README.md
In bioinformatist/ICE: ICE: Impute CircRNA Expression
ICE
ICE (Impute CircRNA Expression) is designed for imputing
circRNA expression using PCGs (protein-coding genes).
Installation
Currently bundled example data may be huge in size. One could try use
proxy and pull it first.
remotes::install_github("bioinformatist/ICE")
Quick start
library() package and load example data
library(ICE)
Currently data used for testing is too large to make a bundle. You
should follow example dataset part to prepare
them.
Performance
ML and DL on 871 circRNAs
library(ICE)
library(ggplot2)
library(data.table)
cv.full.pearson <- list(
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'KNN'),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'RF', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'EN', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'lasso', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'SVM', ncores = 56))
names(cv.full.pearson) <- c('KNN', 'RF', 'EN', 'lasso', 'SVM')
cv.full.pearson <- lapply(names(cv.full.pearson), function(x) data.table(cv.full.pearson[[x]], method = x))
cv.full.pearson <- do.call(rbind, cv.full.pearson)
ggplot() + geom_boxplot(data = cv.full.pearson, aes(x = method, y = CC, color = method))
ggplot() + geom_density(data = cv.full.pearson[method == 'RF'], aes(CC))
ggplot() + geom_density(data = cv.full.pearson[method == 'RF'], aes(CC)) + scale_x_continuous(breaks = seq(0.5, 0.8, 0.025), labels = seq(0.5, 0.8, 0.025), limits = c(0.4, 0.825))
all.CC <- rbind(cv.full.pearson[, .(CC, method)], data.table(DL, method = 'DL'))
require("reticulate")
source_python("pickle_reader.py")
DL <- read_pickle_file("cc_pancancer.pkl")
names(DL) <- 'CC'
all.CC <- rbind(cv.full.pearson[, .(CC, method)], data.table(DL, method = 'DL'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
DL <- read_pickle_file("cc_pancancer_30L_77e_val_mse.pkl")
names(DL) <- 'CC'
all.CC <- rbind(all.CC, data.table(DL, method = 'DL_val_mse'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
DL <- read_pickle_file("final.pkl")
names(DL) <- 'CC'
all.CC <- rbind(all.CC, data.table(DL, method = 'DL_final'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
ML and DL on 2672 circRNAs
cv.full.40.pearson <- list(
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'KNN'),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'RF', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'EN', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'lasso', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'SVM', ncores = 56))
names(cv.full.40.pearson) <- c('KNN', 'RF', 'EN', 'lasso', 'SVM')
cv.full.40.pearson <- lapply(names(cv.full.40.pearson), function(x) data.table(cv.full.40.pearson[[x]], method = x))
cv.full.40.pearson <- do.call(rbind, cv.full.40.pearson)
DL <- read_pickle_file("final_40.pkl")
names(DL) <- 'CC'
all.CC.40 <- rbind(cv.full.40.pearson[, .(CC, method)], data.table(DL, method = 'DL_40'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
Example dataset
Pan-cancer dataset
To perform further analysis on circRNAs, we use convert genome region
(Chromosome:Start-End) as circRNA identifiers.
One pair of circRNA and PCGs matrices is from paper The Landscape of
Circular RNA in
Cancer.
circRNA expression matrix was downloaded from this
website with the metadata and
pre-processed by below code:
library(data.table)
mionco.circ <- fread('v0.1.release.txt')[, `:=`(symbol = NULL, release = NULL, ID = paste0(chr, ':', start, '-', end))]
mionco.circ <- dcast(mionco.circ, formula = sample ~ ID, value.var = 'reads', fun.aggregate = max)
mionco.circ <- as.matrix(mionco.circ, rownames = 'sample')
mionco.pcg <- fread('fpkm_matrix.csv', drop = 1)
mionco.pcg[, Row.names := tstrsplit(Row.names, '\\.')[[1]]]
mionco.pcg[, V2 := NULL]
library(biomaRt)
mart <- useMart(biomart="ensembl", dataset="hsapiens_gene_ensembl")
pcg.id <- getBM(attributes = 'ensembl_gene_id', filters = "biotype", values = "protein_coding", mart = mart)
setnames(mionco.pcg, 'Row.names', 'ensembl_gene_id')
mionco.pcg <- mionco.pcg[pcg.id, on = 'ensembl_gene_id', nomatch = NULL]
mionco.pcg <- as.matrix(mionco.pcg, rownames = 'ensembl_gene_id')
mionco.pcg <- t(mionco.pcg)
sample.used <- intersect(rownames(mionco.circ), rownames(mionco.pcg))
mionco.circ <- mionco.circ[sample.used, ]
mionco.pcg <- mionco.pcg[sample.used, ]
CCLE dataset
The CCLE circRNAs expression was downloaded from CircRic
database, while the RNAseq gene
expression as well as sample (cell line) annotation was downloaded from
CCLE portal of The Broad
Institute. As CCLE data
is processed with hg19, we also need prepare a chain
file
for
liftOver.
