QClus: A droplet-filtering algorithm for enhanced snRNA-seq data quality in challenging samples

Eloi Schmauch, Johannes Ojanen, Kyriakitsa Galani, Juho Jalkanen, Kristiina Harju, Maija Hollmén, Hannu Kokki, Jarmo Gunn, Jari Halonen, Juha Hartikainen, Tuomas Kiviniemi, Pasi Tavi, Minna U. Kaikkonen, Manolis Kellis, Suvi Linna-KuosmanenbioRxiv 2022 · DOI

TL;DR

QClus is a new unsupervised filtering algorithm specifically designed to handle the high levels of cytoplasmic contamination common in human heart single-nucleus RNA sequencing (snRNA-seq). By using tissue-specific marker genes and multi-metric k-means clustering, it outperforms standard "one-size-fits-all" tools, preserving rare cell types while effectively removing empty and contaminated droplets.

What Problem Does This Paper Address?

Single-nucleus RNA sequencing (snRNA-seq) is a powerful tool for studying complex tissues like the heart, but it has a "background noise" problem. When researchers break apart heart tissue to isolate nuclei, the surrounding cytoplasm—which is packed with RNA and mitochondria—often leaks into the droplets.

This results in "empty" droplets containing only ambient RNA or "contaminated" nuclei that look like a mix of different cell types. In heart tissue, this is particularly difficult because cardiomyocytes (heart muscle cells) are massive and full of RNA, creating a massive "cloud" of contamination. Standard QC methods often fail here, either leaving too much noise (masking real biological signals) or being too aggressive (accidentally deleting entire cell populations like fibroblasts or macrophages).

Methods

The authors developed QClus, a five-step pipeline that moves beyond simple thresholding:

  1. Initial Processing: Uses standard tools (Cell Ranger) for a first pass.
  2. Metric Calculation: Instead of just looking at gene counts, QClus calculates six specific metrics, including the fraction of unspliced transcripts (higher in nuclei), mitochondrial content, and—crucially—tissue-specific marker fractions (comparing cardiomyocyte vs. non-cardiomyocyte signals).
  3. K-means Clustering: It uses these six metrics to group droplets into four clusters: high-quality non-cardiomyocytes, high-quality cardiomyocytes, highly contaminated nuclei, and empty droplets.
  4. Outlier Filtering: A distribution-based step to catch remaining low-quality outliers.
  5. Doublet Removal: Integration with the Scrublet algorithm to remove droplets containing two nuclei.

The "secret sauce" is the use of tissue-specific markers as a quality metric, allowing the algorithm to distinguish between a "naturally" RNA-rich cardiomyocyte and a contaminated non-muscle cell.

Datasets

The method was developed and validated using a primary dataset of 32 human heart samples. To prove its robustness, the authors benchmarked QClus against six established methods:

  • EmptyNN, DropletQC, and SampleQC (focused on empty droplet removal).
  • DecontX (a Bayesian decontamination tool).
  • DIEM (a multinomial distribution model for debris).
  • Standard Mitochondrial Thresholding.

Key Insights and Ideas

  • The "Star Shape" Warning: In dimension reduction plots (UMAP), contaminated data often forms a "star shape" where all cell types seem to bleed into a central hub of noise. QClus is specifically designed to collapse this hub.
  • Tissue Specificity Matters: The authors argue that "universal" preprocessing is a myth for challenging tissues. Because heart contamination is dominated by cardiomyocyte debris, a filter that doesn't "know" what a cardiomyocyte looks like will inevitably fail.
  • Clustering vs. Thresholding: Hard cutoffs (e.g., "delete everything with >5% mitochondria") are too blunt. By using k-means clustering on multiple metrics, QClus identifies the signature of a bad droplet rather than just a high number.

Key Results

  • Superior Cell Type Recovery: In head-to-head tests, QClus successfully identified 11 major cell types in samples where competing methods like DIEM only found 3.
  • Avoiding Over-filtering: While methods like DecontX and mitochondrial thresholding often deleted entire populations of fibroblasts and macrophages, QClus retained these populations while still reducing background noise.
  • Robustness: Across all 32 samples, QClus provided consistent results without requiring manual, sample-by-sample parameter tuning—a major win for large-scale studies.
  • Visual Clarity: Post-QClus UMAPs showed clear, distinct clusters for different cell types, effectively removing the "star center" of contamination seen in earlier processing steps.

Limitations and Future Work

The authors acknowledge a few areas for growth:

  • Cell-Type Agnostic Outliers: The final outlier and doublet removal steps are currently agnostic of cell type; the authors suggest a "finer approach" that treats different cell types specifically in these final stages could improve results further.
  • Filtering vs. Decontaminating: QClus is a filtering tool (it removes bad nuclei). It does not currently clean the ambient RNA out of the nuclei it decides to keep. Combining QClus with a decontamination tool (like SoupX or DecontX) could provide the "cleanest" possible data.
  • Adaptability: While optimized for the heart, the framework can be adapted to other tissues (like skeletal muscle) by simply swapping out the marker gene lists used in the metric calculation.

Summary updated

Abstract

Abstract Single nuclei RNA sequencing (snRNA-seq) remains a challenge for many human tissues, as incomplete removal of background signal masks cell-type-specific signals and interferes with downstream analyses. Here, we present QClus, a droplet-filtering algorithm targeted toward challenging samples, using cardiac tissue as an example. QClus uses specific metrics such as cell-type-specific marker gene expression to cluster nuclei and filter empty and highly contaminated droplets, providing reliable cleaning of samples with varying number of nuclei and contamination levels. In a benchmarking analysis against seven alternative methods across six datasets consisting of 252 samples and over 1.9 million nuclei, QClus achieved the highest quality in the greatest number of samples over all evaluated quality metrics and recorded no processing failures, while robustly retaining numbers of nuclei within the expected range. QClus combines high quality, automation, and robustness with flexibility and user-adjustability, catering to diverse experimental needs and datasets.
BibTeX
@misc{schmauch2022qclus,
  author = {Eloi Schmauch and Johannes Ojanen and Kyriakitsa Galani and Juho Jalkanen and Kristiina Harju and Maija Hollmén and Hannu Kokki and Jarmo Gunn and Jari Halonen and Juha Hartikainen and Tuomas Kiviniemi and Pasi Tavi and Minna U. Kaikkonen and Manolis Kellis and Suvi Linna-Kuosmanen},
  title = {{QClus: A droplet-filtering algorithm for enhanced snRNA-seq data quality in challenging samples}},
  year = {2022},
  doi = {10.1101/2022.10.21.513315},
  howpublished = {bioRxiv},
}
Figure