Last updated: 2026-09-10
Checks: 7 0
Knit directory:
immgenT-GP-analysis/analysis/
This reproducible R Markdown analysis was created with workflowr (version 1.7.2). The Checks tab describes the reproducibility checks that were applied when the results were created. The Past versions tab lists the development history.
Great! Since the R Markdown file has been committed to the Git repository, you know the exact version of the code that produced these results.
Great job! The global environment was empty. Objects defined in the global environment can affect the analysis in your R Markdown file in unknown ways. For reproduciblity it’s best to always run the code in an empty environment.
The command set.seed(1) was run prior to running the
code in the R Markdown file. Setting a seed ensures that any results
that rely on randomness, e.g. subsampling or permutations, are
reproducible.
Great job! Recording the operating system, R version, and package versions is critical for reproducibility.
Nice! There were no cached chunks for this analysis, so you can be confident that you successfully produced the results during this run.
Great job! Using relative paths to the files within your workflowr project makes it easier to run your code on other machines.
Great! You are using Git for version control. Tracking code development and connecting the code version to the results is critical for reproducibility.
The results in this page were generated with repository version dca7ffb. See the Past versions tab to see a history of the changes made to the R Markdown and HTML files.
Note that you need to be careful to ensure that all relevant files for
the analysis have been committed to Git prior to generating the results
(you can use wflow_publish or
wflow_git_commit). workflowr only checks the R Markdown
file, but you know if there are other scripts or data files that it
depends on. Below is the status of the Git repository when the results
were generated:
Ignored files:
Ignored: .DS_Store
Ignored: .claude/
Ignored: analysis/.DS_Store
Ignored: analysis/.Rhistory
Ignored: analysis/assets/.DS_Store
Ignored: captions/
Ignored: code/.DS_Store
Ignored: code/other/topic_flashier_20250212.R
Ignored: code/other/topic_wrapper_20250215_alldata_backfit.sh
Ignored: data
Ignored: experiments/
Ignored: figures/.DS_Store
Ignored: figures/final-selected/.DS_Store
Ignored: figures/final-selected/Figure 1/.DS_Store
Ignored: figures/final-selected/Figure 2/.DS_Store
Ignored: figures/final-selected/Figure 5/.DS_Store
Ignored: figures/final-selected/Figure S1/.DS_Store
Ignored: internal/
Ignored: log/
Ignored: output/.DS_Store
Ignored: output/Figure2/
Ignored: plan/
Ignored: tables/
Ignored: tmp/
Note that any generated files, e.g. HTML, png, CSS, etc., are not included in this status report because it is ok for generated content to have uncommitted changes.
These are the previous versions of the repository in which changes were
made to the R Markdown (analysis/Methods_FlashierFit.Rmd)
and HTML (docs/Methods_FlashierFit.html) files. If you’ve
configured a remote Git repository (see ?wflow_git_remote),
click on the hyperlinks in the table below to view the files as they
were in that past version.
