Mapping robust multiscale communities in chromosome contact networks

TL;DR

This paper introduces a method to analyze the stability of multiscale community structures in chromosome contact networks derived from Hi-C data, revealing that some scales yield more reliable biological insights than others.

Contribution

The authors developed a novel approach to map the solution landscape of community partitions in dense Hi-C networks, enabling identification of robust multiscale communities.

Findings

Certain network scales exhibit more stable community structures.

Robust communities can differ significantly across scales.

Method facilitates better interpretation of 3D genome organization.

Abstract

To better understand DNA's 3D folding in cell nuclei, researchers developed chromosome capture methods such as Hi-C that measure the contact frequencies between all DNA segment pairs across the genome. As Hi-C data sets often are massive, it is common to use bioinformatics methods to group DNA segments into 3D regions with correlated contact patterns, such as Topologically Associated Domains (TADs) and A/B compartments. Recently, another research direction emerged that treats the Hi-C data as a network of 3D contacts. In this representation, one can use community detection algorithms from complex network theory that group nodes into tightly connected mesoscale communities. However, because Hi-C networks are so densely connected, several node partitions may represent feasible solutions to the community detection problem but are indistinguishable unless including other data. Because this limitation is a fundamental property of the network, this problem persists regardless of the community-finding or data-clustering method. To help remedy this problem, we developed a method that charts the solution landscape of network partitions in Hi-C data from human cells. Our approach allows us to scan seamlessly through the scales of the network and determine regimes where we can expect reliable community structures. We find that some scales are more robust than others and that strong clusters may differ significantly. Our work highlights that finding a robust community structure hinges on thoughtful algorithm design or method cross-evaluation.