--- title: "Scoring Functions" output: rmarkdown::html_document vignette: > %\VignetteIndexEntry{Scoring Functions} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r set.up, include=FALSE, echo=FALSE, message=FALSE, warning=FALSE} knitr::opts_chunk$set(message=FALSE, collapse = TRUE, comment="") # R packages library(SummarizedExperiment) library(pheatmap) library(devtools) load_all() ``` The **CaDrA** package currently supports four scoring functions to search for subsets of genomic features that are likely associated with a specific outcome of interest (e.g., protein expression, pathway activity, etc.) 1. Kolmogorov-Smirnov Method (`ks`) 2. Wilcoxon Rank-Sum Method (`wilcox`) 3. Conditional Mutual Information Method (`revealer`) 4. K-Nearest Neighbor Mutual Information Estimator (`knnmi`) 5. Correlation Method (`correlation`) 6. Custom - An User Defined Scoring Method (`custom`) Below, we run `candidate_search()` over the top 3 starting features using each of the scoring functions described above. **Important Notes:** - The legacy or deprecated function `topn_eval()` is equivalent to the new and recommended `candidate_search()` function # Load packages ```r library(CaDrA) library(pheatmap) library(SummarizedExperiment) ``` # Load required datasets 1. A `binary features matrix` also known as `Feature Set` (such as somatic mutations, copy number alterations, chromosomal translocations, etc.) The 1/0 row vectors indicate the presence/absence of ‘omics’ features in the samples. The `Feature Set` can be a matrix or an object of class **SummarizedExperiment** from **SummarizedExperiment** package) 2. A vector of continuous scores (or `Input Scores`) representing a functional response of interest (such as protein expression, pathway activity, etc.) ```{r load.data} # Load pre-computed feature set data(sim_FS) # Load pre-computed input scores data(sim_Scores) ``` # Heatmap of simulated feature set The simulated dataset, `sim_FS`, comprises of 1000 genomic features and 100 sample profiles. There are 10 left-skewed (i.e. True Positive or TP) and 990 uniformly-distributed (i.e. True Null or TN) features simulated in the dataset. Below is a heatmap of the first 100 features. ```{r heatmap} mat <- SummarizedExperiment::assay(sim_FS) pheatmap::pheatmap(mat[1:100, ], color = c("white", "red"), cluster_rows = FALSE, cluster_cols = FALSE) ``` # Search for a subset of genomic features that are likely associated with a functional response of interest using each of the scoring methods ## 1. Kolmogorov-Smirnov Scoring Method See `?ks_rowscore` for more details ```{r ks.method} ks_topn_l <- CaDrA::candidate_search( FS = sim_FS, input_score = sim_Scores, method = "ks_pval", # Use Kolmogorov-Smirnov scoring function method_alternative = "less", # Use one-sided hypothesis testing weights = NULL, # If weights is provided, perform a weighted-KS test search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the feature set of top N features that corresponded to the best scores over the top N search ks_topn_best_meta <- topn_best(ks_topn_l) # Visualize best meta-feature result meta_plot(topn_best_list = ks_topn_best_meta) ``` ## 2. Wilcoxon Rank-Sum Scoring Method See `?wilcox_rowscore` for more details ```{r wilcox.method} wilcox_topn_l <- CaDrA::candidate_search( FS = sim_FS, input_score = sim_Scores, method = "wilcox_pval", # Use Wilcoxon Rank-Sum scoring function method_alternative = "less", # Use one-sided hypothesis testing search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the feature set of top N feature that corresponded to the best scores over the top N search wilcox_topn_best_meta <- topn_best(topn_list = wilcox_topn_l) # Visualize best meta-feature result meta_plot(topn_best_list = wilcox_topn_best_meta) ``` ## 3. Conditional Mutual Information Scoring Method from REVEALER See `?revealer_rowscore` for more details ```{r revealer.method} revealer_topn_l <- CaDrA::candidate_search( FS = sim_FS, input_score = sim_Scores, method = "revealer", # Use REVEALER's CMI scoring function search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the ESet of top feature that corresponded to the best scores over the top N search revealer_topn_best_meta <- topn_best(topn_list = revealer_topn_l) # Visualize best meta-feature result meta_plot(topn_best_list = revealer_topn_best_meta) ``` ## 4. K-Nearest Neighbor Mutual Information Estimator from knnmi package See `?knnmi_rowscore` for more details ```{r knnmi.method} knnmi_topn_l <- CaDrA::candidate_search( FS = sim_FS, input_score = sim_Scores, method = "knnmi", # Use knnmi scoring function search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the ESet of top feature that corresponded to the best scores over the top N search knnmi_topn_best_meta <- topn_best(topn_list = knnmi_topn_l) # Visualize best meta-feature result meta_plot(topn_best_list = knnmi_topn_best_meta) ``` ## 5. Correlation Scoring Method See `?corr_rowscore` for more details ```{r correlation.method} corr_topn_l <- CaDrA::candidate_search( FS = SummarizedExperiment::assay(sim_FS), input_score = sim_Scores, method = "correlation", # Use correlation