[Reading Science] Identified 89% of Genes Causing Diseases

Google DeepMind Announces AlphaMissense Development on 19th
Incorporates Machine Learning and Generative AI Features Based on AlphaFold Framework
Confirms 89% of Disease-Related DNA Variants in Human Body

An era has opened where artificial intelligence (AI) can pinpoint gene mutations that cause diseases and enable treatment. Last year, Google DeepMind, which identified and released about 200 million protein structures through AlphaFold, has now developed an AI capable of confirming the disease-causing relevance of human gene mutations with high probability.

On the 19th (local time), Google DeepMind published a paper in the international journal Science announcing the development of this innovative AI network called AlphaMissense.

A significant number of diseases experienced by humans are often caused by genetic factors. DNA consists of four bases?adenine (A), cytosine (C), guanine (G), and thymine (T)?twisted in a specific sequence to perform certain functions. However, if one or two bases are missing or the order is reversed, that is, if a mutation occurs, it can cause disease. Examples include cystic fibrosis, a common genetic disorder among Caucasians, and sickle cell anemia, frequently occurring in Black populations. It is estimated that there are over 70 million mutations causing such genetic diseases, but scientists have only identified a few million so far. Accordingly, various computational tools exist worldwide to predict whether specific gene mutations cause diseases.

Google DeepMind’s AlphaMissense adds machine learning methods to the methodologies of existing tools, resulting in outstanding capabilities. While previous tools identified disease-causing potential in only about 0.1% of all human gene mutations, AlphaMissense has dramatically increased this rate to 89%, according to the research team. The international journal Nature reported, "AlphaMissense appears to have superior ability compared to other tools in distinguishing disease-causing gene mutations from benign ones," and "it performed very well in experiments measuring thousands of gene mutations at once to identify problematic sites." In fact, Google DeepMind’s research team used AlphaMissense to calculate all possible missense mutations in the human genome and found that about 57% are unrelated to disease, while approximately 32% could cause disease.

Pushmeet Kohli, Vice President of Research at Google DeepMind, explained, "AlphaMissense leveraged AlphaFold’s protein structure prediction capabilities to identify locations within gene proteins where disease-causing mutations occur."

Google DeepMind utilized not only the AlphaFold network but also characteristics of generative AI such as ChatGPT and large language models (LLMs) in developing AlphaMissense. Just as ChatGPT learns countless sentences used by humans to predict the next word following a specific term, AlphaMissense was trained through learning millions of protein sequences to detect mutations with a high likelihood of causing disease. This so-called protein language model has proven to be highly effective in predicting protein structures or designing new proteins. Jiga Absek, a researcher at Google DeepMind, told Nature, "AlphaMissense is proficient in various predictions because it has learned which protein sequences are valid and which are not."