Team lcolladotor Journal Club: Crumblr

Team lcolladotor Journal Club: Fast, flexible analysis of differences in cellular composition with crumblr
Gabriel E. Hoffman and Panos Roussos, bioRxiv, 2025
doi: https://doi.org/10.1101/2025.01.29.635498

Presented by: Louise Huuki-Myers, Feb 12, 2025
Recording of our journal club meeting: youtube.com/watch?v=Hi5Eg1rBRnc

Crumblr is a statistical method for testing differential composition analysis, such as testing for changes in cell type frequency across disease state. The method transforms frequency data with CLR and considers the precision of the cell fraction measurement in the modeling. The software package contains tools for variance partition, and both univariate and multivariate modeling.

Transcript

1. Team lcolladotor Journal Club: Fast, flexible analysis of differences in cellular composition with crumblr Gabriel E. Hoffman and Panos Roussos bioRxiv Jan, 2025 10.1101/2025.01.29.635498 Presented by: Louise Huuki-Myers, Feb 12, 2025
2. Problems with Modeling cell fractions • Modeling just the cell fraction (1000/5000 = 1/5) loses important info about precision ◦ Center log ratio (CLR) transformation ◦ Binomial models • Models considering precision - don't control false positives ◦ Poisson ◦ Negative binomial ◦ scCODA
3. count ratio uncertainty modeling based linear regression (crumblr) • Differential cellular composition, modeling variation in precision • Models fraction considering precision while controlling false positive rate Method • Transform cell counts w/ center log ratio (CLR) • Dirichlet-multinomial distribution estimates sampling variance • Univariate & multivariate testing Application • Apply crumblr to four single cell datasets
4. Fig 1: crumblr analysis workflow
5. Fig 2. Performance on simulated cellular compositions A. Models not including precision have low power a. Poisson & Binomial had high FPR B. Evaluating methods for multivariate testing a. Empirical Lin-Sullivan high-precision recall & low FPR
6. Fig 3 crumblr identifies compositional changes associated with aging in blood A. Variance partition a. High variation from batch/pool b. Small variation from sex c. Age: high variation in some (CD8+ αβ T cells) B. Effect Size (vs. age) a. “strongest decreases in frequency of CD8+ αβ T cells” C. Hierarchical clustering of gene expression a. Show effect size for nodes and leaves b. Point size indicates FDR, and ‘+’ indicates FDR < 5% D. CLR frequency CD8+ αβ T cells vs. age scatter plot a. Color shows error from low n cells b. Quadratic fit with (and without) crumblr measurement precision
7. Fig 4: crumblr identifies compositional changes associated with tuberculosis infection in T cell A. Variance partition, large contribution from age B. Effect Size (vs. TB) a. “CD4+ Th17 cluster as having the strongest estimated effect” C. Hierarchical clustering a. 4 Th1 / Th17 cell clusters together b. parent node of these cell types had significantly decreased frequency D. CLR frequency CD4+ Th17 in TB case/control a. Color shows error from low n cells
8. Fig 5: composition changes in bone metastases from prostate cancer A. Variance partition, large contribution from patient & disease status B. Effect Size (solid vs. involved) a. “Pericytes, osteoblasts and endothelial cells showed the strongest increase” C. Hierarchical clustering a. Pericytes, osteoblasts and endothelial cells: parent node increased b. Monocyte subsets parent node decreased frequency D. CLR frequency: Pericytes vs. disease state a. Color shows error from low n cells
9. Fig 6: Identify change from infection A. Variance partition, large contribution from Disease State a. Some contribution from age B. Effect Size (Severe COVID vs. healthy) a. “Decrease in 6 cell clusters, with the largest decreases in MAIT” b. 8 increase C. Hierarchical clustering a. “limited higher-order changes in cell type frequency” D. CLR frequency: Monocytes vs. COVID severity E. CLR frequency vs. cell types heatmap
10. Summary • Differential cellular composition testing considering precision & controlling False Positive rate • Tools for: ◦ Variance partition ◦ Univariate and multivariate testing over continuous or categorical variables ◦ * integrates with SingleCellExperiment • Useful at LIBD: ◦ case/control snRNA-seq datasets