Structure to Function in Single Cells: Predicting Transcription from 3D Cromatin Conformation with Graph Neutral Networks by Laura Gunsalus

Laura Gunsalus

UC San Francisco | Boeing/de Baubigny

I develop and interrogate deep learning models to better understand gene regulatory mechanisms. My current work investigates the relationship between 3D genome organization and gene expression.


The 3D conformation of chromatin, from local loops in DNA to higher-order structures, influences how and when genes are expressed in each cell. The relationship between genome structure and transcription is not yet clear, and few computational techniques exist to represent and model genomes in single cells. Is 3D chromatin conformation predictive of transcription? How can we best represent single cell genomic coordinates for computational analysis? Using matched 3D chromatin position, RNA expression, and distance to nuclear features from multiplexed imaging datasets, we explore the relationship between genome structure and function. We propose representing single cell genomes as fully connected graphs and develop a graph convolutional neural network to predict transcription from genome structure. We find mild predictive relationships between transcription, distance to nuclear speckles, distance to the nuclear lamina, and 3D genome position, highlighting a promising path to disentangle the relationship between genome folding and gene expression.


6 + 13 =