RankVar: Machine Learning to Prioritize Variants in Rare Diseases

Published in Genome Medicine, the algorithm cross-references clinical notes and sequencing to classify candidate variants, including to reinterpret files that have remained undiagnosed.

The genetic diagnosis of rare diseases rarely affects the lack of data: the sequencing of exome or genome produces thousands of variants, only one of which explains the clinical picture. The work published in Genome Medicine offers RankVar, a machine learning algorithm designed to prioritize these candidates by combining the patient's clinical notes and sequencing profile.

The model is based on a random forest driven on about one million variants from the 1000 Genomes project, in which known pathogenic variants were injected to constitute positive examples.

The authors identify three uses. The first is the initial genetic diagnosis. The second, less obvious and probably the most useful in practice, is the reinterpretation of genomes already sequenced in unsolved dossiers: the evolving knowledge, a variant ruled out yesterday can go back in the ranking today. The third is research, with the identification of new candidate genes.