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 learning framework for non-linear archetypal analysis of single-cell RNA sequencing (scRNA-seq) data. scAAnet effectively decomposes complex cellular heterogeneity into underlying gene expression programs (GEPs) and their relative activities, outperforming existing methods in both simulated and real-world biological datasets.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity. While clustering methods are useful for identifying distinct cell types, they struggle to capture continuous spectrums of cell states or identify shared gene expression programs (GEPs) that may span across different cell types. Traditional archetypal analysis, while capable of decomposing expression profiles into GEPs, is linear and cannot account for the complex, non-linear relationships often observed in biological systems. This limitation hinders a deeper understanding of how GEPs contribute to cellular function and disease.

Methods

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

  • Autoencoder Architecture: scAAnet uses an encoder-decoder structure to learn a low-dimensional latent representation of the scRNA-seq data.
  • Count Distribution-Based Loss: Unlike traditional methods that often use Mean Squared Error (MSE), scAAnet employs a loss function that models the count nature of scRNA-seq data using distributions like Zero-Inflated Negative Binomial (ZINB), Zero-Inflated Poisson (ZIP), Poisson, and Negative Binomial (NB). This better accounts for data sparsity and overdispersion.
  • Archetypal Constraint: A crucial addition is an "archetypal loss" that penalizes the distance between inferred archetypes and data points in the latent space. This ensures that the learned archetypes are truly representative of the "pure" cell states or GEPs within the data, a key assumption of archetypal analysis.
  • Downstream Analysis Tools: The authors also provide tools for identifying GEP-specific marker genes and performing differential gene expression (DEG) tests using continuous cell usage variables, offering a more nuanced approach than discrete group assignments.

Datasets

The authors evaluated scAAnet on several publicly available scRNA-seq datasets:

  • Simulated Datasets: Synthetic scRNA-seq data was generated using the Splatter framework, with varying distributions (NB, ZINB) and signal-to-noise ratios to rigorously test method performance.
  • Pancreatic Islet Dataset: A well-annotated dataset with 10 distinct cell clusters, used to demonstrate scAAnet's ability to capture known cell populations.
  • Lung Idiopathic Pulmonary Fibrosis (IPF) Dataset: This dataset, focusing on fibroblasts and myofibroblasts, was used to investigate disease-specific GEPs.
  • Prefrontal Cortex Dataset: This dataset, comprising microglia from individuals with and without Alzheimer's disease (AD) pathology, was used to study GEPs associated with disease states and sex differences.

Key Insights and Ideas

  • Non-linearity is Crucial: Biological systems are inherently non-linear. Capturing this non-linearity in archetypal analysis is essential for accurately decomposing complex cellular states and identifying GEPs.
  • Count Data Matters: Standard MSE loss functions are ill-suited for scRNA-seq count data. Using count-distribution-aware loss functions significantly improves performance.
  • Archetypal Constraint is Key: Simply finding a convex hull in latent space (as some other methods do) is insufficient. Forcing archetypes to be close to the data points in the latent space is critical for identifying meaningful GEPs.
  • Continuous Usage is Powerful: Using continuous cell usage variables for DEG analysis provides a more sensitive and informative approach than relying on discrete cluster assignments.
  • GEPs Bridge Cell Types: scAAnet can reveal GEPs that are shared across different cell types, offering insights into common biological mechanisms or responses.

Key Results

  • Superior Performance in Simulations: scAAnet consistently outperformed competing methods (NMF, scVI, LDVAE, AAnet) in cell usage and GEP inference accuracy, especially when using count distribution-based loss terms.
  • Biologically Meaningful GEPs: The model successfully identified GEPs that corresponded to known cell types in the pancreatic islet dataset, disease-specific programs in IPF, and AD-associated GEPs in microglia.
  • Improved Archetypal Space Visualization: MDS plots showed that scAAnet's inferred archetypes were tightly located at the corners of the data points, indicating a better fit to the archetypal analysis assumptions compared to other methods.
  • Robustness to Parameter Choices: The method showed robustness to changes in the relative weight of the archetypal and reconstruction losses, and reasonable performance even with misspecified numbers of archetypes.

Limitations and Future Work

  • Computational Cost: scAAnet, being a deep learning method, is more computationally intensive and requires more hyperparameter tuning than traditional clustering methods.
  • Number of Archetypes (K): Selecting the optimal number of archetypes remains an unsupervised challenge, though the authors provide stability-based criteria.
  • Complex Trajectories: While scAAnet shows potential for simple trajectory analysis (linear, bifurcation), it may not fully capture highly complex topologies like multifurcations or circular trajectories.
  • Multi-omics Integration: The authors suggest extending scAAnet to integrate multi-modal single-cell data (e.g., transcriptomics with epigenomics) for a more comprehensive understanding of GEPs.
  • Application to Subsets: For large datasets, the authors recommend applying scAAnet to specific cell subsets of interest or those with known shared programs, rather than the entire dataset.

Summary updated

Abstract

Author summary Single-cell RNA sequencing (scRNA-seq) techniques enable the profiling of gene expression at the single-cell level, and thus make it possible to uncover the cellular heterogeneity in a complex cell population which is composed of multiple cell types. Due to the complexity of biological system, different cell types may share underlying gene expression programs (GEPs) at different levels. However, such shared patterns are difficult to study by traditional cluster analysis. Based on the assumption that the expression profile of each cell results from a non-linear combination of multiple GEPs, we develop scAAnet, a deep learning model for non-linear archetypal decomposition of scRNA-seq data. We demonstrate that scAAnet is able to both achieve better decomposition performance in simulated data and identify biologically meaningful GEPs that are either cell-type-specific or disease-enriched in three real scRNA-seq datasets. To help interpret results from scAAnet, we also provide downstream analysis tools for the identification of program-specific marker genes. We expect scAAnet can be applied to explore GEPs shared across cells when scRNA-seq is used to study a complex disease or biological system.
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