On April 7, 2026, Demis Hassabis explained publicly why CRISPR — despite being able to target nearly any DNA sequence — still struggles to cure most genetic diseases: we don’t know which mutation is actually driving the problem, especially in the 98% of the genome that doesn’t code for proteins.
That’s the bottleneck. Not the scissors. The map.
Enter AlphaGenome, DeepMind’s new unified DNA sequence model. It takes up to 1 megabase of raw DNA and predicts thousands of functional genomic tracks at single-base-pair resolution — reliably flagging which non-coding variants are likely driving disease. CRISPR just got a targeting system worthy of it.
The Non-Coding Genome Was Always the Hard Part
Only ~2% of the human genome codes for proteins. The remaining 98% — once dismissed as “junk DNA” — contains the regulatory switches that control when, where, and how much genes are expressed. Mutations here don’t change a protein’s shape. They break the orchestra conductor.
Identifying causal non-coding variants has been notoriously difficult. Traditional methods were slow, low-throughput, and often inconclusive. Hassabis put it plainly: CRISPR can cut almost anywhere, but without knowing the exact disease-driving change in non-coding regions, you’re flying blind.
That gap is what AlphaGenome closes.
What AlphaGenome Actually Does — One Model, the Whole Regulatory Layer
Building on DeepMind’s earlier Enformer model — and complementing AlphaMissense for coding variants — AlphaGenome is a single unified framework that processes long genomic context and outputs functional predictions at a resolution no prior tool matched.
It ingests 1 Mb DNA sequences, long enough to capture distant regulatory interactions that shorter-context models miss entirely. From that input, it predicts thousands of functional genomic tracks: chromatin accessibility, histone marks, transcription factor binding, splicing patterns, and more — all at base-pair resolution.
The benchmark results are stark. AlphaGenome achieved state-of-the-art performance on 22 of 24 genome-track prediction tasks and 25 of 26 variant-effect prediction tasks. The model is already open for research use, with code and weights released via Google DeepMind’s GitHub.
Why This Is the Perfect Partner for CRISPR
Jennifer Doudna’s CRISPR revolution gave us molecular scissors. AlphaGenome gives us the targeting map for the dark genome.
The combination unlocks a workflow that wasn’t practically possible before:
- Identify the precise non-coding mutation causing a patient’s disease
- Design an edit that restores normal regulation — not just disrupts a protein
- Test thousands of potential fixes in silico before ever touching a cell
- Simulate cascade effects across long-range regulatory interactions
For complex, multi-genic diseases — where multiple regulatory variants interact — AlphaGenome’s 1 Mb context window becomes especially critical. Human analysis would miss the cascades. The model doesn’t.
This is exactly the kind of programmable biology at BioComputer’s core: raw genetic sequence turned into predictive, editable computation.
DeepMind Has Now Built the Full Instruction Manual for Life
AlphaFold cracked protein structure. AlphaMissense mapped coding variant effects. AlphaGenome completes the trilogy by reading the regulatory operating system that governs it all.
The progression isn’t coincidental — it’s a deliberate push from “parts list” to “operating system.” And the implications compound: rare genetic diseases, common complex conditions like cancer and autoimmune disorders, personalized medicine based on an individual’s full genomic regulatory landscape.
Researchers anywhere can now query variant effects across massive genomic contexts without building their own multi-million-dollar experimental pipelines. The bottleneck is no longer “can we edit the DNA?” — it’s “do we know which edit to make?”
AlphaGenome answers that question. The era of editing in the dark is over.
References
- Kimmonismus. (2026). Demis Hassabis on AlphaGenome & CRISPR. X. https://x.com/kimmonismus/status/2041537076617794040
- Avsec et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature. https://www.nature.com/articles/s41586-025-10014-0
- Google DeepMind. (2025). AlphaGenome: AI for better understanding the genome. DeepMind Blog. https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/
Related: The Programmable Genome · AI-Biology Convergence · AlphaMissense and the Coding Variant Map
Feature image: AI-generated using Grok.