THE APEX TIMES
A DeepMind research scientist describes how models can evaluate billions of possible single-letter DNA changes and turn the results into a prioritized impact score for researchers.
Google DeepMind is betting that AI can cut years out of early genomic research by helping scientists identify which DNA variants are most likely to cause disease. In a new post on the company’s blog, Žiga Avsec, a DeepMind research scientist overseeing the group’s genomics efforts, described how AlphaGenome and AlphaGenome Atlas are designed to move researchers from decoding DNA sequences toward interpreting which specific changes matter for health and disease.
Human DNA contains about 3 billion base pairs, made up of four chemical “letters” in the sequence: A, T, C and G. Avsec said each letter can be replaced by three alternatives, creating roughly 9 billion possible single-letter changes. Many genetic diseases, he added, are driven by just one of those letter-level changes, but researchers cannot realistically run studies across billions of possibilities to find the causal ones.
AlphaGenome is built for that scale problem. According to Avsec, the model takes a partial DNA sequence of up to 1 million base pairs and produces thousands of predictions. DeepMind then compares the predictions across alternative DNA changes to determine which edits are most likely to affect outcomes, aiming to reduce the “needle in a haystack” nature of variant hunting.
To make those predictions easier to use, DeepMind also created AlphaGenome Atlas, a large dataset meant to let scientists decide what to investigate sooner. Avsec said the Atlas provides a standardized “AlphaGenome Variant Impact” (AVI) score, described as a single, easy-to-read number intended to indicate the potential impact of a variant across the genome, so researchers can prioritize which variants to test next.
The practical bottleneck, Avsec argued, is that while science can read a large share of the human DNA sequence, it still has not fully deciphered the “semantics” of the genetic code, particularly the non-coding portion that he referred to as dark matter, which regulates gene expression and disease risk. He also pointed to the cost and time required for wet-lab experiments, noting that even a single breakthrough can take years.
He illustrated the challenge with a common clinical workflow: sequencing a patient with a rare disease can yield about 4 million genetic variants compared with a reference human genome. Even after narrowing to rare variants, he said researchers may still be left with hundreds or thousands of candidates, making it hard to choose where to focus experimental resources.
AlphaGenome is still described as early in its rollout. Avsec said the original model was launched in June 2025, and AlphaGenome Atlas was introduced in September of the same year. He also cited examples of early usage, including a Medical Research Council fellow at the University of Exeter who used Atlas to narrow down a list of 1 billion genetic variants to focus on the most impactful 1 percent, eventually prioritizing variant groups linked to body mass index.
Avsec also described a separate use case by a Broad Institute fellow who used the AVI score to prioritize variants in research on unsolved rare diseases. In that account, Atlas identified a critical variant tied to a gene linked with neurodevelopmental disorders, which supported subsequent research that helped solve a case. He said DeepMind’s latest model, AlphaGenome 2, is intended to address cell type specificity, moving beyond tissue-level averages to predict how variants disrupt individual cell types that can malfunction in disease.
Still, the blog post underscores that predictions are not the same as proofs. Avsec said DeepMind is working on better ways to estimate how confident the model is in its results, and he cautioned that anyone using AlphaGenome should interpret predictions alongside other information. The post did not provide independent performance metrics, validation studies, or details on how the AVI score maps to specific biological mechanisms, leaving some uncertainty about how often the highest-scoring variants translate into causal findings across different disease areas.
Looking ahead, DeepMind’s stated goal is to use AI to understand the “grammar” of the genome and accurately predict mutation effects, with a longer-term aim of building toward a computational simulation of a human cell. For researchers, the near-term question is whether faster prioritization can consistently improve the odds of identifying causal variants in rare diseases and, by extension, accelerate drug discovery target selection. The next developments to watch are how AlphaGenome 2 performs in cell type-focused settings and what confidence calibration DeepMind introduces for researchers deciding what to test.
A new post-market snapshot suggests Amazon’s stake in the AI startup will be just under 15% after the offering, with an option that could raise exposure, while Google is expected to control nearly 14%, together totaling about 30%.
A Yahoo Finance commentary points to Arista Networks as a notable beneficiary as investors look past Nvidia’s dominance, while noting leadership transition at Arista’s top. The piece argues Arista is chipping away at one of Nvidia’s biggest practical advantages, though key specifics were not detailed in the material provided here.
Market-moving earnings headlines helped lift U.S. futures and drive a rebound in technology shares after Nvidia reported results that exceeded expectations, even as CEO Jensen Huang indicated he is not returning to a previously stated target.
Salesforce (CRM) closed at $194.52 in the latest session, down 9.45% from the prior day, according to a market report published Nov. 17.
A broad risk-off tone returned to Wall Street, with technology shares pressured by worries tied to the AI buildout and fresh disruptions in the tech sector. Salesforce slid about 9% in the session described by Yahoo Finance.
In remarks tied to Nvidia’s latest results, interim CEO Colette Kress pointed to a quarter of $116 billion in revenue, describing the performance as a major year-over-year acceleration and linking it to the company’s longer-range growth outlook.
The company’s board backed interim leadership under Colette Kress and said it is now evaluating the next permanent chief executive, as it also moves to strengthen oversight of semiconductor exports to China.
In a public statement carried by Yahoo Finance, Nvidia’s board chairman pushed back on calls to reverse a leadership decision involving Jensen Huang, saying the company will not reinstate him to the role he previously held.
Ahead of its third-quarter earnings report, Nvidia stock weakened in regular trading, then surged after-hours after the company disclosed results described as record revenue.
The company delivered revenue of $115.8 billion and guided for $130 billion in the forecast period, explicitly excluding China chip sales from that projection, according to a market report. Leadership transition pressures and the rollout of the Vera Rubin AI computing platform are also in focus.
Comments
Community guidelines. Comments may be removed for violations.
Loading comments...