https://www.nature.com/articles/s41586-021-04043-8
The impact of most protein mutations on human health is unknown, and existing prediction algorithms are often unreliable (1).
To this end, Debora Marks of Harvard Medical School collaborated with Oxford University Yarin Gal and other researchers to find a new approach. Instead of relying on existing marker information on the health effects of protein mutations, directly uses machine learning algorithms to analyze how proteins respond to changes in the process of evolution. Its sequence is "constrained" to maintain its adaptability, and then the impact of various protein mutations on human health can be inferred from this "constraint" .
uses evolutionary information to assess the impact of mutations on health (1)
Researchers used this algorithm to predict the health impact of 36 million protein-coding mutations in 3,219 disease-related genes; preliminary evaluation shows that the prediction accuracy of this method exceeds existing algorithms, Comparable accuracy to functional experimental predictions for .
The accuracy of the new algorithm (EVE (evolutionary model of variant effect)) far exceeds the existing algorithm (1)
The accuracy of the new algorithm (EVE (evolutionary model of variant effect)) is comparable to the experimental data (1)
Researchers said The algorithm can independently predict the health effects of protein mutations at very large scales, well complementing strategies to find rare mutations based on large-scale exome sequencing;
In addition, the algorithm relies on the sequencing of an increasing number of species, and this work involves The sequence of approximately 140,000 species, many of which are endangered, also reflects the importance of conservation biology.
This work was published in nature(1) on October 27, 2021. All data and codes of
are available at: https://evemodel.org/, and (1) is updated in real time.
Comments:
This strategy has limited ability to analyze the synergistic or antagonistic effects of multiple mutation sites on the same protein, especially distant sites. In the future, the combined protein structure may be better.
In addition, this algorithm is difficult to analyze protein mutations other than point mutations and , such as early termination, frameshifting, and fusion; it is also a great challenge to analyze non-coding sequences.
Corresponding author introduction:
https://sysbio.med.harvard.edu/debora-marks
https://www.cs.ox.ac.uk/people/yarin.gal/website/
References:
1. J. Frazer et al., Disease variant prediction with deep generative models of evolutionary data. Nature, 1–5 (2021).
Original link:
https://www.nature.com/articles/s41586-021-04043-8
