Run KNN prediction

Performs KNN prediction to annotate cell types based on reference scRNA-seq or bulk RNA-seq data.

Usage

RunKNNPredict(
  srt_query,
  srt_ref = NULL,
  bulk_ref = NULL,
  query_group = NULL,
  ref_group = NULL,
  query_assay = NULL,
  ref_assay = NULL,
  query_reduction = NULL,
  ref_reduction = NULL,
  query_dims = 1:30,
  ref_dims = 1:30,
  query_collapsing = !is.null(query_group),
  ref_collapsing = TRUE,
  return_full_distance_matrix = FALSE,
  features = NULL,
  features_type = c("HVF", "DE"),
  feature_source = "both",
  nfeatures = 2000,
  DEtest_param = list(max.cells.per.ident = 200, test.use = "wilcox"),
  DE_threshold = "p_val_adj < 0.05",
  nn_method = NULL,
  distance_metric = "cosine",
  k = 30,
  filter_lowfreq = 0,
  prefix = "KNNPredict",
  verbose = TRUE
)

Arguments

srt_query: An object of class Seurat to be annotated with cell types.
srt_ref: An object of class Seurat storing the reference cells.
bulk_ref: A cell atlas matrix, where cell types are represented by columns and genes are represented by rows. Either srt_ref or bulk_ref must be provided.
query_group: Column name in the srt_query metadata that represents the cell grouping.
ref_group: Column name in the srt_ref metadata that represents the cell grouping.
query_assay: Assay to be used for the query data. Default is the default assay of the srt_query object.
ref_assay: Assay to be used for the reference data. Default is the default assay of the srt_ref object.
query_reduction: Dimensionality reduction method used for the query data. If NULL, the function will use the default reduction method specified in the srt_query object.
ref_reduction: Dimensionality reduction method used for the reference data. If NULL, the function will use the default reduction method specified in the srt_ref object.
query_dims: Dimensions to be used for the query data. Default is the first 30 dimensions.
ref_dims: Dimensions to be used for the reference data. Default is the first 30 dimensions.
query_collapsing: A boolean value indicating whether the query data should be collapsed to group-level average expression values. If TRUE, the function will calculate the average expression values for each group in the query data and the annotation will be performed separately for each group. Otherwise it will use the raw expression values for each cell.
ref_collapsing: A boolean value indicating whether the reference data should be collapsed to group-level average expression values. If TRUE, the function will calculate the average expression values for each group in the reference data and the annotation will be performed separately for each group. Otherwise it will use the raw expression values for each cell.
return_full_distance_matrix: A boolean value indicating whether the full distance matrix should be returned. If TRUE, the function will return the distance matrix used for the KNN prediction, otherwise it will only return the annotated cell types.
features: Features to be used for the KNN prediction. If NULL, all the features in the query and reference data will be used.
features_type: Type of features to be used for the KNN prediction. Must be one of "HVF" (highly variable features) or "DE" (differentially expressed features).
feature_source: The source of the features to be used. Must be one of "both", "query", or "ref".
nfeatures: Maximum number of features to be used for the KNN prediction.
DEtest_param: A list of parameters to be passed to the differential expression test function if features_type is set to "DE". Default is list(max.cells.per.ident = 200, test.use = "wilcox").
DE_threshold: Threshold used to filter the DE features. If using "roc" test, DE_threshold should be needs to be reassigned. e.g. "power > 0.5".
nn_method: Nearest neighbor search method to use. Options are "raw", "annoy", and "rann". If "raw" is selected, the function will use the brute-force method to find the nearest neighbors. If "annoy" is selected, the function will use the Annoy library for approximate nearest neighbor search. If "rann" is selected, the function will use the RANN library for approximate nearest neighbor search. If not provided, the function will choose the search method based on the size of the query and reference datasets. For finite dense inputs using cosine or Euclidean distance, the raw path uses a bounded-memory compiled top-k kernel unless the full distance matrix is requested. Other metrics and sparse inputs retain the existing proxyC path.
distance_metric: Distance metric to be used for calculating similarity between cells. Must be one of "cosine", "euclidean", "manhattan", or "hamming".
k: A number of nearest neighbors to be considered for the KNN prediction.
filter_lowfreq: Threshold for filtering low-frequency cell types from the predicted results. Cell types with a frequency lower than filter_lowfreq will be labelled as "unreliable". Default is 0, which means no filtering will be performed.
prefix: Prefix to be added to the resulting annotations.
verbose: Whether to print the message. Default is TRUE.

Examples

data(panc8_sub)
keep <- panc8_sub$celltype %in% c("ductal", "alpha", "beta")
reference <- Seurat::NormalizeData(
  panc8_sub[, keep & panc8_sub$tech != "fluidigmc1"],
  verbose = FALSE
)
query <- Seurat::NormalizeData(
  panc8_sub[, keep & panc8_sub$tech == "fluidigmc1"],
  verbose = FALSE
)
genes <- head(intersect(rownames(reference), rownames(query)), 200)
query <- RunKNNPredict(
  query,
  srt_ref = reference, ref_group = "celltype",
  features = genes, nfeatures = length(genes),
  ref_collapsing = FALSE, k = 10, verbose = FALSE
)
query <- RunStandardWorkflow(query, verbose = FALSE, linear_reduction_dims = 10)
#> ℹ [2026-10-11 23:39:10] Skip `log1p()` because `layer = data` is not "counts"
CellDimPlot(
  query,
  group.by = "KNNPredict_classification",
  label = TRUE, legend.position = "bottom"
)

FeatureStatPlot(
  query,
  stat.by = "KNNPredict_prob",
  group.by = "KNNPredict_classification",
  plot_type = "box"
)

Usage

Arguments

See also

Examples