Skip to contents

Run MELD signal smoothing for differential abundance estimation

meld_sc.Rd

This function implements MELD for estimating sample-associated density on a cell manifold. The general idea is to smooth a binary sample indicator matrix over the kNN graph via spectral filtering, yielding per-cell likelihood estimates for each sample condition. For further details on the method, please refer to Burkhardt, et al. This function will take a SingleCells class and return the smoothed density estimates per cell and condition.

Usage

meld_sc(
  object,
  sample_id_col,
  embd_to_use = "pca",
  no_embd_to_use = NULL,
  meld_params = params_meld(),
  seed = 42L,
  .verbose = TRUE
)

Arguments

object

SingleCells class.

sample_id_col

Character. The column in the obs table representing the sample identifier.

embd_to_use

Character. The embedding to use for kNN graph construction. Please use the same here as you used to generate the neighbours. Defaults to "pca".

no_embd_to_use

Optional integer. If you only want to use a subset of the embedding dimensions.

meld_params

A list, please see params_meld(). The list has the following parameters:

  • beta - Numeric. Smoothing strength; larger values produce smoother densities.

  • offset - Numeric. Shift of the filter centre in the rescaled spectrum. Must be in [0, 1].

  • order - Numeric. Filter falloff sharpness; larger values approach a square low-pass.

  • filter - Character. Filter family. One of c("heat", "laplacian").

  • chebyshev_order - Integer. Number of Chebyshev coefficients. Must be >= 2.

  • lap_type - Character. Type of Laplacian. One of c("combinatorial", "normalised").

  • normalise_indicators - Logical. If TRUE, indicator columns are divided by their column sum before filtering.

  • landmark - Logical. Whether to use landmark approximation. Useful for large data sets.

  • n_landmarks - Integer. Number of landmarks to use.

  • knn - List of kNN parameters. See params_knn_defaults() for available parameters and their defaults.

seed

Integer. Seed for reproducibility.

.verbose

Boolean or integer. Controls verbosity and returns run times. FALSE -> quiet, TRUE or 1L -> normal verbosity, 2L -> detailed verbosity.

Value

A list with:

  • raw_scores - The raw MELD scores

  • norm_scores - Negative values were clamped to 0 and the rows L1 normalised. This yields probability-like values.

References

Burkhardt, et al. Nat. Biotechnol., 2021.

Examples

# per cell sample likelihoods over the kNN graph
sc <- demo_single_cells(
  syn_data_params = params_sc_synthetic_data(
    n_cells = 500L,
    n_genes = 50L,
    n_samples = 6L,
    sample_bias = "even"
  )
)
res <- meld_sc(sc, sample_id_col = "sample_id", .verbose = FALSE)
head(res$norm_scores[, 1:3])
#>           sample_1  sample_2  sample_3
#> cell_001 0.1588947 0.1845287 0.1803174
#> cell_002 0.1761598 0.1402652 0.1614138
#> cell_003 0.1657891 0.1734398 0.1571076
#> cell_004 0.1586567 0.1845500 0.1815832
#> cell_005 0.1730026 0.1480340 0.1651779
#> cell_006 0.1658959 0.1731794 0.1563808

unlink(sc@dir_data, recursive = TRUE, force = TRUE)