Abstract
Purpose :
Non-coding genetic variants in cis-regulatory elements can contribute to the development of retinal diseases. However, predicting the functional impact of such variants remains challenging. In order to reduce barriers for identifying variants of interest, we implemented machine learning using single cell epigenomic data from the human retina to systematically quantify cis-regulatory variant impact scores.
Methods :
Using single nucleus ATAC-seq on developing and adult human retina, we developed 18 distinct epigenomic training sets with which to train machine learning (ML) models on putative cis-regulatory elements. With the trained models, we predicted the impact of a wide variety of non-coding variants, including all possible single nucleotide variants, and variants identified in patients with inherited retinal diseases. We further validated these methods through massively parallel reporter assays (MPRA), comparing putative CRE variant activities on a minimal promoter in mouse retinal cells.
Results :
We generated machine learning models for 18 distinct retinal cell types and developmental states that accurately and specifically identify cis-regulatory sequence motifs unique to each cell class with over 90% accuracy. Variant impact scores for each model demonstrated cell class specificity and could be used to identify sequence motifs within regulatory elements. MPRA comparisons to ML predictions demonstrate that models are able to correctly identify sequence features.
Conclusions :
The ML models generated in this study demonstrate the capacity for single nucleus epigenomic data to be used for the prediction of non-coding sequence variant impacts. Derived models demonstrate cell class specificity, and can predict class-specific sequence motifs and impact scores for variants in crucial sequences. These models yield the ability to quickly screen through identified variants in patient data to prioritize the most potentially impactful variants in future analyses.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.