scoring: Scoring of dimensionality reduction and clustering pipeline...

In vittoriofortino84/COPS: Clustering algorithms for Omics-driven Patient Stratification

Scoring of dimensionality reduction and clustering pipeline output

Description

Computes averages of metrics from pipeline output and also returns the best combination based on a weighted sum of metrics.

Usage

clusteval_scoring(...)

scoring(
  res,
  by = c("datname", "drname", "k", "m", "mkkm_mr_lambda"),
  wsum = TrainStabilityJaccard + Silhouette,
  significance_level = 0.05,
  summarise = TRUE,
  format_names = TRUE
)

get_best_result(res, scores)

Arguments

...	passed to scoring
res	result from COPS
by	character vector containing column names to group analysis by
wsum	an expression that indicates how a combined score is computed
significance_level	p-value cutoff for computing rejection rates
summarise	If FALSE, adds "run" and "fold" to by. By default the metrics are averaged across runs and folds.
format_names	If TRUE, formats internally used method names etc. to more user friendly names.
scores	scores from scoring

Details

Metrics are renamed for convenience:

• [Train/Test]Stability[Jaccard/ARI/NMI]

• [NMI/ARI/ChisqRR].<batch>

• [NMI/ARI].<subtype>

• ...

Value

list of all and best scores

Returns a list containing a data.frame $all of all scores and a single row $best with the best score according to wsum.

Functions

• clusteval_scoring(): Alias for scoring

• get_best_result(): Retrieves best clustering from CV results based on scores. In practice retrieves reference fold result from first run matching the best results.

Examples

library(COPS)

res <- COPS(ad_ge_micro_zscore, 
association_data = ad_studies, 
parallel = 1, nruns = 2, nfolds = 5, 
dimred_methods = c("pca", "umap", "tsne"), 
cluster_methods = c("hierarchical", "kmeans"), 
distance_metric = "euclidean",
n_clusters = 2:4)

scores <- scoring(res, wsum = Silhouette - GSE.nmi, summarise = TRUE)

best <- get_best_result(res, scores)
head(best$embedding)
head(best$clusters)

vittoriofortino84/COPS documentation built on Jan. 28, 2025, 3:16 p.m.