Genetic prediction tools show bias in mutation risk assessment

A study led by Dr. Donate Weghorn at the Centre for Genomic Regulation (CRG) in Barcelona has identified biases in 50 of the world’s leading computational tools used to predict the danger of genetic mutations. The research, published in the American Journal of Human Genetics, analyzed 13.5 million mutations across 6,659 human genes.

The findings indicate that most programs tend to overestimate the harm of mutations that occur less frequently than average and underestimate the impact of more common mutations. This bias stems from the software’s reliance on evolutionary conservation; programs often assume that if a DNA region has remained unchanged over millions of years, any mutation in that area must be harmful. However, the study notes that certain regions of the human genome are naturally more prone to mutation, which can skew these predictions.

While the study highlights these inaccuracies, developers of some tools, including Google DeepMind’s AlphaMissense, suggest the impact on clinical diagnostics may be modest. They argue that while the bias might reshuffle the ranking of variants with moderate predicted effects, it is unlikely to change the interpretation of mutations that are clearly damaging.