Massive single-cell RNA-seq analysis and imputation via deep learning

Yue Deng, Feng Bao, Qionghai Dai, Lani F. Wu, Steven J. AltschulerbioRxiv 2018 · DOI

TL;DR

This paper introduces scScope, a novel deep learning framework designed to tackle the challenges of analyzing massive, noisy single-cell RNA sequencing (scRNA-seq) datasets. scScope efficiently handles millions of cells, accurately identifies cell types, and imputes missing gene expression data, paving the way for deeper biological insights from large-scale single-cell experiments.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study complex biological systems at the cellular level, enabling the identification of rare cell populations, developmental programs, and cellular responses to perturbations. However, recent technological advancements have led to the generation of scRNA-seq data at an unprecedented scale, with millions of cells being analyzed. This massive scale introduces two critical challenges:

  1. Data Noise: Single-cell profiles are inherently noisy due to transcript amplification noise and dropout events (where a gene is expressed but not detected). These artifacts become more pronounced in large datasets.
  2. Scalability: Existing analytical tools struggle to handle the computational demands (memory and speed) of analyzing millions of cells and tens of thousands of genes simultaneously.

These challenges necessitate new approaches that can extract meaningful information from extremely noisy, high-dimensional, and massive scRNA-seq datasets.

Methods

scScope is a deep learning-based approach with a novel recurrent network architecture designed for efficient and accurate analysis of large-scale scRNA-seq data. Its key components include:

  • Encoder-Decoder Architecture: This core structure learns low-dimensional representations (feature extraction) of gene expression profiles and reconstructs them, effectively denoising the data.
  • Self-Correcting Imputation Layer: This innovative layer iteratively imputes dropout events (missing gene expression values) by leveraging the co-regulation patterns of genes observed across many cells. This is a significant advancement over standard autoencoders, which struggle with missing data.
  • Batch Effect Correction: scScope includes an optional module to correct for batch effects, ensuring that variations due to experimental conditions do not confound biological findings.
  • Parallelized Training: The framework is designed to utilize multiple GPUs, dramatically reducing training time for massive datasets.

Datasets

The authors evaluated scScope on a variety of datasets, including:

  • Simulated Datasets:
    • Moderate-sized datasets generated by Splatter (2K cells, 3 subpopulations, varying sparsity levels).
  • Large-scale datasets generated by a simulation framework inspired by SIMLR (1M cells, 50 subpopulations, varying sparsity and rare population fractions).
  • Biological Datasets:
    • Mouse Lung Tissue Data: ~7K cells from three batches, used for batch effect correction evaluation.
  • Cord Blood Mononuclear Cells (CBMC) Data: ~8K cells, used for imputation accuracy assessment across different gene expression levels.
  • Mouse Retina Data: ~44K cells, used for identifying minor subpopulations.
  • Mouse Cell Atlas (MCA) Data: ~400K cells from 51 tissues, used to test scalability with increasing gene numbers.
  • Intestinal Data: ~10K cells from mice under different infection conditions, used to explore biological changes.
  • Brain Data: ~1.3 million cells from embryonic mouse brains, used for large-scale analysis and subpopulation discovery.

Key Insights and Ideas

  • Recurrent Imputation for Robustness: The iterative nature of scScope's imputation layer is crucial for accurately filling in missing gene expression values, especially in datasets with high dropout rates. This goes beyond simple denoising by actively correcting for technical artifacts.
  • Scalability is Paramount: The paper demonstrates that many existing methods fail to scale to millions of cells, highlighting the critical need for approaches like scScope that are built for large-scale biological data.
  • Deep Learning for Unsupervised Discovery: scScope's ability to learn informative latent representations without prior cell type labels is key to its power in discovering novel subpopulations and biological insights.
  • Unified Framework: scScope integrates batch effect correction, feature learning, and imputation into a single, parallelizable deep learning framework, streamlining complex scRNA-seq analysis.

Key Results

  • Scalability and Speed: scScope analyzed a 1.3 million cell dataset in under 50 minutes on a single GPU, significantly outperforming non-deep learning methods (which couldn't scale beyond 100K cells) and deep learning methods (which took at least seven times longer).
  • Imputation Accuracy: scScope showed superior imputation accuracy, particularly at high sparsity levels, and maintained small imputation errors across the entire range of gene expression levels.
  • Batch Effect Correction: scScope demonstrated a dramatic improvement in clustering accuracy when batch correction was enabled on the lung tissue dataset.
  • Minor Subpopulation Discovery: scScope accurately identified minor cell subpopulations, showing high consistency with previous findings and identifying statistically significant enrichment of cell-type markers.
  • Gene Number Analysis: scScope achieved the best performance when analyzing 10,000 genes on a large dataset, demonstrating its ability to handle increased gene depth.
  • Biological Discovery: scScope successfully identified novel enterocyte subpopulations in intestinal data and discovered biologically meaningful clusters in large brain datasets, leading to new predictions about gene function and cell responses.

Limitations and Future Work

  • Computational Resources: While scScope is highly scalable, training on extremely large datasets still requires significant computational resources, particularly GPUs.
  • Parameter Tuning: While scScope offers flexibility, users may need to tune parameters like the number of recurrent steps (T) for optimal performance on specific datasets. The paper suggests T=2 offers a good balance between speed and accuracy.
  • Future Work: The authors suggest that scScope's platform can be integrated with any clustering method, implying potential for further development in downstream analysis pipelines. The ability to handle massive and noisy data opens doors for exploring more complex biological questions and discovering rarer cell types.

Summary updated

Abstract

Recent advances in large-scale single cell RNA-seq enable fine-grained characterization of phenotypically distinct cellular states within heterogeneous tissues. We present scScope, a scalable deep-learning based approach that can accurately and rapidly identify cell-type composition from millions of noisy single-cell gene-expression profiles.
BibTeX
@misc{deng2018massive,
  author = {Yue Deng and Feng Bao and Qionghai Dai and Lani F. Wu and Steven J. Altschuler},
  title = {{Massive single-cell RNA-seq analysis and imputation via deep learning}},
  year = {2018},
  doi = {10.1101/315556},
  howpublished = {bioRxiv},
}
Figure