Evaluating cell type deconvolution in FFPE breast tissue: application to benign breast disease

Abstract

Transcriptome profiling using RNA sequencing (RNA-seq) of bulk formalin-fixed paraffin-embedded (FFPE) tissue blocks is a standard method in biomedical research. However, when used on tissues with diverse cell type compositions, it yields averaged gene expression profiles, complicating biomarker identification due to variations in cell proportions. To address the need for optimized strategies for defining individual cell type compositions from bulk FFPE samples, we constructed single-cell RNA-seq reference data for breast tissue and tested cell type deconvolution methods. Initial simulation experiments showed similar performances across multiple commonly used deconvolution methods. However, the introduction of FFPE artifacts significantly impacted their performances, with a root mean squared error (RMSE) ranging between 0.04 and 0.17. Scaden, a deep learning-based method, consistently outperformed the others, demonstrating robustness against FFPE artifacts. Testing these methods on our 62-sample RNA-seq benign breast disease cohort in which cell type composition was estimated using digital pathology approaches, we found that pre-filtering of the reference data enhanced the accuracy of most methods, realizing up to a 32% reduction in RMSE. To support further research efforts in this domain, we introduce SCdeconR, an R package designed for streamlined cell type deconvolution assessments and downstream analyses.

Figure 1.

Overview of the study framework. The diagram illustrates three primary phases of our research: construction of scRNA-seq reference data for breast tissue; comparative analysis of deconvolution methods through simulation experiments; and evaluation of these methods on real data, encompassing both FFPE and FFzn samples (created with BioRender.com).

Figure 2.

Comparative analysis of deconvolution methods using simulated experiments. The benchmarking process entailed three distinct simulation scenarios: (i) baseline simulations; (ii) incomplete references, wherein specific cell types were omitted from the reference data; and (iii) FFPE artifact simulation, grounded on our proprietary FFPE–FFzn sample pair data. (A) RMSE evaluations for three representative methods from each category: MG-based, SC-based and DL-based. (B) Heatmap showing cell type correlations based on scRNA-seq reference data. (C) Aggregated RMSE metrics spanning all three experimental situations for all deconvolution methods.

Figure 3.

Performance assessment of deconvolution methods on real data from BBD FFPE samples. Predictions from each deconvolution methods were compared with CaseViewer annotations. (A) The analysis of gene-filtering stringency on the overall performance of deconvolution methods for all three cell types. Dot color indicates SCC with CaseViewer annotation. Dot size indicates the inverse of RMSE. The Y-axis represents the minimum requisite percentage of expressing cells per cell type. For example, a value of ‘0’ means no gene filtering was applied, while ‘40′ indicates that at least one cell type must have over 40% of its cells expressing the gene. (B) A scatterplot juxtaposing CaseViewer annotation and predictions from Scaden at 20% cutoff. Colors indicate the three major cell types evaluated. SCC and associated P-value were annotated at the top right. (C) Insight into each cell type’s correlation performance with CaseViewer estimates across deconvolution methods. Color and shape of the dot indicate SCC values for individual cell types. (D) Precision of cell type in terms of the inverse of RMSE. Color and shape of the dot indicate 1/RMSE values for individual cell types.

Figure 4.

Overview of SCdeconR. SCdeconR provide functionalities to (i) assemble integrated scRNA-seq reference data, (ii) facilitate the creation of simulated experiments for evaluation of deconvolution methods and (iii) perform cell type deconvolution and downstream analyses, including DEA and GSEA. For those interested in utilizing or exploring SCdeconR, it is currently accessible from CRAN at https://CRAN.R-project.org/package=SCdeconR and GitHub at https://github.com/Liuy12/SCdeconR (created with BioRender.com).