Besca , a single-cell transcriptomics analysis toolkit to accelerate translational research

Sophia C. Mädler, Alice Julien-Laferrière, Luis Wyss, Miroslav Phan, Albert S. W. Kang, Eric Ulrich, Roland Schmucki, Jitao David Zhang, Martin Ebeling, Laura Badi, Tony Kam‐Thong, Petra Schwalie, Klas HatjebioRxiv 2020 · DOI

TL;DR

This paper introduces Besca, a comprehensive Python toolkit designed to streamline and standardize single-cell RNA sequencing (scRNA-seq) analysis. Besca aims to accelerate translational research by providing a robust workflow for quality control, cell annotation (using both signature-based and machine learning approaches), and cell deconvolution from bulk RNA-seq data, making scRNA-seq analysis more accessible, reproducible, and interpretable.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of biological systems, but its analysis is complex and can be a bottleneck for translational research. Key challenges include:

  • Lack of Standardization: Different labs use varying protocols and analysis pipelines, making it difficult to compare results across studies.
  • Cell Type Annotation Bottleneck: Accurately identifying and annotating cell types is crucial but often manual, time-consuming, and prone to inconsistency.
  • Data Integration and Reusability: Integrating data from diverse studies and reusing existing annotations or gene signatures is challenging.
  • Bridging the Gap to Bulk Data: Translating scRNA-seq insights into actionable information from bulk RNA-seq data (which is more common in clinical settings) requires robust deconvolution methods.

Besca aims to address these issues by providing a unified, standardized, and automated toolkit that enhances the efficiency, reproducibility, and interpretability of scRNA-seq data analysis.

Methods

Besca offers a modular approach built upon the popular Scanpy library. Its core components include:

  • Standard Workflow: A generalized pipeline for scRNA-seq data processing, covering quality control (QC), filtering, normalization, dimensionality reduction (PCA, UMAP), and clustering (Leiden algorithm). This workflow is designed to be reproducible and interoperable, producing FAIR data outputs. It also supports CITE-seq data.
  • Sig-annot Module: This module automates cell type annotation using a hierarchical signature-enrichment approach. It leverages pre-defined gene signatures and a hierarchical cell ontology to assign cell types to clusters, mirroring manual annotation but enforcing standardization and comparability.
  • Auto-annot Module: This module employs supervised machine learning (SVM or logistic regression) for cell type annotation. It trains classifiers on well-annotated reference datasets, allowing for efficient annotation of new datasets, especially when transferring knowledge from large cell atlases.
  • Bescape Module: This framework facilitates cell deconvolution from bulk RNA-seq data using scRNA-seq derived gene expression profiles (GEPs). It integrates established deconvolution tools like SCDC and MuSiC, providing a containerized environment for seamless execution and allowing users to generate custom GEPs.

Datasets

Besca was evaluated using a variety of publicly available scRNA-seq datasets, including:

  • Hematopoietic Cells: PBMC3k, Granja2019, Kotliarov2020 (covering peripheral blood, bone marrow, and immune responses to vaccines).
  • Intestinal Cells: Smillie2019, Martin2019, Haber2017 (mouse), Lee2020 (covering colon and small intestine epithelium, lamina propria, and colorectal cancer).
  • Pancreas Cells: Segerstolpe2016, Peng2019, Baron2016 (covering pancreatic islets, pancreatic ductal adenocarcinoma, and general pancreatic samples).

These datasets represent diverse tissues, species, and experimental conditions, demonstrating the broad applicability of Besca.

Key Insights and Ideas

  • Harmonized Annotation is Key: The paper strongly emphasizes that standardized cell type annotation is critical for enabling cross-study comparisons and building comprehensive cell atlases. Besca's Sig-annot and Auto-annot modules directly address this by providing reproducible and harmonized annotation schemas.
  • Machine Learning for Scalability: Auto-annot demonstrates that supervised machine learning can effectively transfer knowledge from well-annotated datasets to new ones, offering a scalable solution for cell type identification, especially as cell atlases grow.
  • Bridging Bulk and Single-Cell Data: Bescape's ability to leverage scRNA-seq data for bulk RNA-seq deconvolution is a significant contribution, enabling researchers to extract cell-type-specific insights from more readily available bulk data.
  • Reusability of Gene Signatures: The integration with the Geneset Management System (GeMS) highlights the importance of managing and reusing gene signatures, a crucial aspect for knowledge accumulation and efficient analysis.

Key Results

  • Besca's standard workflow successfully processed diverse scRNA-seq datasets, generating QC reports and interoperable analysis results.
  • The Sig-annot module provided consistent and harmonized cell type annotations across different datasets, often improving upon or matching original author annotations.
  • The Auto-annot module demonstrated high accuracy in cell type prediction, particularly for B cells, and showed promise in cross-species comparisons.
  • Bescape achieved high correlation (Pearson correlation > 0.9 for many cell types) between predicted and ground truth cell proportions in simulated bulk RNA-seq data, validating its deconvolution capabilities.
  • The tool successfully identified differences in cell proportions in real bulk RNA-seq data, such as lower pancreatic beta cell proportions in type 2 diabetes patients compared to healthy controls.

Limitations and Future Work

  • Annotation Ambiguity: While Auto-annot is powerful, some T cell subtypes showed ambiguity, leading to classification as "unknown" when a probability threshold was applied. This highlights the ongoing challenge of fine-grained cell type resolution.
  • Dataset Specificity: The performance of both Sig-annot and Auto-annot can be influenced by the quality and specific characteristics of the training/reference datasets.
  • Future Work: The authors envision Besca accelerating translational research by enabling more robust cross-study comparisons, facilitating the integration of data into larger cell atlases, and providing a foundation for understanding disease mechanisms and interventions. Further development could focus on expanding the annotation schemas, improving deconvolution accuracy for challenging cell types, and integrating more advanced analysis modules.

Summary updated

Abstract

Abstract Single-cell RNA sequencing (scRNA-seq) revolutionised our understanding of disease biology and presented the promise of transforming translational research. We developed Besca , a toolkit that streamlines scRNA-seq analyses according to current best practices. A standard workflow covers quality control, filtering, and clustering. Two complementary Besca modules, utilizing hierarchical cell signatures or supervised machine learning, automate cell annotation and provide harmonised nomenclatures across studies. Subsequently, Besca enables estimation of cell type proportions in bulk transcriptomics studies. Using multiple heterogeneous scRNA-seq datasets we show how Besca aids acceleration, interoperability, reusability, and interpretability of scRNA-seq data analysis, crucial aspects in translational research and beyond.
BibTeX
@misc{madler2020i,
  author = {Sophia C. Mädler and Alice Julien-Laferrière and Luis Wyss and Miroslav Phan and Albert S. W. Kang and Eric Ulrich and Roland Schmucki and Jitao David Zhang and Martin Ebeling and Laura Badi and Tony Kam‐Thong and Petra Schwalie and Klas Hatje},
  title = {{<i>Besca</i> , a single-cell transcriptomics analysis toolkit to accelerate translational research}},
  year = {2020},
  doi = {10.1101/2020.08.11.245795},
  howpublished = {bioRxiv},
}
Figure