| File | Version | Author | Date | Message |
|---|---|---|---|---|
| Rmd | dca7ffb | Ziang Zhang | 2026-09-10 | Strip the code-narrating prose from every page; fold code chunks by default |
| html | 54718c4 | Ziang Zhang | 2026-09-10 | Build site: the four mean-loading heatmaps without their titles |
| html | 936ed6e | Ziang Zhang | 2026-09-10 | Build site: Extended Data 2d without the DP lineage |
| html | 60125af | Ziang Zhang | 2026-09-10 | Build site: Figure 4b on the shared column colours |
| html | 88ee847 | Ziang Zhang | 2026-09-10 | Build site: Figure 4b without the miniverse clusters |
| html | 5874416 | Ziang Zhang | 2026-09-10 | Build site: main Figure 4 inserted, Extended Data back to 1-7 |
| html | a7a481f | Ziang Zhang | 2026-09-09 | Build site: the rebuilt Figure 1d and the Extended Data renumbering |
| html | 1e88d7e | Ziang Zhang | 2026-09-04 | Build site: published captions and titles across all 24 pages |
| html | 19c977f | Ziang Zhang | 2026-09-02 | Build site: Extended Data 5-7 renumbered, Figure S5 page rebuilt |
| html | dd71400 | Ziang Zhang | 2026-07-31 | Build site: Methods section on cell filtering by total GP membership |
| Rmd | 3f88d67 | Ziang Zhang | 2026-07-31 | Methods: document how L_pm_filtered / F_pm_filtered are obtained |
| html | cbcec52 | Ziang Zhang | 2026-07-30 | Build site: Extended Data Figure naming |
| html | ac650a0 | Ziang Zhang | 2026-07-30 | Build site: Figure S5 (ex-S6a) and Figure S6 as a-f |
| html | ae21d37 | Ziang Zhang | 2026-07-28 | Build site: republish after the reorder commits |
| html | d538aa2 | Ziang Zhang | 2026-07-28 | Build site: reordered Figures 6 / S6 / S3 and the new Figure 7b page |
| html | 029b0ae | Ziang Zhang | 2026-07-28 | Build site. |
| html | 3fc3789 | Ziang Zhang | 2026-07-27 | Republish all 24 pages |
| html | 1390a03 | Ziang Zhang | 2026-07-27 | Republish all 24 pages |
| html | adaef21 | Ziang Zhang | 2026-07-27 | Build site: panel fixes and PDF-derived assets |
| html | 5b19858 | Ziang Zhang | 2026-07-27 | Build site. |
| Rmd | 57b12e3 | Ziang Zhang | 2026-07-27 | Publish the Methods page rewrite; ignore tables/ |
| html | 57b12e3 | Ziang Zhang | 2026-07-27 | Publish the Methods page rewrite; ignore tables/ |
| Rmd | cba8dd4 | Ziang Zhang | 2026-07-12 | Add Methods page: Flashier matrix factorization |
| html | cba8dd4 | Ziang Zhang | 2026-07-12 | Add Methods page: Flashier matrix factorization |
This page documents the upstream matrix factorization step that
produces the gene programs (GPs) analyzed throughout this site, and the
cell filtering applied to its output. The code is shown for reference;
the fit itself was run on the UChicago RCC cluster, and its filtered
loading and factor matrices (L_pm_filtered /
F_pm_filtered) are the starting point for all downstream
analyses.
We fit an empirical Bayes matrix factorization (EBMF) model to the
full ImmGen-T scRNA-seq dataset using the flashier
R package. The model decomposes the cells × genes expression matrix into
a product of a loading matrix (cells × GPs) and a factor matrix (GPs ×
genes), learning up to 200 gene programs from the data.
libs = c("fastTopics", "flashier", "Matrix", "Seurat", "BPCells")
sapply(libs, function(x) suppressMessages(library(x, character.only = TRUE, quietly = TRUE)))
so = readRDS(file = path_to_seurat_object)Each cell’s counts are log-normalized using a cell-specific scale
factor equal to the dataset-wide mean library size
(mean(nCount_RNA)). This “shifted log” normalization keeps
the data on a comparable scale across cells while preserving count
structure.