ccle.circ <- fread('circRNA_expression.csv')[, `:=`(ccl_id = NULL, cancer_type = NULL, circ_gene = NULL)]
anno <- fread('Cell_lines_annotations_20181226.txt')
anno <- anno[, .(CCLE_ID, Name)]
ccle.circ <- merge(ccle.circ, anno, by.x = 'ccl_name', by.y = 'Name')[, ccl_name := NULL]
splitted <- tstrsplit(ccle.circ[[1]], '_|\\|')
temp.19 <- GRanges(seqnames = paste0('chr', splitted[[1]]), ranges = IRanges(start = as.numeric(splitted[[2]]), end = as.numeric(splitted[[3]])), name = ccle.circ[['CCLE_ID']], count = ccle.circ[['number_bsReads']])
chain <- import.chain("hg19ToHg38.over.chain")
temp.38 <- liftOver(temp.19, chain)
temp.38 <- as.data.table(unlist(temp.38))
ccle.circ <- data.table(ID = paste0(temp.38[['seqnames']], ':', temp.38[['start']], '-', temp.38[['end']]), name = temp.38[['name']], count = temp.38[['count']])
ccle.circ <- dcast(ccle.circ, formula = name ~ ID, value.var = 'count', fun.aggregate = max)
ccle.circ <- as.matrix(ccle.circ, rownames = 'name')
ccle.pcg <- fread('CCLE_RNAseq_rsem_genes_tpm_20180929.txt')
ccle.pcg[, `:=`(transcript_ids = NULL, gene_id = tstrsplit(gene_id, '\\.')[[1]])]
mart <- useMart(biomart="ensembl", dataset="hsapiens_gene_ensembl")
pcg.id <- getBM(attributes = 'ensembl_gene_id', filters = "biotype", values = "protein_coding", mart = mart)
setnames(ccle.pcg, 'gene_id', 'ensembl_gene_id')
ccle.pcg <- ccle.pcg[pcg.id, on = 'ensembl_gene_id', nomatch = NULL]
ccle.pcg <- as.matrix(ccle.pcg, rownames = 'ensembl_gene_id')
ccle.pcg <- t(ccle.pcg)
sample.used <- intersect(rownames(ccle.circ), rownames(ccle.pcg))
ccle.circ <- ccle.circ[sample.used, ]
ccle.pcg <- ccle.pcg[sample.used, ]
Up to now, mionco.circ and mionco.pcg could be used as circRNA and
PCG dataset for training or testing, and the ccle.circ and ccle.pcg
ditto.
