So, with all the attention on OpenAI's Millenium Prize solution and an Anthropic safety researcher quitting and claiming that these companies are gambling with our lives, no one is really paying attention to what Google DeepMind just released, even though it's kind of insane. It's called Alpha Genome Atlas, and it's basically a system that predicts the molecular effects of every possible single letter change in the human genome, all 9 billion of them. And researchers have already used it to help uncover a genetic explanation for an unresolved case of severe epilepsy, with lab experiments backing up the prediction. So, what exactly has Google done here? How much does this actually tell us about our DNA? And what are the implications of this going forward? Because there's a pretty big difference between being able to read the human genome and being able to actually understand it. Let's get into it. All right. So, to understand why this matters, we first need to understand what DNA even is. Google describes DNA as the language of life, and mastering it could transform our ability to understand biology and treat disease. We can already read someone's DNA and identify changes in the sequence, but the hard part is figuring out which changes matter and what they actually do. DNA is written using four chemical letters, A, C, G, and T. And a change to even just one of those letters can sometimes have a significant biological effect. Other changes might be harmless though, or contribute to ordinary differences between people, like eye color, for example. So, if you're investigating a genetic disease, finding a variation is only just the beginning. You need to also figure out whether it's relevant and how it could actually be causing a problem. Now, only around 2% of the human genome directly codes for proteins. These are the molecules that do a huge amount of the work inside our cells. And we already have a relatively good understanding of how changes in those protein-coding regions can affect them. But, the other 98% called non-coding DNA doesn't directly code for proteins. And within that enormous stretch of DNA are sequences that help control how the genes work. Think of it like switches and volume controls. Some help determine where a gene is active or how much of a particular protein a cell produces. That doesn't mean every letter in this remaining 98% has some essential function, but there's a lot of important biology happening outside the protein-coding regions that we still don't fully understand, including how those regions control the activity of the genes that actually make proteins. The other 2% that we do understand much better. This is where AlphaGenome comes in. AlphaGenome itself was actually introduced last year. It's a model that predicts how genetic variants affect biological processes, including gene expression and RNA splicing. Gene expression is how the information in a gene gets used, while splicing is part of how a cell processes RNA before making a protein. AlphaGenome can compare an original DNA sequence with a changed version and predict how those processes might be affected, which is already pretty crazy. But now, with AlphaFold Genome Atlas, Google has precomputed those predictions across the entire human genome and made them widely accessible to researchers. Kind of like a giant map of the human genome. And that's where the 9 billion figure comes from. At each position in the genome, the existing DNA letter could be replaced by any of the other three. An A could become a C, a G, or a T. Roughly 3 billion positions with three possible substitutions at each position, which gives you roughly 9 billion single-letter changes. That's the space Google has mapped predictions across. The resulting data set is about 1 petabyte, roughly a million gigabytes. But the size alone isn't what makes this useful. Scientists need a way to find the relevant information within it. So, Google also introduced something called the AlphaFold Genome Variant Impact Score, or simply AVI. This combines AlphaFold Genome's predictions with information from AlphaMissense, which assesses changes that alter proteins, along with other biological features. It brings that information together into one score that helps researchers rank genetic variants by their predicted impact. Essentially, it helps them decide which changes deserve a closer look. Researchers can then dig into what's driving that score. Maybe the model predicts a disruption to gene activity, or maybe it expects a problem with splicing, which is, again, part of how a cell processes RNA before using it to make a protein. That distinction matters because it gives researchers a more specific hypothesis to test. A high score by itself doesn't necessarily mean that a variant causes someone's disease. Here's a much better brief explanation of the AVI score from the Google research scientists who actually worked on it. Check this out. >> And through kind of modeling the language of life, we understand some of the genetic mutations, what they do, and whether they are harmful or not. The question now is how do you combine all these predictions into a single number. >> The AVI score, the variant impact score, is a single number that tells you how deleterious or impactful this variant is. The higher that the number, the more likely it is to [music] have an effect on the phenotype such as cause disease. And we have precomputed this for all 9 billion possible single letter changes [music] in the human genome. One of the core motivations for AVI was how do we go from 10,000 numbers that AlphaFold gives you to [music] a single number that you can then use for efficiently prioritizing. >> It scores every type of variant and tries to give you a AVI score for anything with a big impact. >> we can make