Inherited DNA variants change how open the genome is, cell type by cell type

Small inherited differences in DNA can change how open stretches of the genome are inside individual cells, and those effects depend on what kind of cell it is and what state it is in. That is the conclusion of a study published in Nature by researchers at the Garvan Institute of Medical Research, the University of Sydney, UNSW Sydney and the US company Illumina. The authors link these effects to immune function and disease risk. They also report that adding the genetic information to an artificial-intelligence model improves its ability to map how genes are controlled.

Inside the nucleus, DNA is wound around proteins into a material called chromatin. Some stretches stay tightly coiled. Others loosen, which lets regulatory proteins reach the DNA. Researchers measure this 'chromatin accessibility' across the genome, and the open regions show up in the data as peaks. A chromatin accessibility quantitative trait locus, or caQTL, is a position where a common genetic variant is associated with how open a region is. Its better-known counterpart, the expression QTL or eQTL, is a variant associated with how strongly a gene is expressed. Measuring accessibility one cell at a time, rather than averaging over whole tissue samples, makes it possible to ask whether a variant matters in some cells and not in others. The authors conclude that it does: genetic effects on chromatin are specific to cell type and to cell state. The abstract refers to T cells, the white blood cells that coordinate and carry out much of the immune response.

A persistent problem in genomics is working out which open region controls which gene. So-called multiome experiments measure both chromatin accessibility and gene activity. In unpaired datasets, however, the two measurements come from different cells, so the connections between them, known as peak-gene links, have to be inferred rather than observed. The team trained a graph neural network on both the eQTL and caQTL layers. A graph neural network is a machine-learning model built to learn from data organised as networks of connected points. According to the study, the model inferred peak-gene links in unpaired multiome data with up to 17.2% higher accuracy than an analysis based on data that left the genetic layer out.

The improvement carried through to a second task. Transcription factors are proteins that bind DNA and turn particular genes on or off, and identifying their targets is central to understanding how cells are controlled. With the genetic layers included, the analysis uncovered 128 additional transcription factor-target pairs, which the authors describe as a 22% increase. The approach treats inherited variation as more than a source of disease risk. It also becomes extra evidence for reconstructing the wiring of gene regulation.

Several questions remain open. The accuracy gain of up to 17.2% is a best case rather than a typical figure. The new transcription factor-target pairs are a model's inferences, and each would need its own confirmation. The authors tie cell-specific genetic effects on chromatin to disease risk. The natural next questions are which conditions are most affected and how much of their inherited risk runs through particular immune-cell states.