International Business Department           Liu Bojia           September 25, 2023

  The intricate relationship between genetic mutations and disease has always puzzled scientists. In recent years, it has been possible to sequence DNA and locate genes with mutations, but whether these mutations are disease-causing and how they affect human health remain largely unanswered questions. Uncovering the root causes of disease remains a major challenge for genetics.

  Just as AlphaFold solved the 50-year puzzle of protein structure prediction, this time, artificial intelligence is once again the key to breaking the mold.

  In a paper published in Science this week, the DeepMind team developed a new tool called AlphaMissense based on AlphaFold. The model can be used to study the effects of missense mutations on disease, and can accurately identify pathogenic versus benign missense mutations, as well as provide potential ideas for the diagnosis of genetic diseases.

  Among the different types of mutations in our bodies, missense mutations can be considered a particularly challenging "black box". Simply put, a base substitution in the coding region of an amino acid results in the synthesis of a protein in which the original amino acid is replaced by another. Such changes may have no health consequences, i.e., benign mutations; however, they may also directly disrupt the function of the protein and cause disease. Disease-causing missense mutations may not only directly contribute to a wide range of rare genetic diseases, but also play a role in the study of complex systemic diseases such as type 2 diabetes.

  Each of us carries an average of over 9,000 missense mutations, most of which are neutral and a few of which are dangerous disease-causing mutations. However, it is currently difficult for the scientific community to distinguish between the two. Of the more than 4 million human missense mutations that have been identified, only about 2% have been determined to be either benign or disease-causing, while the vast majority are "mutations of unknown significance". When scientists discover a new missense mutation, we have no way of knowing whether it is benign or pathogenic.

  In the latest study, the DeepMind team built the new AlphaMissense tool based on AlphaFold's method of predicting protein structure from gene sequences. Of course, what AlphaMissense does is not to predict the structural changes and effects of mutated proteins, but to utilize AlphaFold's structural "intuition" to identify regions within proteins where disease-causing mutations may occur.

  Specifically, for input missense mutations, AlphaMissense utilizes a database of relevant protein sequences and variant structures to score the pathogenic risk of the missense mutation.

  Using AlphaMissense, the team predicted the pathogenicity of all 216 million possible single-amino-acid changes from nearly 20,000 classical human proteins, which led to the hypothesis that as many as 71 million missense mutations may exist in the human genome - far more than the more than 4 million known.

  In this catalog of all possible missense mutations, AlphaMissense was able to classify 89% of the missense mutations. Predictions showed that 57% of them could be benign, while 32% could be pathogenic. The next study further validated the accuracy of the prediction: based on examination of a database of known pathogenic mutations, benign and pathogenic missense mutations could be classified with an accuracy of 90%.

  AlphaMissense's predictions are currently being made freely available to the scientific community. In addition to the missense mutation catalog, the DeepMind team also shared extended predictions for 216 million single amino acid substitutions in nearly 20,000 human proteins.

  The team said they are currently exploring ways to help study the genetics of rare diseases. While the predictions don't have direct clinical applications, the work may improve the diagnosis of rare diseases and help find new disease-causing genes.

  Of course, AlphaMissense will need to undergo rigorous testing before it can realize practical applications. In a news story published online by Nature, a number of scientists who were not involved in the study also offered their views on the future prospects of the tool.

  Some scholars pointed out that AlphaMissense is clearly an improvement over the tools already available for predicting the effects of missense mutations, which are the most powerful predictive tools available; however, the importance of AlphaMissense is less significant than AlphaFold's contribution to ushering in a new era of computational biology. Perhaps in two or three years' time, we can welcome a more powerful predictive tool.

  In any case, as the press release on DeepMind's website reads, we look forward to AlphaMissense joining other tools to help researchers better understand diseases and develop life-saving therapies.