Intrinsic entropy model for feature selection of scRNA-seq data

Lin Li, Hui Tang, Rui Xia, Hao Dai, Rui Liu, Luonan ChenJournal of Molecular Cell Biology 2022 · DOI

TL;DR

This paper introduces the "Intrinsic Entropy" (IE) model, a novel feature selection method for single-cell RNA sequencing (scRNA-seq) data. By decomposing a gene's total fluctuation into "intrinsic" (IE) and "extrinsic" (EE) components, the IE model effectively identifies truly informative genes, overcoming the challenges posed by noise and dropout events common in scRNA-seq. This leads to more accurate cell clustering and classification, even identifying novel cell types.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity. However, scRNA-seq data is notoriously noisy, plagued by "dropout events" (where a gene is detected as absent when it's actually present at low levels) and other technical variations. This noise significantly confounds downstream analyses, particularly cell clustering and classification, which are crucial for identifying cell types and states. The accuracy of these analyses heavily relies on selecting the right set of "feature genes" – those that truly represent biological differences between cells, rather than technical artifacts. Existing methods struggle to effectively distinguish between biologically meaningful fluctuations and noise, leading to suboptimal performance.

Methods

The core innovation of this paper is the Intrinsic Entropy (IE) model. The authors propose that the total entropy (TE) of a gene's expression, representing its overall fluctuation, can be decomposed into two components:

  • Intrinsic Entropy (IE): This component reflects the gene's inherent fluctuation due to its own regulatory dynamics within a cellular process. This is considered the "signal."
  • Extrinsic Entropy (EE): This component represents the "noisy" fluctuation caused by external factors, such as interactions with other genes or technical noise. This is considered the "noise."

The model mathematically decomposes TE into IE and EE using information theory principles, specifically mutual information (MI) and conditional entropy (CE). By isolating and prioritizing genes with high IE, the model aims to select genes that are genuinely informative about cell states. The authors also mention using Principal Component Analysis (PCA) to reduce computational complexity when estimating EE.

Datasets

The authors evaluated their IE model using a combination of simulated and real-world scRNA-seq datasets:

  • Simulated Datasets: 600 simulated scRNA-seq datasets were generated using the Splatter package, based on a gamma-Poisson distribution. These datasets were used to compare the IE model against other feature selection methods in terms of clustering accuracy (Adjusted Rand Index - ARI).
  • Real Datasets:
    • 14 published scRNA-seq datasets were used for benchmarking, covering various sequencing protocols (e.g., Smart-seq2, 10x Genomics).
  • The Mammary Gland dataset (from the Tabula Muris project) was used to demonstrate the IE model's performance in clustering and cell type identification.
  • The Chu-time dataset (human progenitors) was used to showcase the model's sensitivity in distinguishing similar cell types.
  • The Tabula Muris dataset (specifically the Large-Intestine and Heart and Aorta subsets) was used for comprehensive evaluation across different clustering and classification methods.
  • A lung cancer-associated myeloid cell dataset (from Lambrechts et al., 2018) was used to identify novel cell subtypes.

Key Insights and Ideas

  • Intrinsic fluctuation is the true signal: The central idea is that a gene's inherent variability (IE) is what truly reflects its biological role and distinguishes cell types, while extrinsic variability (EE) is largely noise.
  • Noise vs. Information: The paper challenges the conventional view that all gene fluctuation is noise. It posits that intrinsic fluctuation is informative, while total fluctuation can be misleading.
  • Robustness across methods: The IE model's effectiveness isn't tied to a specific downstream analysis tool; it consistently improves performance when used with various clustering and classification algorithms.
  • Sensitivity to novel cell types: The IE model demonstrates an ability to identify previously uncharacterized cell subtypes, suggesting its utility in discovering new biological insights.

Key Results

  • Superior Clustering Performance: In simulated datasets, the IE model consistently outperformed other methods (HVG, S-E, and random gene selection) in terms of Adjusted Rand Index (ARI) when selecting a varying number of genes.
  • High Accuracy on Real Data: On 14 real scRNA-seq datasets, feature genes selected by the IE model led to significantly better clustering accuracy compared to other methods.
  • Robustness: The IE model's performance remained strong across different clustering methods (Seurat, SC3, SIMLR) and classification methods (XGBoost, DeepInsight, Random Forest, etc.) on various datasets.
  • Sensitivity: The IE model successfully distinguished between closely related cell types (e.g., H1 and H9 embryonic stem cells) where other methods failed.
  • Novel Subtype Discovery: In a lung cancer myeloid cell dataset, IE-guided clustering identified a novel subtype (M_C3_CCL20) whose expression was associated with worse patient survival, highlighting its potential for biological discovery.

Limitations and Future Work

  • Computational Complexity: While PCA is used to mitigate this, the theoretical framework of IE and EE can be computationally intensive, especially for very large datasets.
  • Future Work: The authors suggest extending the IE model by incorporating network features and dynamic features of informative genes to identify network biomarkers for disease diagnosis, prognosis, and dynamic network biomarker detection. They also highlight the potential for using IE in identifying novel cell types with high biological significance.

Summary updated

Abstract

Recent advances of single-cell RNA sequencing (scRNA-seq) technologies have led to extensive study of cellular heterogeneity and cell-to-cell variation. However, the high frequency of dropout events and noise in scRNA-seq data confounds the accuracy of the downstream analysis, i.e. clustering analysis, whose accuracy depends heavily on the selected feature genes. Here, by deriving an entropy decomposition formula, we propose a feature selection method, i.e. an intrinsic entropy (IE) model, to identify the informative genes for accurately clustering analysis. Specifically, by eliminating the 'noisy' fluctuation or extrinsic entropy (EE), we extract the IE of each gene from the total entropy (TE), i.e. TE = IE + EE. We show that the IE of each gene actually reflects the regulatory fluctuation of this gene in a cellular process, and thus high-IE genes provide rich information on cell type or state analysis. To validate the performance of the high-IE genes, we conduct computational analysis on both simulated datasets and real single-cell datasets by comparing with other representative methods. The results show that our IE model is not only broadly applicable and robust for different clustering and classification methods, but also sensitive for novel cell types. Our results also demonstrate that the intrinsic entropy/fluctuation of a gene serves as information rather than noise in contrast to its total entropy/fluctuation.
BibTeX
@article{li2022intrinsic,
  author = {Lin Li and Hui Tang and Rui Xia and Hao Dai and Rui Liu and Luonan Chen},
  title = {{Intrinsic entropy model for feature selection of scRNA-seq data}},
  journal = {Journal of Molecular Cell Biology},
  year = {2022},
  doi = {10.1093/jmcb/mjac008},
}
Figure