Artificial Intelligence


By Lottie


September 2023

Introducing AlphaMissense - The Future Of Genetic Disease Prediction


Imagine the potential of promptly identifying genetic anomalies linked to diseases. DeepMind, a division of Google, introduces AlphaMissense, a breakthrough AI tool that deciphers DNA mutations and promises to fast-track research on rare diseases.

Rewind a decade. A physics PhD student named Žiga Avsec ventured into the intricate realm of genomics, spurred by a machine learning course. His journey led him to a challenging quest – uncovering a rare mitochondrial disease's genetic root. Such a task was akin to finding a proverbial "needle in a haystack", with countless potential DNA mutations possibly affecting human health. The spotlight was on missense variants - single-letter genetic code changes that modify amino acid production in proteins. Even minor alterations here can unleash significant consequences on our health.

To give you a clearer picture, there are around 71 million missense variants in the human genome. While an individual may carry over 9,000 of them, identifying which ones are harmful remains a substantial challenge. Only a tiny fraction (about 2%) of the 4 million observed missense variants in humans has been classified as benign or pathogenic. Typically, months of rigorous research go into understanding one single variant.

Enter DeepMind's newest innovation. AlphaMissense, the brainchild of the now research scientist Avsec at Google DeepMind, is a game-changer. This machine learning model can swiftly evaluate missense variants, boasting a remarkable 90% accuracy in predicting disease likelihood, surpassing existing technologies.

It's noteworthy to mention that AlphaMissense has its foundations in AlphaFold, another pioneering model by DeepMind. However, while AlphaFold predicts protein structures, AlphaMissense operates similarly to advanced language models, like OpenAI’s ChatGPT. Trained on the intricate language of human and primate biology, it identifies abnormal amino acid sequences, much like how we spot out-of-place words in a text.

Jun Cheng, a collaborator with Avsec, eloquently likens AlphaMissense to a linguistic expert of protein sequences. By contrast, Pushmeet Kohli of DeepMind compares AlphaFold to understanding ingredient binding in recipes, while AlphaMissense forecasts outcomes of incorrect ingredients.

Enjoying what you're reading?

See our work

DeepMind, in collaboration with Genomics England, has also introduced a "pathogenicity score" for each missense variant. The score, ranging from 0 to 1, gauges the mutation's likelihood of causing diseases based on known related mutations.

The vision for AlphaMissense is vast. Researchers believe it will empower the scientific community by offering quicker insights into genetic variations. This will enable them to diagnose and formulate treatments more rapidly. Yet, it's crucial to remember that AlphaMissense's predictions are guiding lights, not definitive answers. They're tools to aid researchers in prioritising and understanding our genetic code more profoundly.

While AlphaMissense's achievements are immense, experts like Ewan Birney of the European Molecular Biology Laboratory consider its predecessor, AlphaFold, to be even more revolutionary. Still, Birney acknowledges the significant benefits AlphaMissense offers, particularly in aiding clinicians to diagnose genetic conditions swiftly.

As we celebrate this innovative tool, it's evident that models like AlphaMissense underscore the vast capabilities of AI in biology. They provide glimpses of the infinite possibilities in understanding our genome and its manifestations.

Enjoyed This Tech Insight?

Don't miss out on the latest trends and developments by subscribing to our Tech Roundup Newsletter where every two weeks, we share the biggest tech headlines from around the world.

Make sure to subscribe here!