Google DeepMind Trains AI to Predict Disease-Causing DNA Mutations in the Human Genome

Google DeepMind, known for its groundbreaking artificial intelligence (AI) programs, has recently made a significant advancement in the medical field. The company announced that it has successfully trained AI to predict disease-causing DNA mutations in the human genome. This breakthrough technology is expected to revolutionize the diagnosis of rare diseases and potentially lead to the development of new drugs.

DeepMind, based in London and acquired by Google a decade ago, rose to prominence for its AI programs that excel at playing video games and mastering complex board games like Go. The company made its foray into the medical domain by introducing AlphaFold, a program that accurately predicts protein structures – a major challenge in biology.

Now, DeepMind claims that it has enhanced its protein prediction model to identify ‘spelling errors’ in human DNA that are either trivial or likely to cause disease. This new AI software, called ‘AlphaMissense’, was publicly released in mid-September through a journal publication.

As part of this project, DeepMind has released millions of prediction results pertaining to disease-causing mutations. However, due to concerns about potential biosecurity risks if the technology is misused on species other than humans, the company does not allow direct downloads of the models.

Currently, doctors utilize computer-aided predictions to identify genetic causes of mysterious syndromes. However, this technology is not used to make direct diagnoses. DeepMind’s research aims to uncover the root causes of diseases, leading to faster diagnoses and life-saving treatments.

This three-year project, led by DeepMind engineers Jun Cheng and Žiga Avsec, has yielded prediction results for 71 million disease-causing mutations. Each mutation represents a missense mutation, where a DNA base sequence is modified, resulting in a change in the protein produced by the gene.

Experts have hailed DeepMind’s announcement as remarkable, despite the undefined commercial value. Alex Zhavoronkov, the founder of AI drug development company Insilico Medicine, commended DeepMind’s ability to conduct incredible research and promote the company.

While the potential for AI-designed drugs is still being explored, investors are optimistic about the development of new proteins. Generate Biomedicines, for instance, has secured $273 million to create antibodies, and a team of former Meta engineers has utilized AI in the pursuit of cancer discovery.

DeepMind’s latest endeavor primarily focuses on aiding doctors in diagnosing rare diseases, particularly in cases where patients exhibit perplexing symptoms. By leveraging AI to predict the pathogenicity of DNA mutations, AlphaMissense can assist in providing answers.

Heidi Rehm, director of clinical laboratories at the Broad Institute in Cambridge, Massachusetts, explains that approximately 25% of cases involve DNA mutations with uncertain health implications, even after extensive genetic sequencing.

Known as ‘variations of uncertain significance’, these mutations appear in genes that have already undergone thorough examination, such as the BRCA1 gene associated with hereditary cancer risk.

DeepMind’s AlphaMissense has the potential to shed light on these mutations by accurately predicting which DNA variations are benign or pathogenic. This AI tool builds upon AlphaFold’s protein prediction model and offers rapid and energy-efficient software.

While the model signifies a significant advancement in AI, it does not definitively determine whether a DNA mutation causes disease. Laboratory findings, patient data, patterns of inheritance, and shared information from mutation databases are necessary to confirm disease-causing mutations.

DeepMind, in collaboration with external biosecurity experts, has disclosed most of its research findings and provided the necessary information to replicate their work. However, due to concerns about misuse, the complete model is not readily available for download, limiting predictions to human protein sequences.

By taking into account the risks associated with the misapplication of AI, DeepMind aims to prevent unauthorized usage and potential risks. The company’s responsible AI approach has been evaluated by both DeepMind Lab and external biosecurity experts.

DeepMind’s breakthrough in predicting disease-causing mutations marks a significant step forward in the medical field. While there is still progress to be made, this advancement has the potential to revolutionize rare disease diagnosis and greatly impact the development of life-saving treatments.

Google DeepMind announced that it has trained artificial intelligence (AI) to predict potentially disease-causing DNA mutations in the human genome. Using this predictive technology is expected to speed up the diagnosis of rare diseases and provide clues for the development of new drugs.

DeepMind, founded in London, England and acquired by Google 10 years ago, is best known for its AI programs that play video games and conquer complex board games like Go. DeepMind entered the medical field by publishing a program called AlphaFold, which can accurately predict the shape of proteins, a problem considered a ‘grand challenge’ in biology.

Now, DeepMind says it can refine its protein prediction model to predict which ‘spelling errors’ found in human DNA are trivial and which are likely to cause disease. This new AI software, called ‘AlphaMissense’ <사이언스(Science)> It was made public in mid-September through a report published in a journal.

DeepMind is releasing tens of millions of prediction results as part of this project, but does not allow direct downloads of the models due to concerns about potential biosecurity risks if the technology is applied to species other than humans.

Doctors already use computer-aided predictions to find genetic causes for mysterious syndromes, but not to make a direct diagnosis. DeepMind said in a blog post that the results of this research are “part of an effort to discover the root of the disease” and “could lead to faster diagnosis and the development of life-saving treatments.”

This project was carried out over three years, led by DeepMind engineers Jun Cheng and Žiga Avsec, and DeepMind released prediction results for 71 million mutations that have the potential to cause disease that were revealed. Each mutation is a missense mutation, which means a mutation where one DNA base sequence is changed to another, leading to a change in the protein produced by the gene.

