The first obstacle wasn't biological, it was computational. When the team led by Žiga Avsec at Google DeepMind began thinking about precomputing predictions for every possible single-letter change in the human genome, early estimates said they would need to improve calculation speed by about 80 times to finish in a reasonable timeframe. The announcement on 8 September, with the release of the AlphaGenome Atlas, turns that effort into a public resource: nine billion precomputed variants in a dataset of about one petabyte.

DNA is often described as an instruction manual for producing proteins, but the protein-coding portion is a minority. Gene activity is largely governed by non-coding regulatory elements that can act at great distances in the genome. Carl de Boer, a genomicist at the University of British Columbia, notes that understanding how DNA changes alter this regulation 'is fundamental to understanding most disease.'

AlphaGenome, announced by DeepMind in 2025, compares an original DNA sequence with an altered one and predicts the effect on gene expression and other regulatory activity. Until now, using the model meant selecting variants, writing code, and running a computationally demanding job. With the Atlas, DeepMind has done that work in advance for all possible single-letter changes to a reference human genome. The online interface lets scientists consult predictions without writing code and without running AlphaGenome locally.

The headline feature is a single-number impact score, designed to show at a glance whether a variant is likely to be meaningful. The Atlas includes eleven different output types, but the synthetic score is the entry point. De Boer cautions that the score has a clear use but risks being easily misinterpreted: 'We are talking about a very complex system with many moving parts.'

There are known limitations. Many diseases are associated with multiple genetic variants. AlphaGenome also looks at a segment of one million base pairs around the variant; some enhancers can regulate genes over much longer distances, beyond the model's field of view. The Atlas is therefore a tool for filtering possibilities and prioritizing lab experiments that could validate its predictions.

The less-told part is the engineering work. The human genome contains roughly three billion base pairs; at each position there are three possible single-nucleotide substitutions, hence the nine billion variants in the Atlas. To produce the full dataset of about one petabyte, the team combined model distillation, GPU kernel optimization, and the elimination of redundant calculations. Avsec describes the shift: 'When we started thinking about this project, it seemed impossible to do that computationally.'

Precomputing everything and making the archive public changes how computational cost is distributed. Instead of running the model in dozens of labs, DeepMind concentrates the computation once and distributes queryable results. This helps groups without access to recent GPUs and reduces the repetition of the same simulations. De Boer considers AlphaGenome the field's leading model, but also very slow and computationally intensive. The Atlas reduces exactly that friction.

The downside is dependence on an archive curated by a single vendor, available free for non-commercial research and with potential for commercial licensing. From a data-control and sovereignty perspective, a petabyte-scale dataset is not trivial to replicate or host in a self-hosted setup; anyone wanting to verify the predictions still faces substantial costs. It is a structural trade-off: centralizing computation lowers access barriers but reduces operational transparency. For those evaluating similar on-premise scenarios, the balance between access and control remains the central question; AI-RADAR publishes analytical frameworks at /llm-onpremise to help weigh these trade-offs.

The main risk is not a lack of predictions, but their apparent simplicity. The impact score may push users to read a result as a verdict, while gene regulation remains a system with many moving parts. The resource will accelerate screening and prioritization, but it will not replace experimental validation.