Skip to content

Quickstart

The PBMC 3k guided-clustering workflow, end to end, next to the Seurat code it mirrors. Every Python line below was run to produce the output shown.

pip install "truecell[analysis]"

Load and build the object

The dataset downloads on first call into ~/.truecell_data/ (~24 MB). Loaders return the raw pieces — a counts matrix, gene names, cell names — and create_truecell_object assembles them, applying the same min_cells / min_features prefilter CreateSeuratObject does.

import truecell
from truecell.datasets import pbmc3k

counts, genes, cells = pbmc3k()
pbmc = truecell.create_truecell_object(
    counts=counts, feature_names=genes, cell_names=cells,
    project="pbmc3k", min_cells=3, min_features=200,
)
truecell.percentage_feature_set(pbmc, pattern=r"^MT-", col_name="percent.mt")

library(Seurat)

pbmc <- CreateSeuratObject(Read10X(data.dir), project = "pbmc3k",
                           min.cells = 3, min.features = 200)
pbmc[["percent.mt"]] <- PercentageFeatureSet(pbmc, pattern = "^MT-")

Quality control

md = pbmc.meta_data
keep = (md["nFeature_RNA"] > 200) & (md["nFeature_RNA"] < 2500) & (md["percent.mt"] < 5)
pbmc = pbmc.subset(cells=list(md.index[keep]))

pbmc <- subset(pbmc, subset = nFeature_RNA > 200 &
                              nFeature_RNA < 2500 & percent.mt < 5)

Both keep 2,638 of the 2,700 cells — the same 2,638 barcodes, not just the same count.

Normalize, select features, scale

truecell.normalize_data(pbmc, normalization_method="LogNormalize", scale_factor=10000)
truecell.find_variable_features(pbmc, selection_method="vst", nfeatures=2000)
truecell.scale_data(pbmc, features=truecell.generics.features(pbmc))

pbmc <- NormalizeData(pbmc, normalization.method = "LogNormalize",
                      scale.factor = 10000)
pbmc <- FindVariableFeatures(pbmc, selection.method = "vst", nfeatures = 2000)
pbmc <- ScaleData(pbmc, features = rownames(pbmc))

scale_data mutates in place

Seurat's functions return a modified object; truecell's write into the one you pass and return None. pbmc = truecell.normalize_data(pbmc) will leave you holding None — a difference worth internalising early. subset is the exception: it returns a new object, as it must.

find_variable_features picks 1,998 of the same 2,000 genes Seurat picks. The two that differ sit at the selection boundary, where the standardized variances agree to three decimals; see Fidelity.

Reduce, cluster, embed

truecell.run_pca(pbmc, n_pcs=50)
truecell.find_neighbors(pbmc, dims=range(10), k_param=20)
truecell.find_clusters(pbmc, resolution=0.5, algorithm=1, random_seed=0)
truecell.run_umap(pbmc, dims=range(10), seed=42)

pbmc <- RunPCA(pbmc, npcs = 50)
pbmc <- FindNeighbors(pbmc, dims = 1:10, k.param = 20, nn.method = "rann")
pbmc <- FindClusters(pbmc, resolution = 0.5, algorithm = 1, random.seed = 0)
pbmc <- RunUMAP(pbmc, dims = 1:10, seed.use = 42)

dims is 0-based here, 1-based in R

range(10) and 1:10 are the same ten PCs. This is the one indexing difference in the API, and it follows Python rather than R on purpose.

Use nn.method = \"rann\" when comparing against R

Seurat's default neighbour search is annoy, which is approximate; truecell's is exact. Leaving the default in place compares two different neighbour tables and reports a difference that belongs to annoy rather than to either implementation. That trap cost one of the verify scripts a false negative of 182 SNN edges.

clusters: [703, 480, 457, 344, 303, 162, 143, 32, 14]

Nine clusters, as Seurat finds on the same data, at ARI 0.928. find_clusters runs Seurat's own modularity optimiser, so the 32-cell dendritic-cell population separates here too.

Markers

markers = truecell.find_all_markers(
    pbmc, only_pos=True, min_pct=0.25, logfc_threshold=0.25,
)

markers <- FindAllMarkers(pbmc, only.pos = TRUE,
                          min.pct = 0.25, logfc.threshold = 0.25)

cluster   gene  avg_log2FC
      1 S100A8    6.642881
      1   FCN1    4.058240
      3  CD79A    6.911221
      3  MS4A1    5.718520
      8 ITGA2B   12.070139
      8 TMEM40   11.633066

find_markers runs all nine of Seurat's tests via test_use=. On a shared cell assignment, the eight that return a p-value reproduce Seurat's top 50 genes exactly and avg_log2FC agrees to 1.8e-15 — the DE vignette has the full table.

Plot

Every plotting function returns a matplotlib Figure.

fig = truecell.dim_plot(pbmc, reduction="umap", label=True)
fig.savefig("umap.png", dpi=150, bbox_inches="tight")

Next