Google DeepMind

AlphaGenome 2: Teaching AI to read the entire human genome

Published 2026-11-11 · Duration 39:47

Summary

This discussion details AlphaGenome 2, an advanced AI model developed by Google DeepMind to decipher the human genome. While only 2% of the genome encodes proteins, the remaining 98% (non-coding DNA) controls gene expression. AlphaGenome 2 advances this understanding by predicting the effect of single mutations down to individual cell types (single-cell resolution), helping researchers pinpoint disease-causing mutations and advancing the field of precision medicine.

Download summary

Key takeaways

  1. AlphaGenome 2 Focuses on Single-Cell Resolution 35:10

    AlphaGenome 2 improves upon previous models by zooming down to individual cell types, allowing predictions of how single mutations drive disease at a much finer resolution than previous 'bulk' tissue analyses. This leverages data from efforts like the Human Cell Atlas.

  2. Predicting Variant Impact via Computational Modeling 20:50

    The model predicts the impact of a single variant (e.g., changing A to T) by training on public data modalities such as gene expression and transcription factor binding, mapping these to the reference genome. It assigns a score (0-80) to rank potential pathogenic variants.

  3. AlphaGenome Atlas for Large-Scale Screening 25:50

    The AlphaGenome Atlas pre-computes the effects of mutation for all 9 billion possible single-letter changes in the human genome, providing a resource for researchers to prioritize variants and narrow down potential causes of rare diseases.

More digests on Genomics

Technical details

  • Genomic Data and Function 480s

    The human genome is approximately 3 billion base pairs long. While only 2% encodes proteins, the remaining 98% (non-coding DNA) contains 'control elements' that regulate gene expression, determining how many proteins are produced in specialized cell types.

  • Model Evolution and Architecture 1020s

    The development progressed from initial transformer models (like Informer) to AlphaFold (focusing on protein structure) and then to AlphaGenome, which expanded the sequence length from 200,000 to 1 million bases and increased resolution from 128 base pairs to 1 base pair.

  • Computational Methodology 1350s

    The model is trained using public data modalities (e.g., gene expression, TF binding) relative to the reference genome. Predictions are made about variants even though the training data is fixed to one reference genome.

  • API and Accessibility 1700s

    The models are made accessible via a public API and a no-code website (AlphaGenome Atlas). Users can plug in their own data (e.g., from a specific lab study) to make variant effect predictions without needing to fine-tune the core model.

Mentioned resources

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.