SAILER: Scalable and Accurate Invariant Representation Learning for Single-Cell ATAC-Seq Processing and Integration

Yingxin Cao, Laiyi Fu, Jie Wu, Qinke Peng, Qing Nie, Jing Zhang, Xiaohui XiebioRxiv 2021 · DOI

TL;DR

This paper introduces SAILER, a novel deep generative model for single-cell ATAC-seq (scATAC-seq) data. SAILER learns cell representations that are robust to technical noise like read depth and batch effects, leading to improved downstream analyses like clustering and imputation. Its scalable architecture makes it suitable for analyzing millions of cells.

What Problem Does This Paper Address?

Single-cell ATAC-seq (scATAC-seq) is a powerful technology for understanding cellular heterogeneity by mapping accessible chromatin regions. However, analyzing this data is challenging due to its high dimensionality, extreme sparsity, complex interdependencies between genomic regions, and sensitivity to technical variations such as differences in sequencing depth and batch effects from multiple experiments. These confounding factors can obscure true biological signals, leading to inaccurate cell clustering and downstream analyses. Existing methods often struggle to effectively disentangle these technical artifacts from genuine biological differences.

Methods

SAILER is built upon a deep generative model, specifically a variational autoencoder (VAE) with an encoder-decoder architecture. The core innovation lies in its "invariant representation learning" approach. SAILER aims to learn a low-dimensional latent representation ($z$) for each cell that captures its intrinsic biological state. To achieve this, it imposes an additional constraint during training: minimizing the mutual information between the learned latent representation ($z$) and known confounding factors ($c$), such as read depth and batch indicators. This effectively forces the model to learn representations that are independent of these technical variations. The model architecture uses fully connected neural networks for both the encoder and decoder. For reconstruction, it uses a binomial likelihood to account for the binary nature of scATAC-seq data, with a weighted cross-entropy loss to handle sparsity.

Datasets

The authors evaluated SAILER using three types of datasets:

  • Simulated Datasets: Three simulated scATAC-seq datasets were generated using SCAN-ATAC-Sim to represent different scenarios:
  • Sim1: Single sample, 5 cell types, including a rare cell type, with varying sequencing depth.
  • Sim2: Single sample, 9 cell types, with varying sequencing depth.
  • Sim3: Two samples with different signal-to-noise ratios and batch effects, including a sample-specific rare cell type.
  • Real Datasets:
    • Mouse Atlas Dataset: A large-scale dataset comprising 81,173 adult mouse cells from 13 tissues and 40 cell types.
  • Mixed Mouse Brain Datasets: Two mouse brain scATAC-seq datasets from different sources (10X Genomics and SnapATAC website) were combined to test batch effect removal.

Key Insights and Ideas

  • Confounding Factors Dominate Heterogeneity: The paper highlights that a significant portion of the observed heterogeneity in scATAC-seq data can be attributed to technical factors like read depth and batch effects, rather than solely biological differences.
  • Invariant Representation is Key: The central idea is that true biological variations should be consistent across samples and technical conditions, while confounding factors vary. SAILER explicitly learns representations that are invariant to these confounding factors.
  • Deep Generative Models for Complex Dependencies: By using a deep generative framework, SAILER can capture complex, non-linear dependencies between genomic regions, which linear methods might miss.
  • Unified Framework: SAILER offers a single framework for denoising, clustering, and imputation, integrating these tasks rather than treating them as separate steps.

Key Results

  • Improved Clustering: SAILER demonstrated significant improvements in clustering performance on simulated datasets, particularly in identifying rare cell types. On the mouse atlas dataset, SAILER achieved a higher Normalized Mutual Information (NMI) of 0.799 compared to SnapATAC (0.748) and SCALE (0.557).
  • Effective Batch Effect Removal: When integrating two mouse brain datasets, SAILER successfully mingled cells from different batches into homogeneous clusters, unlike SCALE which showed distinct batch separation. Mutual information analysis confirmed SAILER's superior ability to remove batch information.
  • Enhanced Imputation: SAILER's imputed data showed better separation of cell types and batches compared to SCALE and MAGIC, and achieved a higher Dice Similarity Coefficient (DSC) of 0.64 compared to SCALE's 0.54 on simulated data.
  • Scalability: SAILER can process millions of cells efficiently due to its GPU-accelerated deep learning architecture, outperforming methods with quadratic scaling.

Limitations and Future Work

  • Limitations: While SAILER effectively removes confounding factors, the paper doesn't explicitly detail limitations in terms of specific types of confounding factors it might struggle with beyond read depth and batch effects. The imputation process relies on fixing confounding variables, which might not always be ideal if those variables also carry subtle biological information.
  • Future Work: The authors suggest that the invariant representation learning framework is general and can be extended to other high-throughput genomic data types (e.g., scRNA-seq, single-cell DNA methylation) and multi-modal single-cell omics data. They propose using multi-modal VAEs to integrate different data modalities, with shared and private latent spaces for cell states and modality-specific issues, respectively.

Summary updated

Abstract

Abstract Motivation Single-cell sequencing assay for transposase-accessible chromatin (scATAC-seq) provides new opportunities to dissect epigenomic heterogeneity and elucidate transcriptional regulatory mechanisms. However, computational modelling of scATAC-seq data is challenging due to its high dimension, extreme sparsity, complex dependencies, and high sensitivity to confounding factors from various sources. Results Here we propose a new deep generative model framework, named SAILER, for analysing scATAC-seq data. SAILER aims to learn a low-dimensional nonlinear latent representation of each cell that defines its intrinsic chromatin state, invariant to extrinsic confounding factors like read depth and batch effects. SAILER adopts the conventional encoder-decoder framework to learn the latent representation but imposes additional constraints to ensure the independence of the learned representations from the confounding factors. Experimental results on both simulated and real scATAC-seq datasets demonstrate that SAILER learns better and biologically more meaningful representations of cells than other methods. Its noise-free cell embeddings bring in significant benefits in downstream analyses: Clustering and imputation based on SAILER result in 6.9% and 18.5% improvements over existing methods, respectively. Moreover, because no matrix factorization is involved, SAILER can easily scale to process millions of cells. We implemented SAILER into a software package, freely available to all for large-scale scATAC-seq data analysis. Availability The software is publicly available at https://github.com/uci-cbcl/SAILER Contact [email protected] and [email protected]
BibTeX
@misc{cao2021sailer,
  author = {Yingxin Cao and Laiyi Fu and Jie Wu and Qinke Peng and Qing Nie and Jing Zhang and Xiaohui Xie},
  title = {{SAILER: Scalable and Accurate Invariant Representation Learning for Single-Cell ATAC-Seq Processing and Integration}},
  year = {2021},
  doi = {10.1101/2021.01.28.428689},
  howpublished = {bioRxiv},
}
Figure