Google's artificial intelligence helps calculate genome changes worthy of (another) Nobel Prize?

AlphaGenome predicts the effect of variations in DNA and can help understand complex diseases.Another Google DeepMind model won the Swedish gold medal two years ago, but this new AI has competition and limitations

Is Google's genome-reading artificial intelligence a Nobel Prize-worthy revolution?

AlphaGenome predicts the impact of DNA variants and can help understand complex diseases.Another Google DeepMind model won the Swedish gold medal two years ago, but this new AI has competition and limitations

Humans have approximately 20,000 genes.While that may seem like a lot, there are actually many more in a rice grain (about 45,000), and they make up just 2% of the genome.Our complexity lies not here, but in how they are organized and activated.The systems that control them are conserved in the remaining 98%, or unexplored parts of the genome.It was known as junk DNA.Interpreting all this information is one of the great unsolved challenges of modern genomics.Doing this is very complex and requires computing power...artificial intelligence could be the key.

Google DeepMind has been trying to become a reference in this field for ten years, and now it has introduced AlphaGenome, an AI tool designed to understand the 98% darkness and how small differences in our DNA can change the function of genes.Achieving this is not only an interesting problem to solve, but also important for diagnosing diseases and developing new treatments.

Two years ago, Google researchers won the Nobel Prize in Chemistry for the development of a similar tool applied to proteins, AlphaFold.But the new AI effort still has technical limitations and seems surprising enough to add another medal to the history of big tech.

mark mutation

While 2% of our genome is made up of genes, the remaining 98% is occupied by various families of repetitive DNA, mobile elements, transposons, and other terms we are not familiar with.Among these, regulatory factors."These are small sequences of DNA that determine when and where proteins are activated and when and where they are deactivated, which determine the function of the gene," Luis Montoliu, a scientific researcher at the National Center for Biotechnology, explains to El Confidential (CNB-CSIC).It should not be done and it can cause congenital diseases."

In 2003, it was possible to obtain the complete sequence of the human genome, but knowing how to read this book and understand the regulatory elements is another story, very difficult to achieve without technology.Until now, there were computer programs that could compare sequences, but "it was done with relatively small sequences of a few thousand nucleotides, with alphagenomes up to a million nucleotides."

"This can help us know where to put the magnifying glass to find mutations that cause 'genetic fatal diseases,'" adds Montolio, who is an international leader in rare disease research."It's a wild development and incredible progress," he believes.

Our breakthrough AI model AlphaGenome helps scientists understand our DNA, predict the molecular impact of genetic changes and drive new biological discoveries.🧬

- Google DeepMind (@googledeepMind) Uni 28, 2026

Find out more at @Nature ↓ https://t.co/jvBLRXYzdj pic.twitter.com/WEL4Ptdv06

These types of models are particularly important for promotion in oncology research."When we sequence a patient with a particular tumor or look for a mutation that causes an inherited cancer, we often find mutations that have not been described before because it is done through laboratory tests," says Dido Carrero, a researcher at the National Center for Cancer Research (CNIO).But AlphaGenome allows us to guess what effect these mutations might have and track them down.This researcher adds: It happens randomly or it is the cause of this disease.

Google released the first results of its AI in DeepMind Nature, a deep learning model that it introduced in June 2025. According to the developer, it was able to improve the predictions of existing models in 25 out of 26 tests. The scientific community agreed that this is a milestone that is very good for some tasks, but still not perfect.

A tool with limitations

"The challenge of research with this AI model is to know what they give reliable answers and what they do not. In the case of AlphaGenome, it can do this in the process of genomic regulation, but at the moment it is not reliable for personalized medicine", Mafalda Dias, who heads the research group on Artificial Intelligence and Genomics at CR.

Not only do genomic studies have specific limitations, but there are also limitations inherent in AI models: lack of interpretability and bias in understanding how the model works. The first "has always been a problem with all mathematical models, but the problem is magnified because AI models are black boxes," says Dias.

AlphaGenome researchers say they are working to prevent this."One of the main ways to do this is to trace the origin: when the model predicts the effect of a particular variant, what is it actually looking at? To find out, we simulate mutations and recalculate the model's scores," Natasha Latysheva and Žiga Avsec from Google DeepMind explain to El Confidencial in a press conference.species to intuition based on what it predicts.

"This is a powerful tool that helps to better understand the basics of biology," Mafalda Díaz from CRG

In the case of biases, part of the problem comes from the data."There are many factors that can cause a biased model: age, gender, environmental conditions... Imagine collecting data from 80-year-old people because they have a certain disease, because you ignore the data from 30-year-old people and the model may not work," explains Beatriz López, coordinator of the research laboratory of the University of Medicine and Health at the University of Medicine and Health.in a conversation with this newspaper.

Limits or not, "we must remember that it is a tool and then the specialized scientist will have to interpret its results, not take it as a solution", underlines López.Diaz calls for ethics and "responsible use."Undoubtedly, by doing so, this scientific discovery can bring "the greatest benefit to mankind"... the condition that Alfred Nobel set for the distribution of his prizes.

And the Nobel goes...

It's been a decade since Google Deepmind began investigating artificial intelligence for molecular biology and medicine.In 2016, he started working with hospitals with clinical data analysis, and in the same year a milestone happened: Alphago beat Profession in science and made it go faster.

And the magic began.The AlphaFold family was launched in 2018 and since then four improved versions have been released.There's also AlphaMissense (2023), which predicts whether a change in a DNA letter will lead to disease, and AlphaProteo (2024), for creating proteins from scratch.Sam Hasabis was one of the Nobel Prize winners for AlphaFold.

Winning a Nobel Prize for a scientific discovery is complicated.This implies international consensus may take years and depends on its novelty and impact, but there is no avoiding the question of whether AlphaGenome is up to the task.

"AlphaFold is a revolution because it was one of the first deep learning models that appeared in the explosion of AI. Now, there are new predictive models every time. We are also developing one," said Carrero of the CNIO.The researcher thinks that it is a very competitive field and it is difficult to stand out: "I hope to win, but I think that these new models will not be noticed by the international community and the general opinion."

"I can't say that this study is a revolution, because it is part of a series of articles and the work of the whole community that develops these types of models," admits Dias.The expert also points out that AlphaFold solves the problem of predicting protein structure, but AlphaGenome has not found a specific key to read the genome.

“AlphaFold was groundbreaking because predicting the structure of proteins with such reliability and using only DNA sequences was something we didn't know how to do.On the other hand, while AlphaGenome is making a 'quantitative leap', what it was achieving was already known how to do, just in smaller pieces."

Scientists agree that although it is very useful, it is currently not at the same level of "surprise" as its predecessor and has limitations.Whether or not he wins the Nobel Prize will be years away;there is no AI model developed by Google that can predict it.

Humans have about 20,000 genes.They may seem like a lot, but in reality even a grain of rice is larger (about 45,000) and occupy only 2% of the genome.Our complexity is not there, but how they are organized and activated.The system that controls them is stored in the remaining 98%, a part of the genome so unexplored that it was called junk DNA.The interpretation of all this information is one of the great challenges ahead in modern genomics.Doing so is so complex that it requires computing power...and understanding that could be crucial.

- The epidemic of artificial intelligence studies is destroying scientist Vitomir Kovanovic

- He solves brutal murders in the Middle Ages with a new method of genetics Omar Cardaudi

- Video worth 4 billion: Google fills the AI ​​bubble with its new investment Manuel Ángel Méndez