Cross-dataset validation
# Keep circRNAs expressed in at least 1% sample
# 非常惨QAQ
train.circ <- ICE:::filter_circ(mionco.circ, 1)
new.circ <- ICE:::filter_circ(ccle.circ, 1)
train.pcg <- ICE:::pre_process(mionco.pcg)
new.pcg <- ICE:::pre_process(ccle.pcg)
# 取完交集就剩一点点了
used.circ.name <- intersect(colnames(new.circ), colnames(train.circ))
fwrite(train.circ[, used.circ.name], 'train_circ.csv')
fwrite(new.circ[, used.circ.name], 'new_circ.csv')
fwrite(train.pcg, 'train_pcg.csv')
fwrite(new.pcg, 'new_pcg.csv')
sfInit(parallel = TRUE, cpu = 56)
cl <- sfGetCluster()
sfExport('train.pcg', 'train.circ', 'new.pcg')
sfExport('ICE', namespace = 'ICE')
pboptions(use_lb = TRUE)
predicted <- pblapply(used.circ.name, function(x) ICE(train.pcg, train.circ, new.pcg, x, method = 'RF', filter = FALSE), cl = cl)
names(predicted) <- used.circ.name
cd.cc <- unlist(lapply(used.circ.name, function(x) cor(predicted[[x]], ccle.circ[, x], method = 'spearman')))
meta <- fread('meta_update.csv')
meta[, count := .N, by = 'Analysis Cohort']
head(unique(meta, by = 'Analysis Cohort')[order(-count)])
fwrite(ICE:::filter_circ(mionco.circ), 'mionco_circ.csv')
fwrite(ICE:::pre_process(mionco.pcg), 'mionco_pcg.csv')
cohort_cor <- function(x){
cohort.sample.used <- intersect(sample.used, meta[x, on = 'Analysis Cohort', ID])
if(length(cohort.sample.used) == 0) return()
circ.cohort <- mionco.circ[cohort.sample.used, ]
pcg.cohort <- mionco.pcg[cohort.sample.used, ]
cv <- list(
ICE_cv_entire(train.pcg = pcg.cohort, train.circ = circ.cohort, cor.method = 'pearson', folds = 3, method = 'KNN'),
ICE_cv_entire(train.pcg = pcg.cohort, train.circ = circ.cohort, cor.method = 'pearson', folds = 3, method = 'RF'))
names(cv) <- c('KNN', 'RF')
cv <- lapply(names(cv), function(x) data.table(cv[[x]], method = x))
cv <- do.call(rbind, cv)
(data.table(cv, cohort = x))
}
cv.full <- lapply(head(unique(meta, by = 'Analysis Cohort')[order(-count)])[['Analysis Cohort']], function(x) cohort_cor(x))
names(cv.full) <- head(unique(meta, by = 'Analysis Cohort')[order(-count)])[['Analysis Cohort']]
cv.full <- Filter(function(x) !is.null(x), cv.full)
cv.full <- do.call(rbind, cv.full)
ggplot() + geom_density(data = cv.full[method == 'RF' & cohort == 'PRAD'], aes(CC))
d_fun <- ecdf (cv.full[method == 'RF' & cohort == 'PRAD', CC])
ggplot(data = cv.full, aes(x = cohort, y = CC, fill = method)) + geom_boxplot() + geom_jitter(aes(fill = method), shape=16, position=position_jitterdodge(0.2))
m6A
library(data.table)
ip <- fread('rpkm_ip_norm_m.txt')
anno <- fread('fix.anno.txt')
Tuning for static ML parameters
library(mlr3)
library(data.table)
# install.packages("mlr3learners.randomforest", repos = "https://mlr3learners.github.io/mlr3learners.drat")
library(mlr3learners.randomforest)
library(paradox)
library(mlr3tuning)
train.circ <- scale(ICE:::filter_circ(mionco.circ))
train.pcg <- ICE:::pre_process(mionco.pcg)
train.pcg <- scale(ICE:::corf(train.pcg, train.circ, "chr7:23611170-23611553", 50))
task.circ <- TaskRegr$new(id = 'circ', backend = data.table(target = train.circ[, "chr7:23611170-23611553"], train.pcg), target = "target")
Tuning for random forest
learner.circ <- lrn('regr.randomForest')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamInt$new('ntree', lower = 1, upper = 500),
ParamInt$new('mtry', lower = 1, upper = 33),
ParamInt$new('nodesize', lower = 1, upper = 50)
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for SVM
learner.circ <- lrn('regr.svm')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamDbl$new('cost', lower = 0.0001, upper = 10000),
ParamDbl$new('gamma', lower = 0.01, upper = 1),
ParamFct$new('type', 'eps-regression'),
ParamFct$new('kernel', 'radial')
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for KNN
learner.circ <- lrn('regr.fnn')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamInt$new('k', lower = 1, upper = task.circ$nrow - 100)
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for lasso regression
library(glmnet)
cv.lasso <- cv.glmnet(train.pcg, train.circ[, "chr7:23611170-23611553"], alpha = 1, family = 'gaussian')
cv.lasso$lambda.min
Tuning for Elastic Net
library(caret)
en.model <- suppressWarnings(train(
train.pcg, train.circ[, "chr7:23611170-23611553"], method = 'glmnet',
trControl = trainControl("cv", number = 10),
tuneLength = 100))
en.model$bestTune
bioinformatist/ICE documentation built on July 5, 2020, 12:20 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
ICE
ICE (Impute CircRNA Expression) is designed for imputing circRNA expression using PCGs (protein-coding genes).