norm_fac = mean(so$nCount_RNA)
so = NormalizeData(so, assay = "RNA", normalization.method = "LogNormalize", scale.factor = norm_fac)
counts = t(so[['RNA']]$counts)
shifted_log_counts = t(so[['RNA']]$data)Before factorization, we remove genes that would confound or dilute the biological signal:
Trbv, Trav,
Trgv, Trdv, …): highly variable but reflect
V(D)J recombination, not transcriptional programs.mt-): reflect
cell quality rather than biology.Rpl,
Rps, Mrpl, Mrps,
Rsl): ubiquitously expressed housekeeping genes.Gm…,
…Rik, …-ps): annotation artifacts with
unreliable quantification.tcr_genes = grepl(rownames(so), pattern = "Trbv|Trbd|Trbj|Trbc|Trav|Traj|Trac|Trgv|Trgd|Trgj|Trgc|Trdv|Trdj|Trdc")
gm_rik_genes = grepl(rownames(so), pattern = "Gm|Rik$|\\-ps$")
ribo_genes = grepl(rownames(so), pattern = "Rpl|Rps|Mrpl|Mrps|Rsl")
mt_genes = grepl(rownames(so), pattern = "^mt-")
genes_not_expressed = rowSums(so[["RNA"]]$counts) == 0
genes_to_keep = !(tcr_genes | gm_rik_genes | ribo_genes | mt_genes | genes_not_expressed)
shifted_log_counts = shifted_log_counts[, genes_to_keep]
counts = counts[, genes_to_keep]To set the noise level parameter S for
flashier, we estimate the standard deviation of
log-normalized Poisson noise. This gives a data-driven estimate of the
baseline technical noise floor, passed to flash() as the
S argument.
n = nrow(counts) # number of cells
x = rpois(1e7, 1/n) # Poisson draws at rate 1/cell
s1 = sd(log(x + 1)) # SD on the shifted-log scaleWe call flash() with prior families chosen to reflect
the expected structure of the loadings and factors:
ebnm_point_exponential for loadings
(cells): enforces non-negativity, so each cell’s contribution to a GP is
a non-negative weight.ebnm_point_laplace for factors
(genes): allows both positive and negative gene weights within a GP,
capturing genes that are up- or down-regulated together.var_type = 2: estimates a separate
residual variance for each gene (column-wise), accommodating the wide
range of expression variability across genes.backfit = TRUE: after the greedy
initialization adds up to 200 factors, a backfitting pass refines all
factors jointly to improve the overall fit.fit = flash(shifted_log_counts,
ebnm_fn = c(ebnm_point_exponential, ebnm_point_laplace),
var_type = 2,
S = s1,
backfit = TRUE,
greedy_Kmax = 200L)
saveRDS(fit, file = sprintf("%s/fit.Rds", output_dir))The next step reads flashier_snmf_summary.rds: the
loading matrix L_pm (cells x GPs), the factor matrix
F_pm (genes x GPs), the ELBO and the PVEs.
The loadings returned by flash() are on an arbitrary
per-GP scale, and a small number of cells load heavily on a great many
programs at once. One step handles both, and it produces the
L_pm_filtered / F_pm_filtered matrices that
every figure and table on this site actually uses.
Scaling. Each column of \(L\) is divided by its maximum, \(\tilde L = L\,\mathrm{diag}(1/d)\) with \(d_k = \max_i L_{ik}\), so that every GP’s loading runs from 0 to 1 and cells are comparable across GPs. A cell’s total membership is then its row sum \(\sum_k \tilde L_{ik}\) – roughly, how many fully-loaded programs’ worth of signal it carries.
Filtering. Cells whose total membership exceeds 10 are dropped. Removing the most extreme cells lowers each column’s maximum, which raises the remaining cells’ rescaled loadings and pushes a few more over the threshold, so the rescale-then-drop step is iterated – here for 12 passes, by which point it has converged.
Rescaling \(F\). The column maxima \(d\) are then recomputed on the retained cells and absorbed into the factor matrix, \(\tilde F = F\,\mathrm{diag}(d)\), so that \(\tilde L \tilde F^{T} = L F^{T}\): the reconstructed expression is unchanged, only the split of scale between loadings and factors moves.
