LRTests: Calculate likelihood ratio tests.

In lmweber/regsplice: L1-regularization based methods for detection of differential splicing

Calculate likelihood ratio tests.

Description

Calculate likelihood ratio tests between fitted models and null models.

Usage

LRTests(rs_results, when_null_selected = c("ones", "full", "NA"))

Arguments

rs_results	RegspliceResults object containing results generated by fitRegMultiple, fitNullMultiple, and fitFullMultiple. If when_null_selected = "ones" or "NA", the "full" models are not required. See RegspliceResults for details.
when_null_selected	Which option to use for genes where the lasso model selects zero interaction terms, i.e. identical to the null model. Options are "ones", "full", and "NA". Default is "ones". See below for details.

Details

The regularized (lasso) fitted models contain an optimal subset of exon:condition interaction terms for each gene, and the "full" fitted models contain all exon:condition interaction terms. The null models contain zero interaction terms, so they are nested within the fitted models.

The likelihood ratio (LR) tests compare the fitted models against the nested null models.

If the regularized (lasso) model contains at least one exon:condition interaction term, the LR test compares the lasso model against the null model. However, if the lasso model contains zero interaction terms, then the lasso and null models are identical, so the LR test cannot be calculated. The when_null_selected argument lets the user choose what to do in these cases: either set p-values equal to 1 (when_null_selected = "ones"); or calculate a LR test using the "full" model containing all exon:condition interaction terms (when_null_selected = "full"), which reduces power due to the larger number of terms, but allows the evidence for differential exon usage among these genes to be distinguished. You can also return NAs for these genes (when_null_selected = "NA").

The default option is when_null_selected = "ones". This simply calls all these genes non-significant, which in most cases is sufficient since we are more interested in genes with strong evidence for differential exon usage. However, if it is important to rank the low-evidence genes in your data set, use the when_null_selected = "full" option.

If when_null_selected = "ones" or when_null_selected = "NA", the "full" fitted models are not required.

Previous step: Fit models with fitRegMultiple, fitNullMultiple, and fitFullMultiple. Next step: Generate summary table of results with summaryTable.

Value

Returns a RegspliceResults object containing results of the LR tests. The results consist of the following entries for each gene:

• p_vals: raw p-values

• p_adj: multiple testing adjusted p-values (Benjamini-Hochberg false discovery rates, FDR)

• LR_stats: likelihood ratio test statistics

• df_tests: degrees of freedom of likelihood ratio tests

See Also

RegspliceResults initializeResults fitRegMultiple fitNullMultiple fitFullMultiple summaryTable

Examples

file_counts <- system.file("extdata/vignette_counts.txt", package = "regsplice")
data <- read.table(file_counts, header = TRUE, sep = "\t", stringsAsFactors = FALSE)
head(data)

counts <- data[, 2:7]
tbl_exons <- table(sapply(strsplit(data$exon, ":"), function(s) s[[1]]))
gene_IDs <- names(tbl_exons)
n_exons <- unname(tbl_exons)
condition <- rep(c("untreated", "treated"), each = 3)

rs_data <- RegspliceData(counts, gene_IDs, n_exons, condition)

rs_data <- filterZeros(rs_data)
rs_data <- filterLowCounts(rs_data)
rs_data <- runNormalization(rs_data)
rs_data <- runVoom(rs_data)

rs_results <- initializeResults(rs_data)

rs_results <- fitRegMultiple(rs_results, rs_data)
rs_results <- fitNullMultiple(rs_results, rs_data)
rs_results <- fitFullMultiple(rs_results, rs_data)

rs_results <- LRTests(rs_results)

lmweber/regsplice documentation built on March 19, 2024, 1:45 p.m.