A genetic variant can be as tiny as a single changed letter of DNA. Whether that change matters is much harder to figure out.
Google DeepMind wants to make that search easier with AlphaGenome. It has just released its AlphaGenome Atlas, which predicts some nine billion possible single-letter variants across the human genome, providing researchers with a way to look at how individual DNA changes may affect molecular processes.
The scale is impressive but the more interesting story is what’s going on around it. AI models are starting to be research infrastructure in biology, rather than tools scientists only run when they have a specific question.# A prequestion map
AlphaGenome is developed to predict the effects of DNA sequence on biological activity. It can look at long stretches of DNA, but still have enough resolution to look at variations on the level of individual bases.
This is important because many medically relevant variants are not located within protein-coding genes. They can be located in regulatory regions that influence the timing of gene activation, the strength of gene expression, or the processing of RNA.
Those effects take time to study experimentally. AlphaGenome can’t replace that work, but it can help narrow the search.
AlphaGenome Atlas takes it a step further. Instead of waiting for a scientist to propose one change after another, DeepMind has already made predictions across a huge number of potential DNA changes.
That changes the work flow.
A researcher studying a suspicious variant can begin with a prediction that has already been made and then decide if it warrants further investigation. AI is used earlier in the process, helping scientists decide where to spend their lab time.# There’s a foundation model race for biology
AlphaGenome is part of a much larger change.
Evo 2 is learning patterns across genomes of different life forms. Other research groups are working on foundation models for proteins, RNA, cells and gene regulation.
These systems are solving different problems but they are starting to form something that looks like a new computing layer for biology.
DNA is only part of the picture. A change in the genetic code can alter the regulation of genes, which can affect RNA, proteins, cell behaviour and ultimately the traits of a person or their risk of disease.
There is no foundation model today that can reliably follow that whole chain.
But the direction is coming clearer. Researchers are building models of each individual layer of biology, and are slowly working out how to link them together.#AlphaFold showed what happens after
That has already occurred.
AlphaFold is famous for predicting protein structures, but its impact grew with the release of those predictions at immense scale, via the AlphaFold Database.
Now, scientists didn’t need to run AlphaFold themselves every time they wanted to see a predicted structure. They could browse an existing resource and use that as a launchpad.
AlphaGenome Atlas does the same for genomic variation.
This could be a common pattern in scientific AI, train a large model, run it over a huge biological space, store the results and make those predictions searchable.
Then the model becomes part of the infrastructure that researchers use day to day.# Another problem: nine billion predictions
Predictions on this scale do not imply that biology suddenly has nine billion answers.
That makes nine billion things that might have to be interpreted.
Artificial intelligence is able to generate predictions orders of magnitude faster than laboratories can experimentally test variants. As these models get better, biology may be stuck with many more computational hypotheses than researchers have the time or resources to validate.
That makes prioritising ever more important.
A useful biological AI system will have to do more than tell us what might happen. Researchers will want to know how confident it is, what evidence underpins the prediction and which result is worth testing first.
That might be one of the next big competitions in AI for science.# Prediction is not yet evidence
There is also a very important limit to what AlphaGenome Atlas stands for.
Its billions of entries are predictions, not billions of lab experiments.
Biology is ultra context dependent. The effect of a variant can be dependent on cell type, genetic background, developmental stage and other factors. Just because a DNA change is predicted to affect molecular activity computationally does not mean that the variant causes a disease.
Experiments and patients still count as evidence.
That distinction becomes all the more important as those tools become more readily available. It can be a powerful starting point for an AI prediction that’s searchable, but it’s not to be confused with biological certainty.## The bigger race
The AI-for-biology race may end up looking very different than the race around large language models.
There is not one model that will dominate everything.
Instead, we may see specialised foundation models for the different layers of life: genomes, proteins, cells, tissues and eventually whole biological systems.
The real breakthrough would come when those layers start to work together.
A genomic model may detect a mutation that may be important. Another model might predict what it does to gene activity. A protein model can take into account downstream effects. A cell model can predict what is happening inside a particular cell type. Then experimental systems could test the most promising results.
Some of that workflow is already there.
AlphaGenome Atlas offers another vital piece of the puzzle by making a huge space of potential human genetic variants computationally searchable, even before scientists know precisely what variant they will need.
That’s why it matters that it made nine billion predictions.
The figure will be published. The more enduring change may be more subtle: AI is beginning to map out biological possibilities before science decides which ones to pursue.**