Google DeepMind Podcast

Avsec & Gayoso on the 98% of DNA we called junk

Žiga Avsec & Adam Gayoso· Researchers at Google DeepMind
·~40 min·English·Other
TrainingAI Company
TL;DR

DeepMind's AlphaGenome reads the 98% of DNA once called junk, and AlphaGenome 2 now resolves it cell type by cell type, to rank which single mutations drive disease, from a rare-epilepsy case to a gene regulated 290,000 bases away.

01The core problem

Three billion letters we can't yet read

We sequenced the human genome in 2003, but reading what a single letter-change actually does to the body is a far harder, still mostly unsolved problem.

But this distance between genotype and phenotype is very long and very complex. It's a very complex relationship.

— Žiga Avsec, Google DeepMind Podcast
Key Insight
AlphaGenome does not try to leap straight from DNA to disease. It deliberately works at the near end of that gap, the molecular layer, where the data is richest and the signal is cleanest, and leaves the long road to the clinic to others.
02What the genome actually is

The 98% we called junk

Only 2% of your DNA codes for proteins; the other 98%, long dismissed as junk, is the regulatory system that decides how many of each protein a cell makes, and when.

But then the rest of the genome, the 98% tells the cells how many of these building blocks should they produce, so that they will assemble into the right machines to carry out the function.

— Žiga Avsec, Google DeepMind Podcast
Key Insight
This reframes the whole target. If disease often hides in the regulatory 98% rather than the protein-coding 2%, then a model that can only read genes is blind to most of where variation acts, which is exactly the gap AlphaGenome was built to close.
03The modeling trick

Trained on one genome, reads them all

AlphaGenome never trains on your DNA; it learns from one fixed reference genome with public measurements aligned to it, yet can predict the effect of variants it has never seen.

we make predictions about different variants in the genome, despite the fact that we only train it with respect to one fixed genome.

— Adam Gayoso, Google DeepMind Podcast
Key Insight
Training against a single reference sidesteps the privacy and scale problems of learning from millions of personal genomes. The bet is that the rules linking sequence to molecular behavior are general enough that one well-annotated genome teaches them, and the variant predictions appear to vindicate it.
04Finding the needle

Ten thousand numbers into one score

For each variant AlphaGenome emits about 10,000 numbers; those compress into a single impact score from 0 to roughly 80 that lets geneticists rank mutations and ignore the ones scoring 0.

it gives you for a single mutation or a single variant, it gives you a score, a number between 0 and roughly 80.

— Žiga Avsec, Google DeepMind Podcast
Key Insight
A rich model output is useless to a geneticist drowning in candidate variants. The single score is the product decision that makes the model usable: it turns a wall of predictions into a sorted to-do list, and the precomputed Atlas means the answer for any of 9 billion possible changes is already waiting.
05What AlphaGenome 2 adds

Now it reads cell type by cell type

Earlier models predicted at the bulk-tissue level; AlphaGenome 2 drops to individual cell types, letting it flag a variant that steers a gene 290,000 bases away, only in gut immune cells, a link later confirmed in the lab.

And the model actually predicted that there was a particular variant that affected a gene that was 290,000 base pairs away, which in genomic terms, that's really far.

— Adam Gayoso, Google DeepMind Podcast
Key Insight
Resolution is the whole story of the 2. A variant's effect can surface inside one specific cell state while washing out in a bulk-tissue average, so predicting at the cell-type level is what let the model propose an unlikely 290,000-base link that the lab then confirmed.
06A rare-disease win

The case of the undiagnosed baby

A child diagnosed with rare epilepsy at six weeks went four to five years with no known genetic cause; AlphaGenome's score ranked the true culprit at the top, it matched the condition, and the lab confirmed it.

And then all the puzzle pieces kind of clicked and connected and they were able to solve that case.

— Žiga Avsec, Google DeepMind Podcast
Key Insight
The takeaway is a division of labor, not a replacement: the model's job is to lift the real culprit to the top of a long candidate list, while the geneticist and the lab still decide whether it truly explains the disease. A tool that only ranks is enough to break a search that had stalled for years.
07Honest limits

Trust the hits, not the silence

The researchers call AlphaGenome high precision but low recall: when it flags a strong effect it is usually right, but it misses a great deal, so a clean result is not an all-clear.

I think these tools are, I would say, high precision, low recall tools, meaning that whenever these models predict there's a strong effect, they're usually right. But there's a lot they miss.

— Adam Gayoso, Google DeepMind Podcast
Key Insight
That asymmetry decides how the tool should be used. A positive hit is worth chasing; a negative is close to meaningless, because the model's blind spots are large. Reading silence as safety would be the dangerous misuse.
08The far goal

AGI before a virtual cell

The genome is just one component; the long-term dream is a virtual cell that predicts how a cell responds to a drug or a gene edit, and the researchers expect AGI to arrive before that cell ever does.

I think we're going to have AGI before we have a virtual cell. That's my prediction.

— Adam Gayoso, Google DeepMind Podcast
Key Insight
The stakes are concrete: only about 10% of drugs survive clinical trials, mostly because the underlying biology is not understood. A cell you could simulate before dosing a patient is a bet on cutting that failure rate, and calling it harder than AGI is a candid admission of how far biology still has to go.