Installation
Currently bundled example data may be huge in size. One could try use proxy and pull it first.
remotes::install_github("bioinformatist/ICE")
Quick start
library() package and load example data
library(ICE)
Currently data used for testing is too large to make a bundle. You should follow example dataset part to prepare them.
Performance
ML and DL on 871 circRNAs
library(ICE)
library(ggplot2)
library(data.table)
cv.full.pearson <- list(
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'KNN'),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'RF', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'EN', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'lasso', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', method = 'SVM', ncores = 56))
names(cv.full.pearson) <- c('KNN', 'RF', 'EN', 'lasso', 'SVM')
cv.full.pearson <- lapply(names(cv.full.pearson), function(x) data.table(cv.full.pearson[[x]], method = x))
cv.full.pearson <- do.call(rbind, cv.full.pearson)
ggplot() + geom_boxplot(data = cv.full.pearson, aes(x = method, y = CC, color = method))
ggplot() + geom_density(data = cv.full.pearson[method == 'RF'], aes(CC))
ggplot() + geom_density(data = cv.full.pearson[method == 'RF'], aes(CC)) + scale_x_continuous(breaks = seq(0.5, 0.8, 0.025), labels = seq(0.5, 0.8, 0.025), limits = c(0.4, 0.825))
all.CC <- rbind(cv.full.pearson[, .(CC, method)], data.table(DL, method = 'DL'))
require("reticulate")
source_python("pickle_reader.py")
DL <- read_pickle_file("cc_pancancer.pkl")
names(DL) <- 'CC'
all.CC <- rbind(cv.full.pearson[, .(CC, method)], data.table(DL, method = 'DL'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
DL <- read_pickle_file("cc_pancancer_30L_77e_val_mse.pkl")
names(DL) <- 'CC'
all.CC <- rbind(all.CC, data.table(DL, method = 'DL_val_mse'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
DL <- read_pickle_file("final.pkl")
names(DL) <- 'CC'
all.CC <- rbind(all.CC, data.table(DL, method = 'DL_final'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
ML and DL on 2672 circRNAs
cv.full.40.pearson <- list(
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'KNN'),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'RF', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'EN', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'lasso', ncores = 56),
ICE_cv_entire(train.pcg = mionco.pcg, train.circ = mionco.circ, cor.method = 'pearson', cutoff = 40, method = 'SVM', ncores = 56))
names(cv.full.40.pearson) <- c('KNN', 'RF', 'EN', 'lasso', 'SVM')
cv.full.40.pearson <- lapply(names(cv.full.40.pearson), function(x) data.table(cv.full.40.pearson[[x]], method = x))
cv.full.40.pearson <- do.call(rbind, cv.full.40.pearson)
DL <- read_pickle_file("final_40.pkl")
names(DL) <- 'CC'
all.CC.40 <- rbind(cv.full.40.pearson[, .(CC, method)], data.table(DL, method = 'DL_40'))
ggplot() + geom_density(data = all.CC, aes(CC, color = method))
Example dataset
Pan-cancer dataset
To perform further analysis on circRNAs, we use convert genome region (Chromosome:Start-End) as circRNA identifiers.
One pair of circRNA and PCGs matrices is from paper The Landscape of Circular RNA in Cancer.
circRNA expression matrix was downloaded from this website with the metadata and pre-processed by below code:
library(data.table)
mionco.circ <- fread('v0.1.release.txt')[, `:=`(symbol = NULL, release = NULL, ID = paste0(chr, ':', start, '-', end))]
mionco.circ <- dcast(mionco.circ, formula = sample ~ ID, value.var = 'reads', fun.aggregate = max)
mionco.circ <- as.matrix(mionco.circ, rownames = 'sample')
mionco.pcg <- fread('fpkm_matrix.csv', drop = 1)
mionco.pcg[, Row.names := tstrsplit(Row.names, '\\.')[[1]]]
mionco.pcg[, V2 := NULL]
library(biomaRt)
mart <- useMart(biomart="ensembl", dataset="hsapiens_gene_ensembl")
pcg.id <- getBM(attributes = 'ensembl_gene_id', filters = "biotype", values = "protein_coding", mart = mart)
setnames(mionco.pcg, 'Row.names', 'ensembl_gene_id')
mionco.pcg <- mionco.pcg[pcg.id, on = 'ensembl_gene_id', nomatch = NULL]
mionco.pcg <- as.matrix(mionco.pcg, rownames = 'ensembl_gene_id')
mionco.pcg <- t(mionco.pcg)
sample.used <- intersect(rownames(mionco.circ), rownames(mionco.pcg))
mionco.circ <- mionco.circ[sample.used, ]
mionco.pcg <- mionco.pcg[sample.used, ]
CCLE dataset
The CCLE circRNAs expression was downloaded from CircRic database, while the RNAseq gene expression as well as sample (cell line) annotation was downloaded from CCLE portal of The Broad Institute. As CCLE data is processed with hg19, we also need prepare a chain file for liftOver.
