AlphaGenome 2: A Step Change in Predicting What DNA Does, and What It Might Add to Drug Discovery - by Professor Sir Peter Donnelly, CEO, Genomics

Today Google DeepMind released AlphaGenome 2, which represents an impressive step change in sequence-to-function models: AI models that read a stretch of DNA and predict what it does inside cells, such as which genes are switched on, how strongly, and in which cell types.

For those of us using human genetics to find drug targets, the key question at any disease-associated locus is what the variants there actually do: which gene they act on, in which direction, and in which cell types. AlphaGenome 2 will certainly help to untangle what is going on at some loci. Its predictions of whether a variant turns a gene up or down, and in which cell types, are markedly better than those of its predecessor. Even with these improvements, AlphaGenome 2 makes confident predictions for only a small subset of the variants that can be queried.

What remains an open question is how much it adds to the other evidence we already use to identify the gene behind a GWAS signal. That evidence includes proximity, colocalisation with expression and protein QTLs, and rare-variant information. Nobody yet knows the answer. At Genomics, we are about to undertake a detailed evaluation across our GWAS data, and we look forward to understanding this much better.

We also plan to make AlphaGenome 2 available through our Mystra platform, so that our partners can use it effortlessly alongside the broad suite of analysis tools on our association, transcriptomic, and proteomic data.

Congratulations to Žiga Avsec, Adam Gayoso, and the AlphaGenome team on a genuinely impressive piece of work. Human genetics has already shown that it can make drug discovery more successful. Tools like AlphaGenome 2 will help us get more value from genetic association data.

Drug development is expensive mainly because of failure. More than 95% of drug development programmes fail, including about 90% of those that reach clinical trials (Hingorani et al., 2019). The most common reason for failure is that the drug does not work in patients – too often the drug target turns out not to matter for the disease in humans.

Human genetics addresses this directly, because it lets us learn about human biology in humans. When a DNA variant that alters a gene's activity also alters disease risk, we are observing a natural experiment that has run over a lifetime, in the species we actually want to treat.

The evidence that this matters is now strong. Multiple studies have shown that drug targets with human genetic support are two to three times more likely to succeed in clinical development. The most recent large analysis put the figure at 2.6 times (Nelson et al., 2015; Minikel et al., 2024).

The starting point is the association data itself. Genome-wide association studies (GWAS) have identified many thousands of places in the genome where variation is linked to disease risk, biomarker levels, and other traits.

At Genomics we have put a great deal of effort into this first ingredient. We bring together tens of thousands of GWAS, then undertake the quality control and alignment of each so that results can be compared consistently, and integrate them into a single resource through our Mystra platform. Working at that scale shows which signals are robust, and which are shared across diseases and traits.

An association signal is not yet a drug target. To act on one, a pharma or biotech team needs three things:

• the gene involved;
• the direction of effect, so they know whether to antagonise or agonise;
• the relevant cell types of action, where that is possible.

Getting from the signal to these answers, often called the signal-to-gene or locus-to-gene problem, is hard. Most signals lie in non-coding DNA, in the regulatory "switches" that control when and where genes are active. Each signal usually spans many correlated variants because of linkage disequilibrium. And a switch can act on a gene hundreds of kilobases away.

Current approaches combine several kinds of evidence. Proximity is the best single predictor: betting on the nearest gene does surprisingly well. Coding variants, colocalisation with expression and protein QTLs, rare-variant burden results, and gene-level models that use co-expression networks then shift those odds.

DNA sequence models offer an important, orthogonal source of information alongside these.

DNA sequence models learn the relationship between DNA sequence and what happens in cells. They come in two broad families:

• Sequence-to-function models. These are trained to predict experimental measurements directly from sequence: gene expression, chromatin accessibility, histone marks, transcription factor binding, and more. AlphaGenome 2 is one of these.
• Genomic language models. These learn patterns from DNA sequence alone, without measurements.

AlphaGenome 2 reads a million letters of DNA. From that sequence it predicts thousands of molecular readouts, at up to single-base resolution.

