Efficient and scalable integration of single-cell data using domain-adversarial and variational approximation

Jialu Hu, Yuanke Zhong, Xuequn ShangbioRxiv 2021 · DOI

TL;DR

This paper introduces DAVAE, a novel framework for integrating diverse single-cell datasets (across samples, technologies, and modalities) without requiring post-hoc processing. DAVAE leverages a domain-adversarial variational autoencoder to learn a shared, low-dimensional representation, demonstrating superior performance in batch effect removal, transfer learning, and cell type prediction.

What Problem Does This Paper Address?

Single-cell sequencing technologies have revolutionized biology, generating massive datasets that promise deep insights into cellular heterogeneity, development, and disease. However, a major hurdle is integrating these datasets, which often come from different experiments, technologies (e.g., scRNA-seq, scATAC-seq, spatial transcriptomics), and even modalities. These differences introduce "batch effects" and other technical variations that obscure true biological signals, making it difficult to compare cells or identify cell types across datasets. Existing methods often require complex pre-processing or struggle with integrating diverse data types. This paper aims to address this by providing a unified, scalable, and effective framework for seamless single-cell data integration.

Methods

The core of DAVAE is a Domain-Adversarial Variational Autoencoder (DAVAE). This framework combines several key components:

  • Variational Autoencoder (VAE): A generative model that learns a compressed, low-dimensional representation (latent space) of the input data. It consists of an encoder that maps input data to the latent space and a decoder that reconstructs the input from the latent representation.
  • Non-linear Mapping: DAVAE uses a deep neural network to model the relationship between the latent space and the expression space, allowing it to capture complex biological relationships.
  • KL Regularizer: This component of the VAE encourages the learned latent representations to follow a standard distribution (e.g., a normal distribution), which helps in regularization and interpretability.
  • Domain-Adversarial Regularizer: This is the novel aspect. A domain classifier is trained to predict the origin (e.g., batch, technology) of the data from its latent representation. Simultaneously, a gradient reversal layer (GRL) is used to train the encoder to produce latent representations that fool this classifier. This adversarial process forces the encoder to learn representations that are invariant to batch effects and other domain-specific variations, effectively aligning different datasets in the latent space.

The framework is trained end-to-end, meaning all components are optimized together without the need for separate pre-processing steps.

Datasets

The authors evaluated DAVAE on five real-world single-cell datasets, showcasing its versatility:

  1. Human Dendritic Cells: Integrating scRNA-seq data from eight different samples to demonstrate batch effect removal.
  2. 293T + Jurkat Cells: Integrating scRNA-seq datasets with different cellular compositions to test the ability to separate and mix cell types correctly.
  3. Human Cell Atlas: Integrating two large-scale scRNA-seq datasets (cord blood and bone marrow) to assess scalability.
  4. Spatial Transcriptomics and scRNA-seq on Mouse Brain: Integrating spatial transcriptomics data with scRNA-seq data to demonstrate cross-modality integration and transfer learning of cell type labels.
  5. scATAC-seq and scRNA-seq on PBMCs: Integrating chromatin accessibility data (scATAC-seq) with gene expression data (scRNA-seq) to show multi-modal integration.

Key Insights and Ideas

  • Unifying Diverse Data: The most significant insight is DAVAE's ability to integrate data from vastly different sources (samples, technologies, modalities) into a single, coherent representation without manual pre-processing. This breaks down data silos and allows for more comprehensive analyses.
  • Adversarial Learning for Batch Correction: The domain-adversarial component is a clever way to explicitly learn batch-invariant representations. By training a classifier to distinguish batches and then forcing the encoder to confuse it, DAVAE effectively "washes out" batch-specific noise.
  • Scalability is Key: The paper emphasizes DAVAE's efficiency and scalability, particularly for large datasets, which is crucial for the growing volume of single-cell data.
  • Transfer Learning Across Modalities: DAVAE enables transferring knowledge (like cell type labels) from one data modality to another, opening up new avenues for annotation and discovery.

Key Results

  • Effective Batch Effect Removal: DAVAE consistently outperformed other leading methods (Scanorama, DESC, Seurat v3) in metrics like Adjusted Rand Index (ARI) for cell type purity and 1-ARI for batch mixing, and kBET acceptance rate across various datasets.
  • Accurate Cell Type Prediction: In cross-modality integration (scATAC-seq and spatial transcriptomics), DAVAE successfully transferred cell type labels from scRNA-seq references, demonstrating its utility for annotation.
  • Scalability: DAVAE showed significant advantages in running time compared to Scanorama when applied to large datasets (up to 600,000 cells), highlighting its efficiency.
  • Preservation of Biological Signal: Despite removing batch effects, DAVAE effectively preserved the underlying biological heterogeneity and cell type distinctions within the integrated data.

Limitations and Future Work

  • Computational Resources: While scalable, training deep learning models like DAVAE can still be computationally intensive, requiring GPUs for optimal performance.
  • Hyperparameter Tuning: As with most deep learning models, DAVAE likely requires careful tuning of hyperparameters (e.g., learning rate, regularization weights) for optimal performance on specific datasets.
  • Future Work: The authors suggest that DAVAE's efficient and scalable nature makes it a valuable tool for distinguishing complex cellular heterogeneity as single-cell genomics continues to evolve. Further exploration of its application to even larger and more diverse datasets, as well as its potential for discovering novel biological insights, would be valuable.

Summary updated

Abstract

Abstract Single-cell data provides us new ways of discovering biological truth at the level of individual cells, such as identification of cellular sub-populations and cell development. With the development of single-cell sequencing technologies, a key analytical challenge is to integrate these data sets to uncover biological insights. Here, we developed a domain-adversarial and variational approximation framework, DAVAE, to integrate multiple single-cell data across samples, technologies and modalities without any post hoc data processing. We fit normalized gene expression into a non-linear model, which transforms a latent variable of a lower-dimension into expression space with a non-linear function, a KL regularizier and a domain-adversarial regularizer. Results on five real data integration applications demonstrated the effectiveness and scalability of DAVAE in batch-effect removing, transfer learning, and cell type predictions for multiple single-cell data sets across samples, technologies and modalities. DAVAE was implemented in the toolkit package “scbean” in the pypi repository, and the source code can be also freely accessible at https://github.com/jhu99/scbean .
BibTeX
@misc{hu2021efficient,
  author = {Jialu Hu and Yuanke Zhong and Xuequn Shang},
  title = {{Efficient and scalable integration of single-cell data using domain-adversarial and variational approximation}},
  year = {2021},
  doi = {10.1101/2021.04.06.438733},
  howpublished = {bioRxiv},
}
Figure