Non-linear Archetypal Analysis of Single-cell RNA-seq Data by Deep Autoencoders

Yuge Wang, Hongyu Zhao

TL;DR

This paper introduces scAAnet, a novel deep autoencoder framework for non-linear archetypal analysis of single-cell RNA sequencing (scRNA-seq) data. scAAnet effectively identifies underlying gene expression programs (GEPs) and their activity across cells, overcoming limitations of traditional clustering methods by capturing continuous cellular states and shared GEPs.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity. While clustering methods are adept at identifying distinct cell types, they struggle to capture the continuous spectrum of cellular states and the underlying gene expression programs (GEPs) that might be shared across different cell populations. Traditional archetypal analysis, though capable of decomposing expression profiles into GEPs, is often linear and may not accurately reflect the complex, non-linear relationships in biological data. This gap leaves a need for methods that can uncover these intricate GEPs and their dynamic usage in a non-linear fashion, especially considering the inherent sparsity and overdispersion of scRNA-seq count data.

Methods

The core of this work is scAAnet, a deep autoencoder designed for non-linear archetypal analysis. Key innovations include:

  • Autoencoder Architecture: scAAnet employs an encoder-decoder structure to map high-dimensional scRNA-seq data into a lower-dimensional latent space and then reconstruct it.
  • Count Distribution-Based Loss: Instead of standard mean squared error, scAAnet uses a loss function based on count distributions (like Zero-Inflated Negative Binomial - ZINB) to better model the sparsity and overdispersion characteristic of scRNA-seq data.
  • Archetypal Constraint: A crucial addition is an archetypal constraint in the loss function. This penalizes the distance between inferred archetypes and data points in the latent space, ensuring that the identified archetypes are truly representative of the data's "extreme" states, a core assumption of archetypal analysis.
  • Regression-Based DEG Test: To identify genes specific to each GEP, the authors developed a DEG test that uses continuous cell usage of GEPs as a predictor, offering a more nuanced approach than discrete group assignments.

Datasets

The authors evaluated scAAnet using:

  • Simulated scRNA-seq Data: Synthetic datasets were generated using a modified Splatter framework to mimic negative binomial and ZINB distributions, allowing for quantitative benchmarking against other methods.
  • Pancreatic Islet Dataset: A publicly available dataset with well-defined cell clusters was used to demonstrate scAAnet's ability to identify biologically meaningful GEPs that correspond to known cell types.
  • Lung Idiopathic Pulmonary Fibrosis (IPF) Dataset: This dataset was used to study fibroblasts and myofibroblasts, revealing an IPF-specific archetype.
  • Prefrontal Cortex Dataset: This dataset, focusing on microglia in the context of Alzheimer's disease, was used to showcase scAAnet's ability to reveal cellular heterogeneity at a higher resolution.

Key Insights and Ideas

  • Non-linearity Matters: The paper strongly advocates for non-linear approaches in archetypal analysis for scRNA-seq, showing that linear methods or those that don't enforce non-linearity can miss crucial biological signals.
  • Count Data is Special: Standard reconstruction losses like MSE are suboptimal for count data. Using distribution-specific losses (like ZINB) significantly improves performance.
  • Archetypal Constraint is Crucial: Simply finding a convex hull in the latent space isn't enough. The explicit archetypal constraint in scAAnet ensures that the inferred archetypes are genuinely representative of the data's extreme points, leading to more interpretable GEPs.
  • Continuous Usage is Informative: Using continuous GEP usage as a predictor for DEG analysis provides richer insights than discrete cell assignments, capturing more nuanced biological differences.

Key Results

  • Superior Performance: scAAnet outperformed existing methods (NMF, scVI, LDVAE, AAnet) in simulations across various metrics, including cell usage inference, GEP inference, and reconstruction error, especially when using count distribution-based losses.
  • Biologically Meaningful GEPs: Applied to real datasets, scAAnet successfully identified GEPs that corresponded well with known cell types and biological states, such as beta cell functions, ER stress response, IPF-specific myofibroblast programs, and microglia heterogeneity in Alzheimer's disease.
  • Improved DEG Identification: The regression-based DEG test using continuous GEP usage demonstrated higher sensitivity in identifying GEP-specific marker genes compared to discrete group assignments.

Limitations and Future Work

  • Archetype Number Selection: The paper acknowledges that selecting the optimal number of archetypes (K) remains a challenge, relying on model stability and reconstruction error as criteria. Future work could explore more automated or robust methods for K selection.
  • Computational Cost: While scAAnet offers advantages, deep learning models can be computationally intensive. Further optimization or exploration of more efficient architectures might be beneficial.
  • Generalizability: While tested on three diverse datasets, further validation on a wider range of scRNA-seq applications and cell types would strengthen the generalizability of scAAnet.
  • Integration with other tools: The authors suggest that the latent representation from scAAnet can serve as a baseline for other visualization tools, hinting at future integration possibilities.

Summary updated

Abstract

Abstract Advances in single-cell RNA sequencing (scRNA-seq) have led to successes in discovering novel cell types and understanding cellular heterogeneity among complex cell populations through cluster analysis. However, cluster analysis is not able to reveal continuous spectrum of states and underlying gene expression programs (GEPs) shared across cell types. We introduce scAAnet, an autoencoder for single-cell non-linear archetypal analysis, to identify GEPs and infer the relative activity of each GEP across cells. We use a count distribution-based loss term to account for the sparsity and overdispersion of the raw count data and add an archetypal constraint to the loss function of scAAnet. We first show that scAAnet outperforms existing methods for archetypal analysis across different metrics through simulations. We then demonstrate the ability of scAAnet to extract biologically meaningful GEPs using publicly available scRNA-seq datasets including a pancreatic islet dataset, a lung idiopathic pulmonary fibrosis dataset and a prefrontal cortex dataset.
BibTeX
@article{wang2022non,
  author = {Yuge Wang and Hongyu Zhao},
  title = {{Non-linear Archetypal Analysis of Single-cell RNA-seq Data by Deep Autoencoders}},
  journal = {PLOS Computational Biology},
  year = {2022},
  url = {https://openreview.net/forum?id=D5Ok6KpBr6m},
}
Figure