Predictive tools can confuse rare mutations with dangerous ones

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: When a patient's DNA is read, it is compared with a reference version of the human genome. This allows geneticists and rare disease experts to look at a list of places where the patient's DNA differs.

Most variations will be harmless and shared with millions of other people, but some can cause illness. Computer software usually helps make that judgment. The program works through the list and scores each difference for how dangerous it appears.

Doctors can use those scores to guide treatment options. Researchers in laboratories around the world can use the knowledge to study potential mechanisms of action. A research team led by Dr.

Donate Weghorn at the Center for Genomic Regulation (CRG) in Barcelona has now tested 50 of the world's top prediction tools used for this purpose against 13.5 million mutations spread across 6,659 human genes. The work, published in the American Journal of Human Genetics, shows that almost all programs overestimate the potential harm of mutations that are less likely to occur than average and underestimate the effect of more likely mutations. The research team found this happens because almost all computer programs look at a region of DNA and ask whether that spot has stayed the same over millions of years of evolution.

If it has, the software decides it has been conserved for a reason, must matter and that changing it must be bad. That includes programs such as AlphaMissense, built by Google DeepMind, along with EVE and popEVE, co-developed by other research groups at the CRG. However, the developers of some of the programs tested expect the effect to be modest in a diagnostic setting.

In their view, it is unlikely to change the interpretation of clearly damaging variants and would more plausibly reshuffle the ranking of variants with moderate predicted effects. The new study finds the models are biased because some regions in the human genome can be much more mutation-prone than other regions. For example, Weghorn has previously shown that the starting points of genes are 35% more prone to mutations than other regions.