Predicting expression-altering promoter mutations with deep learning

Area of Science:

• Genetics
• Bioinformatics
• Genomics

Background:

• Exome sequencing diagnoses only a minority of rare genetic diseases.
• Unrecognized pathogenic variants may exist in noncoding DNA sequences.

Purpose of the Study:

• To develop a deep neural network, PromoterAI, for identifying noncoding promoter variants that dysregulate gene expression.
• To assess the impact of these variants on gene expression and their role in rare diseases.

Main Methods:

• Developed PromoterAI, a deep neural network for predicting expression-altering promoter variants.
• Analyzed RNA and protein expression data from thousands of individuals.
• Investigated negative selection acting on these variants.
• Validated functional impact using reporter assays.

Main Results:

• PromoterAI accurately identifies noncoding promoter variants.
• Variants with predicted expression-altering consequences show outlier expression levels.
• These variants are under strong negative selection in human populations.
• Clinically relevant genes in rare disease patients are enriched for these variants.

Conclusions:

• Noncoding promoter variants significantly contribute to rare genetic diseases.
• PromoterAI is a valuable tool for diagnosing rare genetic disorders.
• Promoter variation accounts for an estimated 6% of the genetic burden in rare diseases.