Pseudocount Extraction for Vector Projection

#-----------------------------------------------------------------------#
# extracting pseudocounts for vector projection

# step 1: set-up and data input
# setting up the work environment
library(edgeR)
# setwd("~/Documents/Alex/Scripts and other coding/edgeR/counts")

# reading in files
files <- read.csv("temp_groups.csv")
# files <- read.table("temp_groups.tsv", sep = "\t", header = TRUE)
counts <- readDGE(files$Files, header=FALSE, labels = files$Summary, group = files$group)

# cleaning up meta tags
noint<-rownames(counts) %in% c("__no_feature", "__ambiguous", "__too_low_aQual", "__not_aligned", "__alignment_not_unique")
counts<-counts[!noint,]

# manual recalculation of library sizes
counts$samples$lib.size <- colSums(counts$counts)

#-----------------------------------------------------------------------#
# step 1.5: design set-up
# model design (additive model)
assays <- as.factor(files$Assay)
regime <- as.factor(files$Regime)
origin <- as.factor(files$Origin)

assays <- relevel(assays, "29")
regime <- relevel(regime, "Ancestral")
origin <- relevel(origin, "California")

design <- model.matrix(~0 + origin + assays + regime)
colnames(design) <- c("Brazil", "California", "Yemen", 
                      "Assay23", "Assay35", 
                      "RegimeCold", "RegimeHot", "RegimeSwi")

#-----------------------------------------------------------------------#
# step 2: filtering
# filtering out very low read counts
keep <- filterByExpr(counts, design) 

counts <- counts[keep, , keep.lib.sizes=FALSE]

# low CPM filtering + features in two samples or more filtering
countsPerMillion <- cpm(counts)

countCheck <- countsPerMillion > 3

keep <- which(rowSums(countCheck) >= 2)

counts <- counts[keep,]

#-----------------------------------------------------------------------#
# step 3: normalisation
counts <- calcNormFactors(object = counts)
counts <- normLibSizes(counts)


#-----------------------------------------------------------------------#
# step 4: dispersion estimation & model fitting
# "traditional" GLM approach
# first need some dispersions/BCVs
counts <- estimateGLMCommonDisp(counts, design)
counts <- estimateGLMTrendedDisp(counts, design)
counts <- estimateGLMTagwiseDisp(counts, design)

# then we can try and route around makeContrasts
fit <- glmQLFit(counts, design, robust = TRUE)


#-----------------------------------------------------------------------#
# step 5: pseudocount extraction and export
pseudocounts <- cpm(counts, normalized.lib.size=TRUE)

write.csv(pseudocounts, "pseudocounts.csv")