ccle.circ <- fread('circRNA_expression.csv')[, `:=`(ccl_id = NULL, cancer_type = NULL, circ_gene = NULL)]
anno <- fread('Cell_lines_annotations_20181226.txt')
anno <- anno[, .(CCLE_ID, Name)]
ccle.circ <- merge(ccle.circ, anno, by.x = 'ccl_name', by.y = 'Name')[, ccl_name := NULL]
splitted <- tstrsplit(ccle.circ[[1]], '_|\\|')
temp.19 <- GRanges(seqnames = paste0('chr', splitted[[1]]), ranges = IRanges(start = as.numeric(splitted[[2]]), end = as.numeric(splitted[[3]])), name = ccle.circ[['CCLE_ID']], count = ccle.circ[['number_bsReads']])
chain <- import.chain("hg19ToHg38.over.chain")
temp.38 <- liftOver(temp.19, chain)
temp.38 <- as.data.table(unlist(temp.38))
ccle.circ <- data.table(ID = paste0(temp.38[['seqnames']], ':', temp.38[['start']], '-', temp.38[['end']]), name = temp.38[['name']], count = temp.38[['count']])
ccle.circ <- dcast(ccle.circ, formula = name ~ ID, value.var = 'count', fun.aggregate = max)
ccle.circ <- as.matrix(ccle.circ, rownames = 'name')
ccle.pcg <- fread('CCLE_RNAseq_rsem_genes_tpm_20180929.txt')
ccle.pcg[, `:=`(transcript_ids = NULL, gene_id = tstrsplit(gene_id, '\\.')[[1]])]
mart <- useMart(biomart="ensembl", dataset="hsapiens_gene_ensembl")
pcg.id <- getBM(attributes = 'ensembl_gene_id', filters = "biotype", values = "protein_coding", mart = mart)
setnames(ccle.pcg, 'gene_id', 'ensembl_gene_id')
ccle.pcg <- ccle.pcg[pcg.id, on = 'ensembl_gene_id', nomatch = NULL]
ccle.pcg <- as.matrix(ccle.pcg, rownames = 'ensembl_gene_id')
ccle.pcg <- t(ccle.pcg)
sample.used <- intersect(rownames(ccle.circ), rownames(ccle.pcg))
ccle.circ <- ccle.circ[sample.used, ]
ccle.pcg <- ccle.pcg[sample.used, ]
Up to now, mionco.circ and mionco.pcg could be used as circRNA and
PCG dataset for training or testing, and the ccle.circ and ccle.pcg
ditto.
Cross-dataset validation
# Keep circRNAs expressed in at least 1% sample
# 非常惨QAQ
train.circ <- ICE:::filter_circ(mionco.circ, 1)
new.circ <- ICE:::filter_circ(ccle.circ, 1)
train.pcg <- ICE:::pre_process(mionco.pcg)
new.pcg <- ICE:::pre_process(ccle.pcg)
# 取完交集就剩一点点了
used.circ.name <- intersect(colnames(new.circ), colnames(train.circ))
fwrite(train.circ[, used.circ.name], 'train_circ.csv')
fwrite(new.circ[, used.circ.name], 'new_circ.csv')
fwrite(train.pcg, 'train_pcg.csv')
fwrite(new.pcg, 'new_pcg.csv')
sfInit(parallel = TRUE, cpu = 56)
cl <- sfGetCluster()
sfExport('train.pcg', 'train.circ', 'new.pcg')
sfExport('ICE', namespace = 'ICE')
pboptions(use_lb = TRUE)
predicted <- pblapply(used.circ.name, function(x) ICE(train.pcg, train.circ, new.pcg, x, method = 'RF', filter = FALSE), cl = cl)
names(predicted) <- used.circ.name
cd.cc <- unlist(lapply(used.circ.name, function(x) cor(predicted[[x]], ccle.circ[, x], method = 'spearman')))
meta <- fread('meta_update.csv')
meta[, count := .N, by = 'Analysis Cohort']
head(unique(meta, by = 'Analysis Cohort')[order(-count)])
fwrite(ICE:::filter_circ(mionco.circ), 'mionco_circ.csv')
fwrite(ICE:::pre_process(mionco.pcg), 'mionco_pcg.csv')
cohort_cor <- function(x){
cohort.sample.used <- intersect(sample.used, meta[x, on = 'Analysis Cohort', ID])
if(length(cohort.sample.used) == 0) return()
circ.cohort <- mionco.circ[cohort.sample.used, ]