Both helpers live in code/R/plot_utils.R:
scale_cols <- function(A, b) t(t(A) * b)
filter_cells_by_total_membership <- function(L, max_val = 10, numiter = 10) {
n <- nrow(L)
rows <- 1:n
for (iter in 1:numiter) {
x <- rowSums(L)
cat(sprintf("%d. Filtered out %d cells.\n", iter, sum(x > max_val)))
i <- which(x <= max_val)
L <- L[i, ]
rows <- rows[i]
d <- apply(L, 2, max)
L <- scale_cols(L, 1 / d)
}
return(rows)
}and the step itself is code/pipeline/01b_filter_cells.R:
flashier_snmf_summary <- readRDS(paste0(data_path, "flashier_snmf_summary.rds"))
seurat_meta <- readRDS(paste0(data_path, "igt1_96_withtotalvi20260206_clean_ADTonly.Rds"))@meta.data
cells_flashier <- rownames(flashier_snmf_summary$L_pm)
cells_seurat <- seurat_meta$cellID
# Keep only cells that are also present in the processed Seurat object.
L_pm <- flashier_snmf_summary$L_pm[cells_flashier %in% cells_seurat, ]
F_pm <- flashier_snmf_summary$F_pm
# Scale every GP to a maximum loading of 1, then iteratively drop cells whose
# total membership across GPs exceeds 10.
D <- diag(1 / apply(L_pm, 2, function(x) max(x)))
L <- L_pm %*% D
cells <- filter_cells_by_total_membership(L, numiter = 12)
# Re-apply the column scaling on the retained cells ...
L_pm_filtered <- L_pm[cells, ]
d <- apply(L_pm_filtered, 2, max)
L_pm_filtered <- scale_cols(L_pm_filtered, 1 / d)
# ... and absorb the same factors into F, leaving L %*% t(F) unchanged.
F_pm_filtered <- scale_cols(F_pm, d)
saveRDS(L_pm_filtered, paste0(data_path, "L_pm_filtered.rds"))
saveRDS(F_pm_filtered, paste0(data_path, "F_pm_filtered.rds"))Applied to this fit, flashier_snmf_summary.rds holds
682,953 cells and 19,805 genes over 200 GPs. Eighteen of its cells are
absent from the current Seurat object – the fit and the object are
different data vintages – so 682,935 cells enter the filter. The twelve
passes drop 69, 178, 157, 299, 232, 175, 193, 140, 46, 22, 1 and finally
0 cells: 1,512 cells in total, 0.22% of the data. What remains is
L_pm_filtered (681,423 cells x 200 GPs) and
F_pm_filtered (19,805 genes x 200 GPs).
These two matrices, together with the Seurat cell metadata subset to
rownames(L_pm_filtered), are the inputs loaded by
load_gp_data() (code/R/setup_data.R) at the
top of every figure script on this site.
sessionInfo()R version 4.5.1 (2025-06-13)
Platform: aarch64-apple-darwin20
Running under: macOS Sequoia 15.6.1
Matrix products: default
BLAS: /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/4.5-arm64/Resources/lib/libRlapack.dylib; LAPACK version 3.12.1
locale:
[1] en_CA/en_CA/en_CA/C/en_CA/en_CA
time zone: America/Chicago
tzcode source: internal
attached base packages:
[1] stats graphics grDevices utils datasets methods base
loaded via a namespace (and not attached):
[1] vctrs_0.7.3 cli_3.6.6 knitr_1.50 rlang_1.2.0
[5] xfun_0.55 stringi_1.8.9 otel_0.2.0 promises_1.5.0
[9] jsonlite_2.0.0 workflowr_1.7.2 glue_1.8.1 rprojroot_2.1.1
[13] git2r_0.36.2 htmltools_0.5.9 httpuv_1.6.16 sass_0.4.10
[17] rmarkdown_2.30 evaluate_1.0.5 jquerylib_0.1.4 tibble_3.3.0
[21] fastmap_1.2.0 yaml_2.3.12 lifecycle_1.0.5 whisker_0.4.1
[25] stringr_1.6.0 compiler_4.5.1 fs_1.6.6 Rcpp_1.1.1-1.1
[29] pkgconfig_2.0.3 later_1.4.4 digest_0.6.39 R6_2.6.1
[33] pillar_1.11.1 magrittr_2.0.5 bslib_0.9.0 tools_4.5.1
[37] cachem_1.1.0