Topological Data Analysis of Single-cell Hi-C Contact Maps

Mathieu Carrière, Raul RabadanarXiv 2018 · OpenReview

Abstract

In this article, we show how the recent statistical techniques developed in Topological Data Analysis for the Mapper algorithm can be extended and leveraged to formally define and statistically quantify the presence of topological structures coming from biological phenomena in datasets of CCC contact maps.

Introduction

Chromosomes are dynamic structures whose three-dimensional organization shifts during critical biological processes such as cell division and differentiation. High-throughput Chromosome Conformation Capture (Hi-C) techniques allow researchers to generate "contact maps"—matrices representing the spatial proximity of genomic regions. However, single-cell Hi-C data presents significant analytical challenges: it is high-dimensional, extremely sparse, and subject to technical biases. While patterns like circularity in dimensionality-reduced data have suggested cell-cycle progression, these observations often lacked formal mathematical definitions or statistical validation. This work introduces a rigorous framework using Topological Data Analysis (TDA) to bridge this gap, providing tools to formally define and statistically verify these biological "shapes."

Methods: TDA and the Mapper Algorithm

The core of this approach is the Mapper algorithm, a TDA tool that summarizes high-dimensional data as a simplified graph (a simplicial complex) capturing the dataset's underlying topology.

The CCC Pipeline

The generation of contact maps involves several biochemical steps to capture DNA interactions in situ. The process begins with cross-linking to preserve spatial proximity.

Initial cross-linking step of the CCC procedure.

Following cross-linking, the DNA is fragmented, ligated to join interacting segments, and then the cross-links are reversed to release the DNA for sequencing.

Subsequent steps in the CCC procedure: fragmentation, ligation, and reverse cross-linking.

Analytical Framework

To apply Mapper effectively to Hi-C data, the authors implemented several key innovations:

  • Stratum-Adjusted Correlation Coefficient (SCC): Unlike standard Euclidean distance, SCC accounts for the distance-dependent interaction frequency inherent in genomic data, providing a more biological distance metric.
  • Filter Functions (Lenses): The dataset is viewed through "lenses"—specifically the first two eigenfunctions—which guide the slicing and clustering process of the Mapper algorithm.
  • Extended Persistence Diagrams: These diagrams quantify the significance of topological features (e.g., loops). Features farther from the diagonal in a 2D persistence plane are considered more significant.
  • Multivariate Bootstrapping: To provide statistical guarantees, the authors extended bootstrapping to the multivariate Mapper. By resampling the data 100 times and measuring bottleneck distances, they can establish confidence intervals for identified topological features.

Datasets

The framework was validated using a landmark dataset of 1,171 single mouse embryonic stem (ES) cells. Contact matrices were reconstructed at a 500kb resolution, providing a benchmark for studying cell-cycle dynamics.

Key Insights: Topology as a Robust Biological Summary

TDA offers a "coordinate-free" summary of data, making it robust to the noise and small deformations typical of biological variation. A primary discovery was the identification of a large loop in the Mapper graph. By coloring the nodes with independent biological markers—such as the "repli-score" (DNA replication timing) and the "mitotic band" (ratio of long-range interactions)—the authors observed a continuous flow around the loop. This confirms that the abstract topological feature directly represents the physical progression of the cell cycle.

Results and Statistical Validation

The Mapper algorithm successfully reconstructed the circular progression of the cell cycle from raw contact maps. Beyond visual identification, the multivariate bootstrap extension allowed for the calculation of a 93% confidence level for the loop feature. This statistical rigor ensures that the observed circularity is a genuine biological signal rather than a discretization artifact or noise.

Conclusion and Future Work

This study demonstrates that TDA, combined with robust statistical methods, can provide formal guarantees for biological structures in sparse genomic data. While this work focused on 1D loops, future research will explore higher-dimensional features like "voids" or "cavities" in more complex datasets. Additionally, while the multivariate bootstrap works well empirically, further theoretical study is required to fully characterize its statistical properties. The application of this framework to other processes, such as cell differentiation, remains a promising avenue for uncovering hidden biological drivers.

Summary updated

Abstract

In this article, we show how the recent statistical techniques developed in Topological Data Analysis for the Mapper algorithm can be extended and leveraged to formally define and statistically quantify the presence of topological structures coming from biological phenomena in datasets of CCC contact maps.
BibTeX
@misc{carriere2018topological,
  author = {Mathieu Carrière and Raul Rabadan},
  title = {{Topological Data Analysis of Single-cell Hi-C Contact Maps}},
  year = {2018},
  eprint = {1812.01360},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.01360},
}
Figure