Cell-type composition

Does a condition change which cells a sample holds? propeller tests the proportion of each cell type across samples, as speckle does Phipson et al., 2022, from the number of cells that a cube records for every sample and cell type (Cube.n_cells); Composition tests gives the steps. The simulated PBMC study knows the answer: 12 donors, each sampled before and after a stimulation that multiplies the ratio of monocytes to every other cell type by $e^{0.7}$ , about 2, and changes nothing else (the simulated_pbmc dataset of the Datasets page).

import matplotlib.pyplot as plt

import cytepool
from cytepool.composition import propeller
from cytepool.design import Design
from cytepool.io import read_cube
from cytepool.plotting import plot_composition, plot_proportions

cube = read_cube("simulated_pbmc.cube.zarr")
print(f"CytePool {cytepool.__version__}")
print(cube)
plot_composition(cube, by="condition")
plt.show()

CytePool 0.1.dev54+g3f0cbb52e
Cube with 24 samples x 6 groups x 2,000 features
  layers: counts (uint32, lazy), detected (uint32, lazy), log_sum (float64, lazy), log_sum_sq (float64, lazy), linear_sum (float64, lazy)
  filled: 144 of 144 sample x group entries have cells (100.0%)
  samples: sample_id, donor, condition, sex
  groups: group_id
  features: feature_id, symbol

Stacked bars of the proportions of six cell types in each of the 24 samples, the 12 control samples on the left and the 12 stimulated ones on the right, B cells at the top and dendritic cells at the bottom. Monocytes (pink) take about 0.2 to 0.37 of each control sample and about 0.35 to 0.57 of each stimulated one, and every other cell type takes a little less after the stimulation.

Test every cell type¶

Each donor gives a sample in each condition, so the design pairs them, as for the genes (Genes that change in each cell type); propeller fits fixed effects, and the donor enters the formula.

design = Design(
    cube.samples,
    "~ donor + condition",
    levels={"condition": ["ctrl", "stim"]},
    contrasts={"stim_vs_ctrl": "conditionstim"},
)
result = propeller(cube, design)
table = result.table("stim_vs_ctrl")
for group, ratio, adjusted in zip(
    table["group_id"], table["proportion_ratio"], table["adjusted_p_value"]
):
    print(
        f"{group:<16} stimulated / control {ratio:.2f}, adjusted p-value {adjusted:.2g}"
    )
plot_proportions(result, by="condition")
plt.show()

B cells          stimulated / control 0.81, adjusted p-value 6.4e-05
CD4 T cells      stimulated / control 0.78, adjusted p-value 1.7e-07
CD8 T cells      stimulated / control 0.77, adjusted p-value 3.6e-07
NK cells         stimulated / control 0.78, adjusted p-value 1.2e-05
Monocytes        stimulated / control 1.60, adjusted p-value 4.6e-12
Dendritic cells  stimulated / control 0.81, adjusted p-value 0.0052

The proportion of each of six cell types in every sample, as points, controls in blue and stimulated samples in orange, with a black bar at each condition's mean and propeller's adjusted p-value above each cell type. Monocytes rise from a mean of about 0.27 to 0.43 (adjusted p-value 4.6e-12); B, CD4 T, CD8 T and NK cells fall a little, from about 0.12 to 0.10, 0.33 to 0.25, 0.15 to 0.11 and 0.12 to 0.09, each with an adjusted p-value below 1e-4; dendritic cells stay near 0.02 (adjusted p-value 0.0052).

Every cell type tests significant, although the simulation changed only the monocytes: proportions add up to one, so when the monocytes gain, the shares of all the others fall. The ratios show what happened: the five other cell types fall by about the same factor, 0.77 to 0.81, which is what a rise of the monocytes alone does to them.

Test within a lineage¶

The denominator decides the question. Among the lymphocytes alone, the B, T and NK cells, a rise of the monocytes does not move the shares, and the test asks whether the stimulation changes the lymphocytes’ make-up. Cube.select keeps those cell types.

lymphocytes = cube.select(groups=["B cells", "CD4 T cells", "CD8 T cells", "NK cells"])
within = propeller(lymphocytes, design)
plot_proportions(within, by="condition")
plt.show()

The proportions of B, CD4 T, CD8 T and NK cells among those four cell types in every sample, controls in blue and stimulated samples in orange, with a bar at each condition's mean: about 0.17, 0.45, 0.21 and 0.17 in both conditions, each with an adjusted p-value of 0.89.

Within the lymphocytes nothing changes, as simulated: the rise of the monocytes lowered the lymphocytes’ share of all cells, not their make-up.

The design and the transform¶

Without the pairing, the differences between donors’ compositions count as noise. The arcsine square root is the alternative transform to the logit.

unpaired = Design(
    cube.samples,
    "~ condition",
    levels={"condition": ["ctrl", "stim"]},
    contrasts={"stim_vs_ctrl": "conditionstim"},
)
tests = {
    "paired, logit": propeller(cube, design),
    "paired, asin": propeller(cube, design, transform="asin"),
    "unpaired, logit": propeller(cube, unpaired),
}
tables = {name: test.table("stim_vs_ctrl") for name, test in tests.items()}
print("adjusted p-values of the stimulation")
print(f"{'':<17}" + "".join(f"{name:>17}" for name in tables))
for row, group in enumerate(tables["paired, logit"]["group_id"]):
    values = "".join(
        f"{table['adjusted_p_value'][row]:>17.2g}" for table in tables.values()
    )
    print(f"{group:<17}{values}")

adjusted p-values of the stimulation
                     paired, logit     paired, asin  unpaired, logit
B cells                    6.4e-05          3.8e-06            0.083
CD4 T cells                1.7e-07          6.1e-16             0.05
CD8 T cells                3.6e-07          5.5e-09            0.079
NK cells                   1.2e-05          3.5e-07            0.086
Monocytes                  4.6e-12            3e-32          5.1e-06
Dendritic cells             0.0052            0.046             0.22

Pairing the samples matters more than the transform. Without it, the monocytes still stand out, but the falls of the other cell types are no longer clear, with adjusted p-values from 0.05 to 0.22. The arcsine square root gives smaller p-values than the logit here, and the same calls.

Choices¶

The design. Pair the samples of a donor when each donor gives one in every condition; propeller refuses random effects, so the donor is a fixed effect.
The denominator. Proportions are shares of the cell types tested: a change in one cell type moves the shares of all the others. Test the cell types of one lineage with Cube.select when the question is about that lineage, and read the proportion ratios of the result, not only its p-values.
The transform. The logit, speckle’s default, or transform="asin"; they agree on clear changes.
Several conditions. f_test="condition" tests every level of the condition at once, and plot_proportions then shows the F-test’s adjusted p-values.
Not yet available. Precision-weighted mixed models and count models of composition (Composition tests).

References¶

Phipson, B., Sim, C. B., Porrello, E. R., Hewitt, A. W., Powell, J., & Oshlack, A. (2022). propeller: testing for differences in cell type proportions in single cell data. Bioinformatics, 38(20), 4720–4726. 10.1093/bioinformatics/btac582