Pharma
Deployment record
In the tracker →- Organisation
- Genomics Limited
- System
- AlphaGenome 2
- Function
- drug target evaluation
- 01DeepMind's AlphaGenome 2 helps researchers identify which DNA variants affect genes and in which cells, facilitating drug discovery.
- 02The model was tested on inflammatory bowel disease, predicting that one of seven DNA variants increases ETS2 gene activity in specific immune cells.
- 03AlphaGenome 2 improves precision and recall in predictions compared to its predecessor, scoring 0.62 against existing CRISPR interference screens.
Show all 5Show less
- 04PharmaGenomics Limited plans to integrate AlphaGenome 2 into its Mystra platform for evaluating drug targets, although no timeline for results has been provided.
- 05The long-term goal of AlphaGenome 2 is to simulate cellular responses to genetic edits or drugs, potentially accelerating experimental processes significantly.
Genomics Limited
Follow Deployed on Google
A genetics study narrows an inflammatory bowel disease signal to seven DNA variants. The gene that matters could be hundreds of thousands of DNA letters away. Before a drug discovery team can test it, the team has to work out which variant affects the gene and in which cells.
Recently, Google DeepMind released AlphaGenome 2 to help researchers make that call. The model is available through an application programming interface (API) for noncommercial research and in private preview for commercial research.
Alphabet and Google CEO Sundar Pichai said that the model could help researchers connect genetic findings to disease “in the specific cells where they actually unfold.”
From a DNA Signal to an Experiment
DeepMind tested the model on the inflammatory bowel disease case. AlphaGenome 2 singled out one of the seven variants and predicted it would raise the activity of ETS2, a gene about 290,000 DNA letters away, in monocytes and macrophages. That matched what earlier lab experiments had found.
The original AlphaGenome made predictions for a fixed set of tissues and cell lines. AlphaGenome 2 takes a DNA sequence and a cell type or state as inputs. A lab can use its own single-cell gene expression data to make predictions for newly profiled cells without retraining the model, DeepMind researchers Žiga Avsec and Adam Gayoso said in a conversation with Hannah Fry on the Google DeepMind podcast.
The week's drug discovery, trials and health-system deployments. Wednesdays.
DeepMind benchmarked the model against existing CRISPR interference screens, which suppress activity at selected DNA sites to see which genes respond. For DNA switches 100,000 to 2.5 million letters from a gene, AlphaGenome 2 scored 0.62 on a measure of precision and recall, up from 0.48 for its predecessor. The paper says that at a stringent confidence threshold, the model still misses most variants thought to be causal.
The Test Inside Pharma
Genomics Limited CEO Peter Donnelly said the company plans to incorporate AlphaGenome 2 into Mystra, its platform for using human genetics to evaluate drug targets. Donnelly said “it is difficult to move from an association signal to what a drug programme needs.”
Pharma companies already use Mystra through early access, according to Genomics. The AlphaGenome 2 launch gave no timeline for the integration or results from a drugmaker using the model to guide a new screen.
The economics come down to whether a team can confirm as many hits with fewer experiments or reach a decision sooner. A strong prediction could move an experiment up the queue. A low score alone gives a team little basis for dropping a target.
One Step Toward a Virtual Cell
On the podcast, the researchers described a longer-term goal of simulating how a cell responds to a genetic edit or a drug. AlphaGenome 2 predicts how DNA affects gene regulation in a given cell state. A virtual cell would also need to predict how that state changes after an intervention.
Google DeepMind CEO Demis Hassabis described the potential payoff in an interview with Lex Fridman last year. “Maybe you could 100x speed up experiments by doing most of it in silico, the search in silico, and then you do the validation step in the wet lab.”
Last updated
