Conceitos essenciais
Comprehensive benchmarking of 11 cell type deconvolution methods using synthetic and real spatial transcriptomics data, revealing top-performing methods and key factors impacting their performance.
Resumo
The study presents a comprehensive benchmarking of 11 cell type deconvolution methods for spatial transcriptomics data. The authors developed a novel simulation engine called synthspot to generate 63 synthetic "silver standard" datasets with varying tissue patterns and cell type compositions. They also used 3 "gold standard" datasets from imaging-based spatial transcriptomics technologies. The methods were evaluated using root-mean-squared error (RMSE), area under the precision-recall curve (AUPR), and Jensen-Shannon divergence (JSD).
The key findings are:
- RCTD and cell2location were the top-performing methods across the synthetic and real datasets.
- Over half of the methods did not outperform the baseline non-negative least squares (NNLS) algorithm or the bulk deconvolution method MuSiC.
- Method performance was significantly impacted by the abundance pattern of the tissue, with dominant or rare cell types posing challenges.
- The choice of reference scRNA-seq dataset also greatly affected method stability and performance.
- The authors provide a reproducible Nextflow pipeline to generate synthetic data, run deconvolution methods, and benchmark them on user datasets.
The study highlights the importance of comprehensive benchmarking to identify the strengths and limitations of deconvolution methods, and provides guidelines for users to select appropriate methods for their spatial transcriptomics data.
Estatísticas
The number of genes per cell type in the reference dataset significantly impacts deconvolution performance.
Datasets with fewer than 1,000 genes per cell type generally performed worse than those with 3,000-9,000 genes per cell type.
Citações
"RCTD and cell2location were the top-performing methods across all metrics in the silver standards, followed by SpatialDWLS, stereoscope, and MuSiC."
"Most methods had worse performance in the two abundance patterns with a dominant cell type, and there was considerable performance variability between replicates due to different dominant cell types being selected in each replicate."
"Except for MuSiC, we see that methods with better performance–cell2location, RCTD, SpatialDWLS, and stereoscope–were also more stable against changing reference datasets."