Loading Curie Briefs...
Getting the latest healthcare news for you
Getting the latest healthcare news for you
Popular genetic mutation prediction tools have a hidden bias problem. A new study tested 50 leading DNA variant prediction tools — including Google DeepMind's AlphaMissense — and found nearly all of them systematically overestimate the danger of rare mutations and underestimate the risk of common ones. The culprit? They ignore the fact that some regions of the genome are naturally more mutation-prone than others.
When doctors sequence a patient's DNA, they rely on computer programs to flag which genetic variants might be harmful. But a new study published in the American Journal of Human Genetics reveals a widespread flaw in nearly all of these tools: they don't account for the fact that some parts of the genome mutate far more frequently than others.
Researchers at the Centre for Genomic Regulation (CRG) in Barcelona tested 50 top prediction tools against 13.5 million mutations across 6,659 human genes. They found that almost all programs — including Google DeepMind's AlphaMissense — overestimate the danger of mutations in low-mutation-rate regions and underestimate the risk in high-mutation-rate regions. The root cause: most tools rely on evolutionary conservation as a proxy for importance, without correcting for baseline mutation rates. The fix, according to the authors, is to incorporate genome-wide mutation rate maps into these software tools.
Key Takeaways:
Why it matters: These tools guide real-world diagnostic and treatment decisions for rare disease patients. Systematic bias — even if modest — could lead to misclassification of variants, potentially delaying or misdirecting care for vulnerable patients.
