In a revelation of a breakthrough in artificial intelligence (AI), Google DeepMind scientists built a program with the ability to predict if millions of genetic mutations can likely cause disease or are harmless. 

The program, called AlphaMissense, which catalyzes research and diagnosis of rare diseases, predicts so-called missense mutations, meaning a single letter is misspelled in the DNA code, and even though in some cases, they tend to be harmless, in others, they can also interfere with protein work and cause illnesses, such as cystic fibrosis, sickle-cell anemia, cancer, and brain development problems.

AlphaMissense was applied to assess all 71 million single-letter mutations affecting human proteins. As soon as the program precision was set to 90%, it predicted that 57% of missense mutations were mostly harmless and 32% were probably harmful. The rest were deemed uncertain.

The system was also developed to have the capacity to flag mutations that were not previously linked to specific disorders and point doctors to better treatments.

Thus, the scientists released a free online catalog of the results to aid geneticists and clinicians who are studying how mutations cause diseases or diagnosing patients with rare disorders.

Read more: Google testing AI called 'Genesis' to replace news, article writing

A typical person has around 9,000 missense mutations in their genome, and out of the 4 million in humans, only 2% are classified as being benign or pathogenic. Although doctors work with programs that predict the diseases caused by mutations, they can only offer supporting evidence for making a diagnosis due to the predictions being inaccurate.

Dr. Jun Cheng - who is an author for Science - alongside others, demonstrates how AlphaMissense works better than current "variant effect predictor" systems as it should aid experts in locating better which mutations are causing diseases.

Just like the English language

The AI program is an adaptation of DeepMind’s system that predicts the 3D structure of human proteins from their chemical makeup called AlphaFold program.

AlphaMissense was given data on DNA from humans and closely related primates to learn the common missense mutations and most likely benign and those that are rare and potentially harmful. Simultaneously, the system adapted itself to the "language" of proteins by studying millions of protein sequences and knowing what a "healthy" protein looks like.

Then, when the AI is fed a mutation, it produces a score reflecting the risk of the genetic change, but it cannot say how the mutation causes any problems.

"This is very similar to human language," Cheng said, adding, "If we substitute a word in an English sentence, a person familiar with English can immediately see whether the word substitution will change the meaning of the sentence or not."

Prof. Joe Marsh, a computational biologist at Edinburgh University who was not part of the study, claimed that AlphaMissense had "great potential".

"We have this issue with computational predictors where everybody says their new method is the best," he stated, noting, "You can’t really trust people, but [the DeepMind researchers] do seem to have done a pretty good job."

Read next: 'Godfather of AI' warns of tech dangers; quits Google

Marsh suggested that if clinical experts conclude that AlphaMissense is reliable, the program's predictions carry more weight and mean more in future disease diagnoses.

Professor Ben Lehner, senior group leader in human genetics at the Wellcome Sanger Institute, argued, however, that the predictions would still need verification by other scientists, but it seemed good at identifying which DNA changes drive diseases and which don't. 

"One concern about the DeepMind model is that it is extremely complicated," Lehrer said.

"A model like this may turn out to be more complicated than the biology it is trying to predict. It’s humbling to realise that we may never be able to understand how these models actually work. Is this a problem? It may not be for some applications, but will doctors be comfortable making decisions about patients that they don’t understand and can’t explain?"

“The DeepMind model does a good job of predicting what is broken," he added.

"Knowing what is broken is a good first step. But you also need to know how something is broken if you want to fix it. Many of us are very busy generating the massive data needed to train the next generation of AI models that will tell us not only which changes in DNA are bad but also exactly what the problem is and how we might go about fixing things."