vinay-swamy / distill Goto Github PK

Distill is a variant pathogenicity classifier for rare mendelian disease (no there is no crazy acronyn involved).

Random Forest + xgboost + Deep LSTM - based pathogenicity classifier of human genetic variation. Uses curated set of ClinVar, solved UK10K eye mendelian disorders, and gnomAD DNA variants with a broad set of in silico, population, gene conseequence, and conservation based metrics and annotations to build highly specific variant classification score.

There are a few minor and one major main distinctions in how this algorithm is built and assessed compared to popular pathogenic scoring systems like REVEL, CADD, and FATHMM.

Minor differences:

• Deep set of putative non-pathogenic mutations, including real WGS from rare mendelian disease
• Rich set of annotations ranging from the scores like CADD and FATHMM, to non-coding annotations like LINSIGHT, to in silico consequences from VEP, and to population allelle frequency metrics.
• Ensembl model from three high performing frameworks (Random Forest, xgboost, Keras)

Major difference:

• Training of algorithms under more realistic situations that genetic clinics will face for mendelian disorders for singletons and trios.
  • 250:1 not_pathogenic:pathogenic ratio