Ketrin Gjoni

UC San Francisco

“Deep-learning prioritization of autism spectrum disorder variants that disrupt three-dimensional genome folding”

I study how variants change the structure of DNA folding and disrupt interactions between important genes and their enhancers. To explore this in autism spectrum disorder, I use a machine learning tool and epigenomic data to predict how structural variants change genomic contacts at nearby regulatory regions.


The spatial organization of DNA into its three-dimensional (3D) structure facilitates gene regulation through the topologically associated domains (TADs). The re-organization of TADs can change the frequency of genomic contacts between distal regulatory elements and cause aberrant transcription. We hypothesize that disruption of genome folding is a prevalent but overlooked mechanism of variants in autism spectrum disorder (ASD), for which 80% of the genetic causality is unknown. Experimentally testing this requires genome editing and chromatin capture experiments, which are cost-prohibitive to perform at scale across a disease cohort. To overcome this barrier, I propose to extend an existing convolutional neural network– Akita, which predicts chromatin contacts from genome sequence alone– by incorporating epigenomic data. This allows me to score variants for disrupted folding in a cell type-specific context. Using neuronal epigenetic and transcriptional annotations, I aim to prioritize candidate variants that disrupt cis-regulatory interactions of genes implicated in ASD and validate the top hits in vitro.

5 + 2 =