Because these models work from sequence, they can score any variant, including ones that have never been measured in an experiment. For a specific variant, the recipe is simple:

1. Predict with the reference allele.
2. Predict again with the alternative allele.
3. Compare the two predictions.

The difference contains the model's estimate of what the variant does to each nearby gene, in a particular cell type. AlphaGenome 2 only makes a confident call in a small subset of these cases, but where it does, this can potentially give a candidate gene, a direction, and a cell type: exactly the three things a drug discovery team needs.

AlphaGenome 1 predicted a fixed catalogue of experimental readouts, each tied to a tissue or cell line that had been measured in the laboratory. These include which stretches of DNA are open and accessible in that tissue, where particular proteins bind, and whether each gene is switched on and how strongly. The volume of predictions is larger than most people would guess: because these readouts are predicted along a million letters of DNA, at up to single-letter resolution, a single stretch of DNA in a single cell type could yield millions of predicted values.

AlphaGenome 2 also makes many predictions, but unlike its predecessor, it makes each prediction for a specified cell type or state, and that specification shapes the whole network, not just its outputs. For AlphaGenome 2, the cell type or state of interest is an input. Two further changes go with this:

• It is trained on large single-cell atlases as well as bulk data.
• A particularly striking new feature is that it can make predictions for cell types it has not seen in its training data, provided it is given single-cell expression data for them. No retraining is needed.

The paper tests the model extensively, and the improvements over AlphaGenome 1 are substantial. Three stood out for me:

• Cell-type resolution. Within tissues, the share of differences in gene expression between cell types that the model explains rises from about 23% to 34%.
• Direction of effect. At 90% precision and accuracy, it calls the correct direction for roughly twice as many fine-mapped single-cell eQTLs as before (43% against 23%).
• Long-range regulation. Linking enhancers to genes more than 100 kb away improves markedly, with AUPRC rising from 0.48 to 0.62 on CRISPR-validated pairs.

The comparisons are with AlphaGenome 1 fine-tuned on the same single-cell data, where relevant. Taken together, the model sees cell types much more sharply, and is much better at predicting whether a variant turns a gene up or down, which is what matters most for drug discovery.

The paper's testing of AlphaGenome 2 on the signal-to-gene problem itself is relatively limited. That is partly because publicly available truth data are limited: there are few disease loci where we know the causal gene, the direction of effect, and the cell type with certainty. The paper includes encouraging examples, such as the ETS2 locus in inflammatory bowel disease, and good agreement with the direction of rare-variant burden effects. But a systematic assessment across many diseases is still to come.

That is why we are excited, at Genomics, to be undertaking a programme of work to evaluate AlphaGenome 2 for signal-to-gene across our GWAS data. We will look at three questions:

• how often it adds information beyond existing evidence such as proximity, colocalisation, and burden results;
• how reliable its direction-of-effect calls are;
• when its predictions can be trusted.

One feature is worth understanding from the outset: AlphaGenome 2 is selective. It only makes confident calls at a subset of variants, so for many GWAS signals it won’t make a call. At the thresholds used in the paper, AlphaGenome 2 makes confident calls for about 2% of the roughly 6.8 million variants scored. That is less discouraging than it sounds: most of those variants were included simply because they are common, not because there is any reason to think they do anything. Among variants that fine-mapping identifies as likely to be causal, the call rate in the relevant cell type is around 15 times higher.

But when it does make a call, that call is strongly enriched for real effects, and so it can be highly informative. Its silence should be read as "no information", not as "no effect". Used that way, a minority of confident, reliable calls could add real value to the evidence for a target.

Our Mystra platform gives pharma and biotech teams access to our data, together with a suite of sophisticated tools for analysing it, including an agentic layer, MystraAI. Mystra is in many ways the ideal platform to integrate AlphaGenome 2, combining its insights across cell types with a wealth of relevant molecular information, including transcriptomic and proteomic annotations. We see that integration as a way to maximize the value of AlphaGenome 2 for our Mystra customers and we look forward to delivering its predictions inside our platform.