why DeepMind attached an unusually blunt disclaimer to this release. Their language is that Alpha Genome has not been validated for and is not approved for any clinical use,

6:12whatsoever, full stop. I want to sit on that sentence for a moment because it is carrying a tremendous amount of weight. This is a research instrument and not a diagnostic device, and those are genuinely different categories of thing. Completely different categories, and the scientists actually using it well, understand that perfectly. The Broad Institute team, for example, used it to help prioritize candidate variants in epilepsy cases, which is fundamentally a filtering and ranking job. Filtering is the honest frame for this. You have a patient with a long list of candidate variants and no principled way to rank them. And this gives you a defensible ordering for what to test in the lab first. And that ranking matters enormously because laboratory validation is slow and genuinely expensive. If a model moves the correct variant from position 400 on your list up to position three, you have just saved somebody months of bench work. Let's talk about how they physically built this thing, because pre-computing nine billion predictions with a model that has to read a million base pairs every single time sounds computationally impossible on its face.

7:14It very nearly was impossible. DeepMind has said they needed something like an 80-fold speed-up to make the project feasible at all. And they got there through three separate pieces of engineering rather than one clever trick. Give me all three, because I suspect the engineering story here is actually more interesting than the science headline that most outlets ended up running with when this landed yesterday. Model distillation first, which means training smaller and faster model to imitate the behavior of the big one. Then graphics processor kernel optimization and finally eliminating redundant calculations that were being repeated across millions of overlapping sequence windows. That last one is the classic scientific computing story, and I love it every time. Somebody finally profiles the workload, notices it is recomputing the identical thing millions of times over, and deletes an enormous amount of pure waste. And the real payoff from all that engineering is a shift in who is allowed to do this kind of work at all. Before the Atlas existed, running a model like this across the whole genome was simply

8:16unaffordable for most academic labs. So the actual product here is not really the predictions themselves. The actual product is that a graduate student with a laptop now has access to something that used to require winning a substantial compute grant. Two researchers in Barcelona made precisely that point in the coverage. Saying the Atlas removes the computational barriers that had previously made genome-wide application of these models completely unfeasible for most working scientists. How many people were actually using the underlying model before this landed? Because that number tells us whether the demand was real or whether this is a very expensive solution hunting for a problem. Roughly 9,000 researchers had accessed alpha genome through its programming interface since it came out last year, which for a highly specialized scientific tool is a genuinely serious and committed user base. That is real adoption by any standard. Now tell me about the licensing, because free for academics with a commercial tear attached is a very specific strategic posture coming from a very large company.

9:16The Atlas is free through a web portal for academic and non-commercial research. It is reachable through a programming interface and commercial access runs through Google Cloud rather than through that free public channel. So the goodwill flows to the research community while the revenue flows through the cloud business, which is honestly the same structure that made the protein structure database such a strategically clever piece of work. It is a remarkably effective pattern. You become the default reference layer for an entire scientific field. And then every commercial application built in that field has a very natural reason to sit on top of your infrastructure. Let's be fair to the criticism though, because I do not want the two of us, sounding like we are simply reading a press release aloud. What are the honest technical limits on this thing? There are three big ones worth knowing. The first is that most common diseases are polygenic, meaning dozens or hundreds of variants each contribute a small amount and a per variant score does not capture that combination at all. That is a serious limitation. If a condition is genuinely the sum of 200 small effects,

10:18then ranking any single one of those effects tells you almost nothing useful about the person sitting in front of you. The second limit is the viewing window itself. One million base pairs sounds enormous, but some regulatory elements are known to act from even further away. And anything outside that window is simply invisible to the model. And the third one I assume is the limitation that never goes away with any of these systems, which is that a prediction is not a measurement no matter how impressive the benchmark scores happen to look. That is exactly it. Alpha Genome 122 of 24 benchmark comparisons when it first launched, which is genuinely impressive. But a benchmark victory is simply not the same thing as a validated experimental result in a living cell. One of the scientists quoted in the nature coverage put it about as well as it can be put, calling this a useful and generous way to scale up access while saying plainly that it will not replace experiments. Another described it as a searchable dictionary for non-coding DNA. And I think that is this single best description

11:20anybody has offered of what this resource actually is and what it is genuinely good for. Addictionary is a very good metaphor because addictionary does not tell you what a sentence means. Tells you what the individual words could mean. And you still have to do the reading yourself. That is the perfect note to land on. This is a reference work rather than an oracle and the scientists who treated exactly that way are the ones who are going to get real results out of it. So what should our listeners actually be watching for over the next several months as this filters through the research community and out into whatever the commercial world decides to build with it? I would watch two things closely. First, whether independent labs publish validations that genuinely hold up. And second, whether that variant impact score starts quietly appearing inside clinical pipelines long before anybody has approved it for that purpose. That second one is the risk I would track most carefully. The gap between a research tool and a clinical decision aid is exactly where a great deal of harm has historically happened in medical technology.

12:21And the genuinely good news is that deep mind said the quiet part loudly this time. Attaching a disclaimer that leaves very little room for anybody to later claim they were somehow misled about what this thing can do. That's all for today's episode of the DX Today Podcast. Thanks for listening and we'll see you next time.