AlphaMissense: An AI-powered Catalogue of Missense Mutations

Joud Baddawi

October 8, 2023

Google's AI company and research laboratory, DeepMind, has recently presented a new AI tool that can predict whether a genetic mutation is pathogenic or benign.

An individual has around 9,000 mutations, most of which are benign. However, some of these mutations are responsible for diseases such as cystic fibrosis, sickle-cell anaemia, or cancer, DeepMind said.

This breakthrough can help scientists dive into the root causes of disease, which can lead to faster diagnoses and developing treatments.

Back to Biology

Missense variants or mutations are a form of genetic mutations happening at the DNA level where a single nucleotide - one letter; A, T, G, or C - changes in the DNA sequence of a gene.

This results in the substitution of one amino acid in the protein encoded by that gene with a different amino acid. Missense variants can be either benign (having no impact on protein function) or pathogenic (leading to changes in protein function and potentially causing disease).

How AlphaMissense Classifies Variants

AlphaMissense is based on an AI model called AlphaFold. AlphaFold, developed by DeepMind, can predict the 3D structure of any protein based on its amino acid sequence at scale and in minutes, down to atomic accuracy.

AlphaFold was presented by Google DeepMind in 2018, which published the largest database of proteins with over 200 million structures available.

To train AlphaMissense, AlphaFold was fine-tuned on labels distinguishing variants seen in human and closely related primate populations. Variants commonly seen are labelled as benign, and variants never seen are labelled as pathogenic.

AlphaMissense uses databases of related protein sequences and structural context (3D shapes) of variants to produce a score between 0 and 1. This score rates the likelihood of a variant being pathogenic. The score allows researchers to choose a threshold for classifying variants as pathogenic or benign.

AlphaMissesne Results

The three-year project was led by DeepMind engineers Jun Cheng and Žiga Avsec, and the company said it is publicly releasing predictions for 71 million possible variants. AlphaMissense predicted that 57 percent of all mutations were likely benign and 32 percent were likely pathogenic, with the rest remaining uncertain.

The AlphaMinnesne database is available easily to any researcher through Github. However, DeepMind will not be releasing the whole model for immediate download and use by others. The authors of the research paper, published in Science magazine stated this is due to biosecurity risks.

"As part of our commitment to releasing our research breakthroughs safely and responsibly, we will not be sharing model weights, to prevent use in potentially unsafe applications."

In conclusion, AlphaMissense represents a significant advancement in genetic research and healthcare. It can lead to quicker diagnosis of genetic disorders, more personalised and targeted therapies for individuals with pathogenic mutations, and can help researchers gain deeper insights into the underlying mechanisms of diseases.

Because AlphaMissense's predictions are available to researchers, this will accelerate genetic research efforts, ultimately benefiting patients and the healthcare industry as a whole.