Zero-inflated Smoothing Spline (ZISS) Models for Individual-level Single-cell Temporal Data

Yifu Tang, Yi Zhang, Yue Wang, Jingyi Zhang, Xiaoxiao SunarXiv 2024 · DOI · arXiv

Abstract

Recent advancements in single-cell RNA-sequencing (scRNA-seq) have enhanced our understanding of cell heterogeneity at a high resolution. With the ability to sequence over 10,000 cells per hour, researchers can collect large scRNA-seq datasets for different participants, offering an opportunity to study the temporal progression of individual-level single-cell data. However, the presence of excessive zeros, a common issue in scRNA-seq, significantly impacts regression/association analysis, potentially leading to biased estimates in downstream analysis. Addressing these challenges, we introduce the Zero Inflated Smoothing Spline (ZISS) method, specifically designed to model single-cell temporal data. The ZISS method encompasses two components for modeling gene expression patterns over time and handling excessive zeros. Our approach employs the smoothing spline ANOVA model, providing robust estimates of mean functions and zero probabilities for irregularly observed single-cell temporal data compared to existing methods in our simulation studies and real data analysis.

Introduction

In single-cell genomics, researchers utilize "pseudotime" to measure a cell’s progress through biological processes like differentiation. However, scRNA-seq data is notoriously noisy due to zero-inflation (or "dropout"). This phenomenon occurs when a gene appears to have zero expression because sequencing technology failed to capture its mRNA, rather than the gene being biologically silent.

Existing tools, such as Zero-Inflated Generalized Additive Models (ZIGAMs), often struggle with "knot selection," requiring researchers to manually determine the flexibility of the regression curves. When data is irregularly spaced or heavily zero-inflated, these models frequently produce biased estimates or fail to converge.

Methodology

The Zero-inflated Smoothing Spline (ZISS) method introduces a two-component mixture model to separate biological signals from technical noise:

  1. The Mean Function (μ): ZISS utilizes a smoothing spline ANOVA model to represent true gene expression. This "knot-free" approach uses a penalty functional to ensure smoothness, making it highly robust for irregularly sampled data.
  2. The Dropout Probability (p): To account for the fact that dropout rates vary over time, ZISS employs B-spline basis functions to model the likelihood of observing a zero at any given pseudotime point.

These components are integrated using an Expectation-Maximization (EM) algorithm, which iteratively estimates whether a zero is biological or technical while simultaneously updating the expression curve and dropout model.

Experimental Results

The researchers validated ZISS using both synthetic simulations and real-world data from the Human Kidney Cell Atlas (comprising ~19,500 epithelial cells from 27 patients).

Comparison of ZISS and NZSS (Non-Zero Smoothing Spline) performance. Panel (a) shows how ZISS (red) accurately follows the data trend compared to the biased NZSS (blue). Panel (b) illustrates the high frequency of zeros across pseudotime bins.

Key Findings

  • Superior Accuracy and Stability: In high-sparsity simulations, ZISS achieved a Mean Square Error (MSE) of 0.169, significantly outperforming ZIGAM (0.368) and standard non-zero fits (0.550). It also demonstrated the lowest standard deviation across repeated experiments, indicating high reliability.
  • Dynamic Dropout Modeling: The study demonstrated that zero-inflation probabilities are not constant across pseudotime. ZISS successfully captures these dynamic rates, whereas models assuming fixed dropout produce biased results.
  • Knot-Free Robustness: By moving away from manual knot selection, ZISS effectively handles "gaps" in pseudotime where few cells are captured, maintaining biologically plausible smooth curves where other methods produce erratic jumps.

Conclusion and Future Work

ZISS represents a significant advancement in modeling single-cell temporal data by effectively distinguishing between biological and technical zeros. While currently optimized for 1D pseudotime, the authors suggest the framework could be extended to spatial transcriptomics (2D/3D space). Future iterations may also explore Negative Binomial distributions to better account for the over-dispersion often found in biological systems.

Summary updated

Abstract

Recent advancements in single-cell RNA-sequencing (scRNA-seq) have enhanced our understanding of cell heterogeneity at a high resolution. With the ability to sequence over 10,000 cells per hour, researchers can collect large scRNA-seq datasets for different participants, offering an opportunity to study the temporal progression of individual-level single-cell data. However, the presence of excessive zeros, a common issue in scRNA-seq, significantly impacts regression/association analysis, potentially leading to biased estimates in downstream analysis. Addressing these challenges, we introduce the Zero Inflated Smoothing Spline (ZISS) method, specifically designed to model single-cell temporal data. The ZISS method encompasses two components for modeling gene expression patterns over time and handling excessive zeros. Our approach employs the smoothing spline ANOVA model, providing robust estimates of mean functions and zero probabilities for irregularly observed single-cell temporal data compared to existing methods in our simulation studies and real data analysis.
BibTeX
@misc{tang2024zero,
  author = {Yifu Tang and Yi Zhang and Yue Wang and Jingyi Zhang and Xiaoxiao Sun},
  title = {{Zero-inflated Smoothing Spline (ZISS) Models for Individual-level Single-cell Temporal Data}},
  year = {2024},
  eprint = {2401.15309},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.15309},
}
Figure