TLDR
Google DeepMind researchers Giga and Adam discuss AlphaGenome and AlphaGenome 2, AI models that predict the molecular effects of DNA mutations across the entire genome, including the non-coding 98%. The models leverage public genomic data to score variant impacts at single-cell resolution, helping prioritize causal mutations in rare diseases and complex traits. The long-term vision is a 'virtual cell' that simulates cellular responses to genetic and chemical perturbations for precision medicine.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
The model can accurately predict effects of variants 290,000 base pairs away from their target gene if the variant sits in accessible chromatin.
Only 2% of the genome codes for proteins; the remaining 98% regulatory 'dark matter' is where AlphaGenome operates.
Researchers can plug their own single-cell data into AlphaGenome 2 without fine-tuning to get variant predictions for their specific system.
The AVI score compresses 10,000 molecular predictions plus conservation and AlphaMissense scores into a single prioritization number.
AlphaGenome was trained on a single reference genome yet generalizes to predict effects of unseen variants across the population.
Key takeaways
AlphaGenome reads DNA sequence and predicts molecular consequences of any single-letter change, focusing on regulatory non-coding regions.
AlphaGenome 2 adds cell-type-specific resolution, predicting variant effects in specific cell states (e.g., inflammatory macrophages).
The AlphaGenome Atlas pre-computes variant impact scores (AVI) for all 9 billion possible human genome mutations, freely accessible via a no-code website and API.
Models are high-precision but low-recall: when they flag a variant as impactful they are usually right, but they miss many true effects.
A real-world case: AlphaGenome's AVI score identified a causal splicing variant in a baby with rare epilepsy after years of failed diagnostics.
Training uses only the reference genome and public functional genomics data (e.g., ENCODE, Human Cell Atlas), not individual patient genomes.
Performance depends on chromatin accessibility: variants in open chromatin are predicted well even at long distances; variants in closed chromatin are poorly predicted.
Ultimate goal is a 'virtual cell' that simulates how a cell's state changes given its genome and perturbations (drugs, mutations), but this remains many years away.
Notable quotes
“we bought the book, but it's quite hard to read”
▸ Notable quotes · 0:24“Alpha Genome is designed to read a stretch of the DNA, or the genetic material, the instruction manual in the cell, and tell us what's happening inside the cell at that location in the genome”
▸ Notable quotes · 1:34“The distance between genotype and phenotype is very long and very complex”
▸ Notable quotes · 5:16“We're still very, very far from simulating biology”
▸ Notable quotes · 30:58“I think we're going to have AGI before we have a virtual cell”
▸ Notable quotes · 38:33“Biology is very humbling”
▸ Notable quotes · 37:50Chapters
0:00 — Introduction: The Genome as an Unread Book: Host Hannah Fry introduces the challenge of interpreting the human genome and the role of AI in making it legible, referencing the Human Genome Project and Eric Lander's quote.
1:34 — What Is AlphaGenome?: Giga explains AlphaGenome reads DNA sequence and predicts molecular effects of any single-letter change, serving as a tool to study variants at the molecular level.
2:05 — Personal Motivations: Giga (physicist turned ML genomics) and Adam (computer science, family cancer experience) share their paths to working on genome interpretation.
4:07 — Genotype, Phenotype, and the Non-Coding Genome: Discussion of genotype vs. cellular phenotypes, the 2% coding vs. 98% regulatory genome, and how cell-type-specific regulation works.
8:44 — Evolution of DeepMind's Genomic Models: History from Informer to AlphaMissense (coding variants) to AlphaGenome (non-coding, 1M base context) and the Atlas resource with AVI scores.
26:29 — AlphaGenome 2: Cell-Type Resolution: Adam describes AlphaGenome 2's ability to predict variant effects in specific cell types, leveraging Human Cell Atlas data, and a case study on macrophages in inflammatory bowel disease.
30:58 — Validation, Limitations, and Accessibility: Models are high-precision but low-recall; performance depends on chromatin accessibility; public API and no-code Atlas website enable broad use; researchers can plug in their own data.
36:53 — Future Vision: The Virtual Cell: Long-term goal is a computational virtual cell that simulates cellular responses to genetic and chemical perturbations to improve drug success rates; biology is harder than AGI.
Speakers
Hannah Fry — Host
Giga — Research Scientist, Google DeepMind
Adam — Research Scientist, Google DeepMind