Image by Kate Patterson (Garvan Institute of Medical Research)
An in-depth study by Professor Chris Goodnow, Deputy Director of the Garvan Institute of Medical Research, has assessed how accurately we can predict the health consequences of mutations that change single letters in our genetic code – so-called “missense mutations”. His work demonstrates an important disjunction between the predicted and actual impacts of such mutations.
Within the 3 billion DNA base pairs of each of our genomes lurk thousands of missense mutations – single-letter “mis-spellings” in our DNA sequence that can affect the structure and function of proteins.
Some missense mutations cause disease, yet other changes have no apparent effect. So, being able to assess the likely clinical impact of these mutations is a major goal in 21st century medicine.
Computer programs (such as PolyPhen2 and CADD) are presently the most widely used way to assess the likely impact of a mutation. These programs leverage vast amounts of existing scientific knowledge of protein structure, variations in DNA sequence, and well-characterised disease-associated mutations. What isn’t clear, though, is how well the programs perform when tasked with determining the health impact of a previously unseen missense mutation.
Prof Goodnow and colleagues tested the predictive power of several popular computer programs. The work drew on a remarkably extensive collection of over 200,000 missense mutations in the mouse genome, which was developed by Prof Goodnow and researchers at John Curtin School of Medical Research (ANU) and University of Texas Southwestern Medical Center (funded by the US National Institutes of Health).
Prof Goodnow says, “We drew up a list of 23 critical genes for the immune system, whose function we understood in great detail in humans and mice – so we knew precisely how a loss-of-function mutation in that gene would affect the immune system in a mouse carrying it.”
“We identified 33 mice where a new missense mutation had arisen in one of the 23 genes. We bred each mouse line to homozygosity – that is, until we had mice that were carrying two copies of the mutation. Then, we tested whether each line had a measurable deficiency in immune function.”
Intriguingly, the researchers found that all the popular computer programs appeared to over-predict the harmfulness of mutations. For every five mutations predicted to be harmful by PolyPhen2, for instance, only one led to a measurable immune-system abnormality.
“At first,” Prof Goodnow explains, “we worried that the computer algorithms were just performing poorly – but further tests revealed that the programs can’t yet distinguish a severely damaged gene from one that is damaged very slightly.”
To explore the function of the programs in more detail, the researchers analysed the impact of all possible missense mutations of p53, the protein most often mutated in human cancer. They compared the actual effect of over 2000 p53 mutations with the predictions of made by the different computer programs.
They realised that the mutations being classed as harmful by the tools fell into two groups – those that had an evidently harmful effect in an individual, and those whose effect was barely measurable.
The barely measurable – or “nearly neutral” – mutations were still sufficiently damaging that, over many generations, they would still eventually be weeded out of the pool.
“What we came to see,” says Prof Goodnow, “is that the programs can pick up mutations that are damaging to an individual, but also those that have effects only on an evolutionary timescale.”
“So it’s not that the algorithms are no good – it’s that there are, in effect, two types of damaging mutation. And the challenge for us now is to learn how to distinguish between the two in a clinical setting. At the moment, the computer programs aren’t a substitute for direct experimental measurements at the laboratory bench.”
Prof Goodnow’s research is published in the leading journal Proceedings of the National Academy of Sciences of the United States of America.