the haystack, the more likely we are to find the needle. >> So, yeah. I think you get the idea. What's incredible is that Google has now made this accessible through a web interface for any scientist or researcher to use. And we actually already have an example of what researchers can potentially do with it. Scientists at the Broad Institute, working with the Gregor Consortium, used the AVI score to investigate unresolved rare disease cases. In one case involving severe epilepsy, it helped them prioritize a previously overlooked variant affecting a gene called DNM1. This gene is already strongly linked to epileptic encephalopathy. So, finding a potentially disruptive change there gave the researchers something worth investigating. AlphaFold genome predicted that the variants created an incorrect splice site, causing an abnormal extension of the resulting protein. Basically, in simple terms, the cell would process the RNA incorrectly because it was being spliced in the wrong area, leaving extra material in the instructions used to make that protein. The researchers then tested the predicted effect in the lab, and the experiments supported it, while also identifying nearby variants with similar effects. Together with the other evidence, this was enough to recommend classifying the variants as likely pathogenic, meaning likely to cause disease. So, yeah. The result here was genuine progress toward explaining the condition and understanding exactly what's causing it. Developing a treatment, though, would still be another challenge. But, obviously, for a rare disease case that had remained unresolved, finding a likely explanation is a huge step, and it shows how these predictions can guide an investigation toward something that holds up experimentally in the real world. Now, Google also showed how this could help with research across much larger populations because when you're looking for rare genetic variants associated with a particular trait, the huge number of harmless changes can create a lot of statistical background noise. The relevant signal can get buried. Gareth Hawkes at the University of Exeter applied Atlas to genomic data for more than 54,000 UK Biobank participants studying links between rare variants and the levels of proteins circulating in their blood. By grouping variants according to their predicted molecular effects, he uncovered 22% more non-coding genetic associations. This is from the 98th percent. Basically, they found more links between genetic changes and the levels of particular proteins in people's blood using data they already had and were sitting on. Hawks also used Atlas to investigate links with a body mass index. By focusing on the 1% of non-coding variants predicted to have the greatest impact, he identified 19 genetic regions that could guide further research into that trait. So, yeah. The reason this is so crazy is because we already have the data. Researchers already have enormous amounts of genetic data. Think about 23andMe. And so, tools like this could help them get more useful information out of it, basically overnight, and help them make better decisions about what to test next. Multiply that across different research groups around the world working on different conditions, and you can see why Google wants to make this resource widely accessible. Which brings us to the implications of all of this. Google says these tools could help researchers find new therapeutic targets, better understand genetic disorders, and decide which experiments to run next. So, that could mean uncovering the cause of another unexplained condition, like we saw with epilepsy, or identifying a protein involved in a disease that researchers could try targeting with a new drug. Google also describes Atlas as a starting point. They say as AlphaFold improves, these maps should become more comprehensive and precise. But, there's actually an even bigger idea here. In the video, they explain that ultimately, they want to bring all these different Alpha models together into one giant system. You'll have AlphaGenome helping predict what a DNA change does, AlphaMissence assessing changes that alter proteins, AlphaFold helping us understand protein structures, and perhaps AlphaGo keeping the researchers entertained while the AI does everything. But, in all seriousness, the idea is to bring together those capabilities and create a full picture of the human biology, and eventually to recreate that biology perfectly accurately in silicon, so that you can run infinite tests and edits on it. So, yeah, this is why I wanted to cover this. Google is quietly building toward a general-purpose biology system, and no one is even talking about it. I mean, this is genuinely levels ahead of anything we've seen come out of the other top AI labs. Bringing all these scientific models together and giving researchers access to them that could be world-changing, literally. What excites me the most though is that this goes beyond whatever discoveries Google's own team can make. They're building the underlying tools to help the scientific community make discoveries faster, not just keeping it for themselves. So, you have researchers around the world now working on different diseases that can potentially use these systems to find things they would have otherwise missed. So, yeah. Let me know what you guys think about this. Do you think Google is taking the right direction here? And do you think they have the right intentions at heart? And also, what disease or area of medicine are you most hoping AI helps us make progress in next? Let me know down below. But anyways, thanks for watching. Feel free to hit that like button if you enjoyed. And as always, I'll be catching you guys tomorrow for our weekly AI news live stream every Friday at 3:00 p.m. Eastern.