scAdapt: Virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species

Xiang Zhou, Hua Chai, Yuansong Zeng, Huiying Zhao, Ching‐Hsing Luo, Yuedong Yang

TL;DR

This paper introduces scAdapt, a novel deep learning framework that leverages virtual adversarial domain adaptation to accurately classify cell types in single-cell RNA sequencing (scRNA-seq) data, even when datasets come from different platforms, species, or experimental conditions. scAdapt effectively bridges these "batch effects" by aligning data distributions and preserving cell type structures, outperforming existing methods in various challenging scenarios.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) allows us to study cellular heterogeneity at an unprecedented resolution. A crucial step in analyzing this data is identifying cell types. Traditionally, this involved manually matching gene expression patterns to known markers, a process that is time-consuming, requires expert knowledge, and can be difficult to reproduce. While automated methods have emerged, they often struggle when applied to datasets generated under different conditions (e.g., different labs, technologies, or even species). These differences, known as "batch effects," cause variations in data distribution, making it hard for models trained on one dataset to generalize to another. This paper tackles the challenge of robust cell type classification in the face of these significant data variations.

Methods

scAdapt is built upon a deep learning architecture that combines several key techniques:

  • Virtual Adversarial Training (VAT): This semi-supervised learning approach enhances classifier robustness by making it invariant to small perturbations in the input data. It helps the model learn a more stable decision boundary, especially when dealing with unlabeled target data.
  • Domain Adversarial Adaptation: This technique aims to make the learned feature representations domain-invariant. A domain classifier tries to distinguish between source and target data, while the feature extractor tries to fool it, forcing the learned embeddings to be indistinguishable across datasets. This addresses the global distribution mismatch.
  • Semantic Alignment: To preserve the discriminative structure of cell types, scAdapt aligns the centroids of labeled source cell clusters with the centroids of pseudo-labeled target cell clusters. This "category-level alignment" ensures that similar cell types from different datasets are mapped closely in the embedding space, even when explicit labels are unavailable for the target data.
  • Joint Optimization: Crucially, scAdapt jointly optimizes these components (classification, domain adaptation, and semantic alignment) to achieve both accurate cell type classification and effective batch correction simultaneously.

Datasets

The authors evaluated scAdapt on a diverse range of datasets:

  • Simulated Datasets: Generated using the "Splatter" R package to control for varying degrees of batch effects and clustering difficulty.
  • Cross-Platform Datasets: Human Peripheral Blood Mononuclear Cells (PBMC) data from seven different sequencing platforms.
  • Cross-Species Datasets: Human and mouse pancreas data, requiring gene name harmonization.
  • Spatial Transcriptomic Datasets: Mouse brain data combining spatially resolved MERFISH with dissociated scRNA-seq (10x Chromium) data, featuring a limited number of overlapping genes.

These datasets collectively form a comprehensive benchmark for evaluating the robustness of cell type classification and batch correction methods across various biological and technical variations.

Key Insights and Ideas

  • The Power of Joint Learning: The most significant insight is that simultaneously addressing cell type classification and batch correction within a single, unified framework (scAdapt) leads to superior performance compared to methods that tackle these issues separately or sequentially.
  • Semantic Alignment for Robustness: Explicitly aligning cluster centroids between labeled source and pseudo-labeled target data is critical for preserving cell type structure and achieving accurate classification, especially when batch effects are strong.
  • VAT for Unlabeled Data: Virtual Adversarial Training proves highly effective in leveraging unlabeled target data to improve classifier performance and robustness, a common scenario in real-world scRNA-seq analysis.
  • Beyond Global Alignment: While global domain adaptation is important, scAdapt's addition of semantic alignment demonstrates that preserving class-specific structures is equally vital for accurate downstream tasks.

Key Results

  • Superior Classification Accuracy: scAdapt consistently outperformed seven other classification methods across simulated, cross-platform, and cross-species datasets, achieving average accuracies of up to 0.98 on simulated data and 0.93 on cross-species data.
  • Effective Batch Correction: When evaluated as a batch correction method, scAdapt achieved the highest silhouette scores (indicating good cell type separation) and lowest divergence scores (indicating good cell mixing) on simulated and cross-platform datasets, outperforming methods like Seurat, fastMNN, Harmony, and LIGER.
  • Interpretability: The activation maximization technique successfully identified cell-type-specific marker genes, many of which were validated against existing databases and literature, demonstrating the model's biological interpretability.
  • Handling Limited Genes: scAdapt showed promising results on spatial transcriptomic data with a limited number of overlapping genes, achieving an average accuracy of 0.87.

Limitations and Future Work

  • Subtype Identification: The authors acknowledge that distinguishing very similar subtypes within a cell type remains a challenge. They suggest that fine-grained classification techniques from computer vision and deep metric learning could be explored for future improvements.
  • Intra-class Variance: The current softmax loss tends to collapse cells of the same type into a single cluster. Future work could incorporate methods that better preserve intra-class variance for more nuanced clustering.
  • Reference Data Quality: The performance of scAdapt relies on the source dataset containing a reasonable number of cells per cell type to adequately represent their transcriptional programs. The authors recommend at least 10 cells per cell type for optimal results.

Summary updated

Abstract

Abstract Motivation In single cell analyses, cell types are conventionally identified based on known marker gene expressions. Such approaches are time-consuming and irreproducible. Therefore, many new supervised methods have been developed to identify cell types for target datasets using the rapid accumulation of public datasets. However, these approaches are sensitive to batch effects or biological variations since the data distributions are different in cross-platforms or species predictions. Results We developed scAdapt, a virtual adversarial domain adaptation network to transfer cell labels between datasets with batch effects. scAdapt used both the labeled source and unlabeled target data to train an enhanced classifier, and aligned the labeled source centroid and pseudo-labeled target centroid to generate a joint embedding. We demonstrate that scAdapt outperforms existing methods for classification in simulated, cross-platforms, cross-species, and spatial transcriptomic datasets. Further quantitative evaluations and visualizations for the aligned embeddings confirm the superiority in cell mixing and preserving discriminative cluster structure present in the original datasets. Availability https://github.com/zhoux85/scAdapt . Contact [email protected] or [email protected]
BibTeX
@article{zhou2021scadapt,
  author = {Xiang Zhou and Hua Chai and Yuansong Zeng and Huiying Zhao and Ching‐Hsing Luo and Yuedong Yang},
  title = {{scAdapt: Virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species}},
  journal = {Briefings in Bioinformatics},
  year = {2021},
  doi = {10.1093/bib/bbab281},
}
Figure