Courtesy of Cheng
Curing incurable diseases, detecting undetectable medical events, and playing chess are just a few of the missions of Google DeepMind, an artificial intelligence (AI) research company looking to solve pressing world issues by pursuing artificial general intelligence (AGI) and other specialized AI projects. Described by The Guardian as “the most advanced AI research outfit in the world,” DeepMind has projects devoted to machine learning, language processing, and game theory. However, its recent groundbreaking work in genetics has catapulted DeepMind to be a popular topic within healthcare discourse. On September 19, 2023, DeepMind released its catalog of missense variations (or, changes in DNA due to swapping amino acids) developed by its novel AI AlphaMissense. AlphaMissense is a deep-learning AI that aims to determine whether or not specific amino acid mutations will cause adverse effects.
“It categorised 89% of all 71 million possible missense variants as either likely pathogenic or likely benign. By contrast, only 0.1% have been confirmed by human experts.”
The foundation of AlphaMissense comes from AlphaFold, another DeepMind program that “[predicts] 3D models of protein structures and is accelerating research in nearly every field of biology.” Every protein is made up of a string of amino acids folded in a certain way. The structure determines the protein’s function and is incredibly important in understanding the protein’s role in the human body. There is one catch: determining the structure is incredibly difficult. It is so difficult that it has given rise to a critical issue in biological modeling known as the “protein folding problem,” which questions how the structure of a protein is predicted just from its amino acid sequence. Just like a long string of randomly arranged letters, there is no valuable information unless you can rearrange these letters into the structure of words. This is what AlphaFold does, only with amino acid sequences and proteins instead of letters and words. It solved the protein folding problem in 2020 and has only kept growing. It currently has over 200 million protein structure predictions in its public database. According to DeepMind co-founder Demis Hassabis, this includes “almost every organism on the planet that has had its genome sequenced.”
Noting this incredible feat, it is important to understand exactly what missense mutations are and why they matter. Think back to the example of letters and proteins. Human beings are made up of an assortment of proteins, each of which performs a specific function. A missense variation changes a letter and creates a new word, and this new word has the power to change a person’s health completely.
Regarding DNA and protein, “a single letter substitution…changes which amino acid is translated, which can affect the function of a protein.” Many of these changes are harmless. Others can cause devastating diseases like cancer, sickle cell anemia, and cystic fibrosis. The work of AlphaMissense detects which missense mutations are likely to be pathogenic, helping prioritize research and contributing to the study of rare diseases.
However, this work is not without criticism. As international concern over the misuse of AI grows, so must the oversight and ethical commitments of AI companies such as Google DeepMind. DeepMind co-founder Mustafa Suleyman, in an interview with David Shariatmadari, acknowledges the potential for danger that comes with the use and development of AI. In regards to his grim predictions surrounding the impact of AI on society, he says, “I hope I’m wrong…I don’t think [I am].” A Safety and Ethics group at DeepMind is currently working with groups worldwide to develop careful, equitable safeguards for AI. They emphasize technical safety and responsibility and have signed pledges vowing their commitment (for example, not using AI in autonomous lethal weapons). While these policies are only as strong as their self-enforcement, DeepMind appears to be a leader in AI ethics.
Another concern, writes Tina Hesman Saey, author for Science News, is false expectations. She states that AlphaFold, the protein-folding AI, uses “predictions, not actual structures.” DeepMind’s website has a disclaimer encouraging healthcare providers and researchers to use other research sources in conjunction with AlphaMissense. While this breakthrough is a huge step forward in the biotechnology field, more work must be done. In the podcast DeepMind, host Hannah Fry and research scientist Kathryn Tunyasuvunakool discuss using AlphaFold (the basis for AlphaMissense) to study neglected tropical diseases. “It basically changes the situation from…waiting several years for the experiments to be finished, and now we have this middle way of…[getting] some actionable structural information within ten to fifteen minutes,” Tunyasuvunakool claims (30:40). Experiments that typically require years of human resources and millions of dollars to conduct are finished in days; diseases that would take the lives of millions over those years are that much closer to a cure. “It’s a huge saving,” says Tunyasuvunakool.
Google DeepMind has big dreams, but it also has a history of success and astounding innovation. This trend of beating the odds will have to be weighed against its obligation to ethical development, especially as AI systems are quickly becoming more complex and powerful. As it stands now, though, DeepMind has one goal: continuing its groundbreaking research in AI to improve the world.