AlphaGenome brings DeepMind’s AI to non-coding DNA

On June 25, 2025, Google DeepMind introduced AlphaGenome to predict variant effects on regulatory signals. The original source supports the documentary core of the event; a molecular prediction is neither a clinical diagnosis nor proof of causation for an individual.

The announcement matters because most disease-associated variants are not found inside genes that produce proteins. Understanding what they do requires determining whether a mutation changes when, where or how strongly a gene is activated. AlphaGenome does not diagnose diseases or replace a laboratory experiment, but it aims to narrow down which variants should be researched first.

The problem lies in the 98% of the genome

Only about 2% of the human genome codes for proteins. The remaining 98% contains regulatory instructions: sequences that control gene activity depending on the cell type, tissue or stage of development. For decades, some of this territory was imprecisely referred to as “junk DNA.” It is now known to contain many important signals, although deciphering them remains a complex task. Document supporting the figure.

A single variant may leave a protein unchanged and still have significant effects. It can alter RNA production — the molecule between DNA and protein — modify how that RNA is spliced or prevent a regulatory protein from binding to DNA. These consequences often depend on sequences located far from the affected gene, which is why it matters that the model retain context across a broad region.

AlphaGenome combines convolutional layers, which detect local patterns in the sequence, with transformers, an architecture capable of connecting distant positions. Based on that analysis, it predicts thousands of molecular properties: where genes start and end, how much RNA is produced, RNA splicing, DNA accessibility, contacts between regions of the genome and the binding of regulatory proteins.

One model versus specialized tools

DeepMind trained AlphaGenome on public data from consortia including ENCODE, GTEx, 4D Nucleome and FANTOM5, which experimentally measured gene regulation across hundreds of human and mouse cell types and tissues.

The company says its model outperformed the best external system in 22 of 24 prediction evaluations on DNA sequences. In tests designed to estimate the regulatory effect of variants, it matched or exceeded the best models in 24 of 26 evaluations. The comparison includes tools specialized in specific tasks; AlphaGenome’s bet is to bring those predictions together in a single system. Document supporting the figure.

That unification may matter more than a one-off improvement in a results table. A lab studying a suspicious variant typically combines different programs to estimate gene expression, DNA accessibility or changes in splicing. DeepMind proposes obtaining those signals with a single API call by comparing predictions for the original sequence with those for the mutated sequence. According to the company, the calculation can be completed in one second.

The model follows in the footsteps of Enformer, another DeepMind system for predicting gene activity from sequence. It also complements AlphaMissense, which classifies variants in protein-coding regions. AlphaGenome broadens the focus to non-coding regions, where there are far more variants and fewer known rules for interpreting them.

From prioritizing mutations to designing DNA

DeepMind illustrates the system’s usefulness with mutations linked to T-cell acute lymphoblastic leukemia. AlphaGenome predicted that certain non-coding variants could activate the TAL1 gene by creating a binding motif for the MYB protein, reproducing a previously known disease mechanism. The value of the example is not that it discovered the case from scratch, but that it shows the model can connect a specific DNA change with a plausible gene and regulatory mechanism.

Potential applications include researching rare diseases, where an uncommon variant can have a major effect, and searching for therapeutic targets. It could also be useful in synthetic biology for designing sequences that activate a gene in a specific cell type, such as neurons, but not in other tissues.

The distinction between prediction and demonstration is worth preserving. The model learns from available measurements and offers hypotheses about what may happen in a cell; experimental validation will still be necessary, especially when a conclusion is intended to guide clinical research. In addition, more distant regulatory interactions and individual-specific effects remain difficult problems for models based solely on sequence.

AlphaGenome is available in preview through an API for non-commercial research. DeepMind plans to release the model later. If its results hold up outside benchmarks, it could reduce one of the most costly obstacles in modern genomics: deciding which of the millions of possible differences in DNA are worth taking to the lab.

Turning the headline into a check

An experimental result begins with its observable variable. Record what counts as success, which behavior triggers a label and which cases fall outside scope. a molecular prediction is neither a clinical diagnosis nor proof of causation for an individual. If the phenomenon cannot be recognized without interpreting a model's intent, the conclusion needs even greater caution and a reproducible definition.

Protocol matters as much as score. Document instructions, tools, time, compute budget, number of attempts, example selection and grading rule. Changing any one may alter the result without the model learning anything new. Comparing two headlines therefore begins by checking that they measure the same axis.

A strong replication tries to break the conclusion. Add unseen data, small variants, negative controls and tasks where abstention is correct. Preserve failures as well as selected successes. To assess how to separate input scope, predicted output, validation and clinical use, the set must resemble the intended use and reflect the cost of each error class.

What the record must preserve

A study can reveal a pattern without settling an entire field. Honest wording preserves domain, sample and date, and avoids turning 'we observed' into 'we proved forever.' Evidence becomes more valuable when another team can repeat it with available materials or state what is missing. That traceability is more useful than a sweeping label.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to separate input scope, predicted output, validation and clinical use. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.

Before closing, another person should be able to reconstruct the conclusion without knowing the headline. Give them the sources, conditions and negative case, then ask what they would accept and reject. If they need an assumed intent, a figure without a denominator or an undated later fact, the chain still has a gap. That short review catches errors that fluent prose can conceal.