Stephen Hsu, a physicist at Michigan State University who studies genetic problems using AI technology, said, “The aim of this study was to make a change to a protein and then predict the shape of the protein. “The aim is to find out if it can cause disease in people, ” he explained. “In most cases, even if a mutation occurs in a protein, we don’t know if that mutation causes disease.”

Outside experts said DeepMind’s announcement, like other previous announcements, was a big surprise, although its commercial value was still unclear. Commenting on this announcement, Alex Zhavoronkov, founder of Insilico Medicine, an AI drug development company, said, “DeepMind has produced results worthy of DeepMind,” adding, “DeepMind is not only excellent at promoting the company, but “We are also excellent. conduct incredible research on AI.”

“The real test of modern AI is whether it can lead to new treatments, and that hasn’t happened yet,” Zavoronkov said. However, some AI-designed drugs are in the experimental stage, and investors believe that research to develop useful new proteins is an area of ​​particular interest. According to Forbes, a company called Generate Biomedicines has raised $273 million to create antibodies, and a team of former Meta engineers is using AI to ‘discover cancer’. He founded EvolutionaryScale with the idea that he could create cells that could be programmed to destroy themselves.

better model

However, DeepMind’s latest effort has less to do with developing new drugs and more to do with doctors diagnosing rare diseases, especially in patients with mysterious symptoms, such as newborns with persistent rashes or adults with sudden weakness.

With the development of gene sequencing, doctors can now decode a person’s genome and then comb through the DNA data to discover the cause of the disease. Sometimes the cause is obvious, like the mutation that causes cystic fibrosis, but “we do extensive genetic sequencing,” says Heidi Rehm, director of clinical laboratories at the Broad Institute in Cambridge, Massachusetts. “In about 25% of cases, even when scientists discover a suspicious DNA mutation, they don’t fully understand the health implications of that mutation,” he explained.

Scientists call these mysterious mutations ‘variations of uncertain significance,’ and they occur even in genes that have already been thoroughly studied, such as the BRCA1 gene, which is famous for being linked to the risk of developing hereditary cancer. “Any gene can have these mutations,” Lehm said.

DeepMind says AlphaMissense can help find answers by using AI to predict which DNA mutations are “benign” and which are “pathogenic,” meaning they have the potential to cause disease. Before AlphaMissense, there were programs used for prediction similar to this model, such as PrimateAI.

“A lot of research has already been done in this area, and the overall quality of these computer predictors has gotten much better,” says Lehm. “Computer predictions are only ‘one piece of evidence’, and on their own they do not indicate whether DNA mutations actually cause disease.” “It’s hard to be sure it can cause this,” he said.

Typically, experts do not know which mutation ’causes disease’ until they have laboratory findings, such as actual patient data, evidence showing patterns of inheritance within families, and information shared through public mutation sites such as ClinVar. declare it to be a mutation.

“Models are getting better, but no model is perfect yet, and those predictive models don’t tell us definitively whether a DNA mutation causes a disease,” Lem said. He added that he was “disappointed” because he appeared to be exaggerating the medical certainty.

fine tuning

According to DeepMind, this new model is based on AlphaFold, an early model for predicting protein shape. Pushmeet Kohli, DeepMind’s vice president of research, said that while AlphaMissense performs a very different task than AlphaFold, it “leverages intuition about biology” gained from previously predicting protein shapes. Because this model is based on AlphaFold, it takes relatively little time to run the software and uses less energy than if the model were built from scratch without a base model.

From a technical point of view, AlphaMissense is pre-trained and then adapted to new tasks through an additional step called ‘correction’. For this reason, Patrick Malone, a physician and biologist at KdT Ventures, describes the model as “one of the most important recent methodological developments in the field of AI.”

“This model shows that refined AI can leverage prior learning, an area in computational biology that is often limited in accessing data at a sufficient scale,” Malone continued. “It’s particularly useful,” he emphasized.

biosecurity risks

DeepMind said it has released all its predictions about the human genome for free, as well as all the details needed to fully reproduce its work, including the computer code. However, because there may be a ‘biosafety’ risk if this model is applied to genetic analysis of species other than humans, the entire model is not available for immediate download and use.

“In order to safely and responsibly disclose the results of our research, we will not disclose the model’s weight to prevent the model from being used for potentially unsafe purposes,” the authors wrote in fine print in the paper.

It is not clear what ‘unsafe use’ is referred to here, or what ‘species other than humans’ the researchers have in mind. Although DeepMind did not specify, the risks of such abuse could include using AI to design more dangerous germs or biological weapons.

But at least one outside expert we spoke to described DeepMind’s limitations as “a transparent effort to prevent anyone else from quickly adapting and using this model.” This outside expert requested anonymity because Google invests in several companies he founded.

DeepMind denied speculation that there could be reasons other than security for restricting the use of models. A DeepMind spokesperson said, “This study was evaluated by Google’s DeepMind Lab, which studies responsible AI, and ‘external biosecurity experts’.”

“The limitations on the model are intended to limit protein sequence predictions for species other than humans,” DeepMind said in a statement. “Not disclosing the weight could prevent people from downloading the model and using it on species other than humans.” “This reduces the likelihood of malicious actors exploiting this model.”

#DeepMind #develops #find #genetic #diseases #MIT #Technology #Review