scoring function cmethod = "spearman", # Use spearman correlation scoring function top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the feature set of top N feature that corresponded to the best scores over the top N search corr_topn_best_meta <- topn_best(topn_list = corr_topn_l) # Visualize best meta-feature result meta_plot(topn_best_list = corr_topn_best_meta) ``` ## 6. Custom - An User Defined Scoring Method See `?custom_rowscore` for more details ```{r custom.method} # A customized function using ks-test customized_ks_rowscore <- function(FS, input_score, weights=NULL, meta_feature=NULL, alternative="less", metric="pval"){ metric <- match.arg(metric) alternative <- match.arg(alternative) # Check if meta_feature is provided if(!is.null(meta_feature)){ # Getting the position of the known meta features locs <- match(meta_feature, row.names(FS)) # Taking the union across the known meta features if(length(locs) > 1) { meta_vector <- as.numeric(ifelse(colSums(FS[locs,]) == 0, 0, 1)) }else{ meta_vector <- as.numeric(FS[locs, , drop=FALSE]) } # Remove the meta features from the binary feature matrix # and taking logical OR btw the remaining features with the meta vector FS <- base::sweep(FS[-locs, , drop=FALSE], 2, meta_vector, `|`)*1 # Check if there are any features that are all 1s generated from # taking the union between the matrix # We cannot compute statistics for such features and thus they need # to be filtered out if(any(rowSums(FS) == ncol(FS))){ verbose("Features with all 1s generated from taking the matrix union ", "will be removed before progressing...\n") FS <- FS[rowSums(FS) != ncol(FS), , drop=FALSE] # If no features remained after filtering, exist the function if(nrow(FS) == 0) return(NULL) } } # KS is a ranked-based method # So we need to sort input_score from highest to lowest values input_score <- sort(input_score, decreasing=TRUE) # Re-order the matrix based on the order of input_score FS <- FS[, names(input_score), drop=FALSE] # Check if weights is provided if(length(weights) > 0){ # Check if weights has any labels or names if(is.null(names(weights))) stop("The weights object must have names or labels that ", "match the labels of input_score\n") # Make sure its labels or names match the # the labels of input_score weights <- as.numeric(weights[names(input_score)]) } # Get the alternative hypothesis testing method alt_int <- switch(alternative, two.sided=0L, less=1L, greater=-1L, 1L) # Compute the ks statistic and p-value per row in the matrix ks <- .Call(ks_genescore_mat_, FS, weights, alt_int) # Obtain score statistics from KS method # Change values of 0 to the machine lowest value to avoid taking -log(0) stat <- ks[1,] # Obtain p-values from KS method # Change values of 0 to the machine lowest value to avoid taking -log(0) pval <- ks[2,] pval[which(pval == 0)] <- .Machine$double.xmin # Compute the scores according to the provided metric scores <- ifelse(rep(metric, nrow(FS)) %in% "pval", -log(pval), stat) names(scores) <- rownames(FS) return(scores) } # Search for best features using a custom-defined function custom_topn_l <- CaDrA::candidate_search( FS = SummarizedExperiment::assay(sim_FS), input_score = sim_Scores, method = "custom", # Use custom scoring function custom_function = customized_ks_rowscore, # Use a customized scoring function custom_parameters = NULL, # Additional parameters to pass to custom_function weights = NULL, # If weights is provided, perform a weighted test search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 3 starting points for the search max_size = 10, # Allow at most 10 features in meta-feature matrix do_plot = FALSE, # We will plot it AFTER finding the best hits best_score_only = FALSE # Return all results from the search ) # Now we can fetch the feature set of top N feature that corresponded to the best scores over the top N search custom_topn_best_meta <- topn_best(topn_list = custom_topn_l) # Visualize best meta-feature result CaDrA::meta_plot(topn_best_list = custom_topn_best_meta) # Evaluate results across top N features you started from CaDrA::topn_plot(custom_topn_l) ``` For validation purposes, compare the custom and built-in function. ```{r} topn_res <- CaDrA::candidate_search( FS = sim_FS, input_score = sim_Scores, method = "ks_pval", # Use Kolmogorov-Smirnov scoring function method_alternative = "less", # Use one-sided hypothesis testing weights = NULL, # If weights is provided, perform a weighted-KS test search_method = "both", # Apply both forward and backward search top_N = 3, # Evaluate top 7 starting points for each search max_size = 10, # Maximum size a meta-feature matrix can extend to do_plot = FALSE, # Plot after finding the best features best_score_only = FALSE # Return all results from the search ) ## Fetch the meta-feature set corresponding to its best scores over top N features searches topn_best_meta <- topn_best(topn_res) # Visualize the best results with the meta-feature plot meta_plot(topn_best_list = topn_best_meta) # Evaluate results across top N features you started from topn_plot(topn_res) ``` # SessionInfo ```{r RsessionInfo} sessionInfo() ```