pcg.cohort <- mionco.pcg[cohort.sample.used, ]
cv <- list(
ICE_cv_entire(train.pcg = pcg.cohort, train.circ = circ.cohort, cor.method = 'pearson', folds = 3, method = 'KNN'),
ICE_cv_entire(train.pcg = pcg.cohort, train.circ = circ.cohort, cor.method = 'pearson', folds = 3, method = 'RF'))
names(cv) <- c('KNN', 'RF')
cv <- lapply(names(cv), function(x) data.table(cv[[x]], method = x))
cv <- do.call(rbind, cv)
(data.table(cv, cohort = x))
}
cv.full <- lapply(head(unique(meta, by = 'Analysis Cohort')[order(-count)])[['Analysis Cohort']], function(x) cohort_cor(x))
names(cv.full) <- head(unique(meta, by = 'Analysis Cohort')[order(-count)])[['Analysis Cohort']]
cv.full <- Filter(function(x) !is.null(x), cv.full)
cv.full <- do.call(rbind, cv.full)
ggplot() + geom_density(data = cv.full[method == 'RF' & cohort == 'PRAD'], aes(CC))
d_fun <- ecdf (cv.full[method == 'RF' & cohort == 'PRAD', CC])
ggplot(data = cv.full, aes(x = cohort, y = CC, fill = method)) + geom_boxplot() + geom_jitter(aes(fill = method), shape=16, position=position_jitterdodge(0.2))
m6A
library(data.table)
ip <- fread('rpkm_ip_norm_m.txt')
anno <- fread('fix.anno.txt')
Tuning for static ML parameters
library(mlr3)
library(data.table)
# install.packages("mlr3learners.randomforest", repos = "https://mlr3learners.github.io/mlr3learners.drat")
library(mlr3learners.randomforest)
library(paradox)
library(mlr3tuning)
train.circ <- scale(ICE:::filter_circ(mionco.circ))
train.pcg <- ICE:::pre_process(mionco.pcg)
train.pcg <- scale(ICE:::corf(train.pcg, train.circ, "chr7:23611170-23611553", 50))
task.circ <- TaskRegr$new(id = 'circ', backend = data.table(target = train.circ[, "chr7:23611170-23611553"], train.pcg), target = "target")
Tuning for random forest
learner.circ <- lrn('regr.randomForest')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamInt$new('ntree', lower = 1, upper = 500),
ParamInt$new('mtry', lower = 1, upper = 33),
ParamInt$new('nodesize', lower = 1, upper = 50)
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for SVM
learner.circ <- lrn('regr.svm')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamDbl$new('cost', lower = 0.0001, upper = 10000),
ParamDbl$new('gamma', lower = 0.01, upper = 1),
ParamFct$new('type', 'eps-regression'),
ParamFct$new('kernel', 'radial')
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for KNN
learner.circ <- lrn('regr.fnn')
# learner.circ$param_set
tune.ps <- ParamSet$new(list(
ParamInt$new('k', lower = 1, upper = task.circ$nrow - 100)
))
instance <- TuningInstance$new(
task = task.circ,
learner = learner.circ,
resampling = rsmp('cv'),
measures = msr('regr.mse'),
param_set = tune.ps,
terminator = term("stagnation", threshold = 1e-5)
)
tuner <- tnr("grid_search", resolution = 10)
result <- tuner$tune(instance)
Tuning for lasso regression
library(glmnet)
cv.lasso <- cv.glmnet(train.pcg, train.circ[, "chr7:23611170-23611553"], alpha = 1, family = 'gaussian')
cv.lasso$lambda.min
Tuning for Elastic Net
library(caret)
en.model <- suppressWarnings(train(
train.pcg, train.circ[, "chr7:23611170-23611553"], method = 'glmnet',
trControl = trainControl("cv", number = 10),
tuneLength = 100))
en.model$bestTune
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.