THUNDER: A reference-free deconvolution method to infer cell type proportions from bulk Hi-C data

Bryce Rowland, Ruth Huh, Zoe Hou, Jia Wen, Yin Shen, Ming Hu, Paola Giusti‐Rodríguez, Patrick F. Sullivan, Yun LibioRxiv 2020 · DOI

Abstract

Hi-C data provide population averaged estimates of three-dimensional chromatin contacts across cell types and states in bulk samples. Effective analysis of Hi-C data entails controlling for the potential confounding factor of differential cell type proportions across heterogeneous bulk samples. We propose a novel unsupervised deconvolution method for inferring cell type composition from bulk Hi-C data, the Two-step Hi-c UNsupervised DEconvolution appRoach (THUNDER). We conducted extensive simulations to test THUNDER based on combining two published single-cell Hi-C (scHi-C) datasets. THUNDER more accurately estimates the underlying cell type proportions compared to supervised and unsupervised methods (e.g., MuSiC, TOAST, and NMF). We further demonstrate the practical utility of THUNDER to estimate cell type proportions and identify cell-type-specific interactions in Hi-C data from adult human cortex tissue samples. THUNDER will be a useful tool in adjusting for varying cell type composition in population samples, facilitating valid and more powerful downstream analysis such as differential chromatin organization studies. Additionally, THUNDER estimated contact profiles provide a useful exploratory framework to investigate cell-type-specificity of the chromatin interactome while experimental data is still rare.

Introduction

Bulk Hi-C sequencing provides a "population average" of how DNA is packed and organized within a sample. However, biological tissues are rarely homogeneous; a brain sample, for instance, is a complex mixture of neurons, astrocytes, and microglia. If the proportions of these cells vary between individuals, it can lead to "Simpson’s Paradox," where researchers might mistake a change in cell-type composition for a meaningful change in chromatin organization. To conduct accurate 3D-chromatin-interactome wide association studies (3WAS), researchers must control for these proportions. The challenge is that Hi-C data is high-dimensional and high-quality, cell-type-specific reference profiles are currently scarce and expensive to produce.

Methods

The authors introduce THUNDER (Two-step Hi-c UNsupervised DEconvolution appRoach). Its architecture relies on Non-negative Matrix Factorization (NMF), which decomposes a mixture matrix ($V$) into a profile matrix ($W$) and a proportion matrix ($H$).

The THUNDER workflow: (a) initial NMF on bulk Hi-C matrices, (b) selection of informative bin-pairs, and (c) refined deconvolution to estimate cell-type proportions and profiles.

The "Two-Step" innovation is key to THUNDER's performance:

  1. Initial Deconvolution: An initial NMF is performed on the entire dataset to generate rough cell-type profiles.
  2. Feature Selection: THUNDER analyzes these profiles to identify "informative bin-pairs"—genomic locations that show high cell-type specificity or high variation across inferred types.
  3. Refined Deconvolution: The algorithm subsets the original bulk data to include only these informative bin-pairs and runs NMF a second time. This removes noise from stable genomic regions and focuses the mathematical "attention" on the regions that actually differentiate cell types.

Unlike previous attempts, THUNDER is designed to integrate both intrachromosomal (same chromosome) and interchromosomal (different chromosomes) contact information simultaneously.

Key Results and Insights

THUNDER was evaluated using simulated bulk Hi-C (created from single-cell data), various cell lines, and real-world human cortex samples. Several key findings emerged:

  • Superior Accuracy: In simulations with varying cell counts, THUNDER significantly reduced the Mean Absolute Deviation (MAD) from the ground truth compared to existing methods. For example, it reduced error by 57% compared to CIBERSORT and 48% compared to standard NMF without feature selection. It also consistently outperformed 3CDE across multiple simulation scenarios.
Comparison of Mean Absolute Deviation (MAD) between THUNDER and 3CDE across various simulation scenarios, demonstrating THUNDER's consistently lower error rates.
  • The Power of Sparsity: THUNDER typically selects only ~1-2% of total bin-pairs as "informative." Focusing on this tiny, highly variable subset leads to much more accurate proportion estimates than using the entire genome.
  • Interchromosomal Contacts Matter: While often dismissed as noise, interchromosomal contacts proved highly informative for deconvolution, particularly in cancer cell lines where translocations create unique signatures.
  • Computational Efficiency: Despite the massive size of Hi-C data, THUNDER is parallelizable. At 10Kb resolution, it averaged 3.4 hours of computing time per chromosome, making it viable for genome-wide studies.

Limitations and Future Work

While THUNDER represents a significant step forward, some limitations remain. Due to the sparsity of current single-cell Hi-C data, simulations were limited to a coarse 10Mb resolution; further validation at higher resolutions (like 10Kb) is needed as ground-truth data improves. Additionally, the algorithm requires users to specify the number of cell types ($k$) beforehand and relies on biological intuition to label the resulting unsupervised clusters (e.g., identifying which cluster represents neurons vs. glia).

Future adaptations of THUNDER could extend to other spatial data types like HiChIP or PLAC-seq, which offer higher resolution at lower sequencing costs, further refining our ability to map the 3D chromatin landscape of specific cell types within complex tissues.

Source: THUNDER: A reference-free deconvolution method to infer cell type proportions from bulk Hi-C data (Rowland et al., 2020). Link to Paper

Summary updated

Abstract

Abstract Hi-C data provide population averaged estimates of three-dimensional chromatin contacts across cell types and states in bulk samples. Effective analysis of Hi-C data entails controlling for the potential confounding factor of differential cell type proportions across heterogeneous bulk samples. We propose a novel unsupervised deconvolution method for inferring cell type composition from bulk Hi-C data, the Two-step Hi-c UNsupervised DEconvolution appRoach (THUNDER). We conducted extensive simulations to test THUNDER based on combining two published single-cell Hi-C (scHi-C) datasets. THUNDER more accurately estimates the underlying cell type proportions compared to supervised and unsupervised methods (e.g., MuSiC, TOAST, and NMF). We further demonstrate the practical utility of THUNDER to estimate cell type proportions and identify cell-type-specific interactions in Hi-C data from adult human cortex tissue samples. THUNDER will be a useful tool in adjusting for varying cell type composition in population samples, facilitating valid and more powerful downstream analysis such as differential chromatin organization studies. Additionally, THUNDER estimated contact profiles provide a useful exploratory framework to investigate cell-type-specificity of the chromatin interactome while experimental data is still rare.
BibTeX
@misc{rowland2020thunder,
  author = {Bryce Rowland and Ruth Huh and Zoe Hou and Jia Wen and Yin Shen and Ming Hu and Paola Giusti‐Rodríguez and Patrick F. Sullivan and Yun Li},
  title = {{THUNDER: A reference-free deconvolution method to infer cell type proportions from bulk Hi-C data}},
  year = {2020},
  doi = {10.1101/2020.11.12.379941},
  howpublished = {bioRxiv},
}
Figure