tests/testthat/test_integration5.R

In satijalab/seurat: Tools for Single Cell Genomics

# Tests for IntegrateLayers
set.seed(42)


# checks that the absolute value of `x` and `y` are within `tolerance`
expect_abs_equal <- function(x, y, tolerance = 1.0e-04) {
  expect_equal(abs(x), abs(y), tolerance = tolerance)
}


# setup shared fixtures
# update `pbmc_small` to use `Assay5` instances
test.data <- pbmc_small
suppressWarnings(
  test.data[["RNA"]] <- CreateAssay5Object(
    counts = LayerData(
      test.data,
      assay = "RNA",
      layer = "counts"
    )
  )
)
# split the assay into multiple layers
test.data[["RNA"]] <- split(test.data[["RNA"]], f = test.data$groups)


context("IntegrateLayers")

# setup fixtures for standard integration workflow
test.data.std <- NormalizeData(test.data, verbose = FALSE)
test.data.std <- FindVariableFeatures(test.data.std, verbose = FALSE)
test.data.std <- ScaleData(test.data.std, verbose = FALSE)
test.data.std <- suppressWarnings(
  RunPCA(test.data.std, verbose = FALSE)
)

test_that("IntegrateLayers works with HarmonyIntegration", {
  skip_on_cran()
  skip_if_not_installed("harmony")

  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.std,
      method = HarmonyIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    0.3912
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.6668
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.7248
  )
})

test_that("IntegrateLayers works with CCAIntegration", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.std,
      method = CCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    0.9174
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    1.4885
    )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.5442
  )
})

test_that("IntegrateLayers works with RPCAIntegration", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.std,
      method = RPCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    0.1785
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.5832
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.5442
  )
})

test_that("IntegrateLayers works with JointPCAIntegration", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.std,
      method = JointPCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    0.4092
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.3246
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.5442
  )
})

test_that("IntegrateLayers fails when expected", {
  # an error should be raised if a v3 assay is passed in
  expect_error(
    IntegrateLayers(
      pbmc_small,
      method = CCAIntegration,
      orig.reduction = "pca",
      assay = "RNA",
      new.reduction = "integrated"
    )
  )

  # an error should be raised if a nonexistent `assay` is specified
  expect_error(
    IntegrateLayers(
      test.data.std,
      method = CCAIntegration,
      orig.reduction = "pca",
      assay = "DNA",
      new.reduction = "integrated"
    )
  )

  # an error should be raised if a nonexistent `orig.reduction` is specified
  expect_error(
    IntegrateLayers(
      test.data.std,
      method = CCAIntegration,
      orig.reduction = "lda",
      new.reduction = "integrated"
    )
  )
})


context("IntegrateData with SCTransform")

# setup fixtures for SCTransform workflow
test.data.sct <- suppressWarnings(
  SCTransform(
    test.data, 
    # use v1 to avoid potentially different
    # return values depending on if `glmGamPoi`
    # is installed or not
    vst.flavor="v1", 
    # set seed for reproducibility
    seed.use = 12345, 
    verbose = FALSE
  )
)
test.data.sct <- suppressWarnings(
  RunPCA(test.data.sct, verbose = FALSE)
)

test_that("IntegrateLayers works with HarmonyIntegration & SCTransform", {
  skip_on_cran()
  skip_if_not_installed("harmony")

  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.sct,
      method = HarmonyIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    1.1520
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    1.0302
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.1886
  )
})

test_that("IntegrateLayers works with CCAIntegration & SCTransform", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.sct,
      method = CCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10,
      # `CCAIntegration` needs to know how the data was normalized
      normalization.method = "SCT"
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    1.6113
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.6927
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.0855
  )
})

test_that("IntegrateLayers works with RPCAIntegration & SCTransform", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.sct,
      method = RPCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10,
      # `RPCAIntegration` needs to know how the data was normalized
      normalization.method = "SCT"
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    1.6492
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.7343
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.0855
  )
})

test_that("IntegrateLayers works with JointPCAIntegration & SCTransform", {
  integrated <- suppressWarnings(
    IntegrateLayers(
      test.data.sct,
      method = JointPCAIntegration,
      orig.reduction = "pca",
      new.reduction = "integrated",
      verbose = FALSE,
      # since `k.weight` must be less than the number of samples in the
      # smallest layer being integrated, it must be set to accommodate the
      # small dataset used for testing
      k.weight = 10,
      # `JointPCAIntegration` needs to know how the data was normalized
      normalization.method = "SCT"
    )
  )
  # the integrated reduction should have the same dimensions as the original 
  expect_equal(dim(integrated[["integrated"]]), dim(integrated[["pca"]]))
  # spot-check a few of the integrated values - since the integrated 
  # reductions sporadically flip sign only compare absolute values
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[5, 5],
    0.3427
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[40, 25],
    0.1015
  )
  expect_abs_equal(
    Embeddings(integrated[["integrated"]])[75, 45],
    0.0855
  )
})