Mercurial > repos > fubar > rglasso_1_9_8
diff rglasso_cox.xml @ 20:bb725f6d6d38 draft default tip
planemo upload for repository https://github.com/galaxyproject/tools-iuc/tree/master/tools/rglasso commit 344140b8df53b8b7024618bb04594607a045c03a
| author | iuc |
|---|---|
| date | Mon, 04 May 2015 22:47:29 -0400 |
| parents | 0e87f636bdd8 |
| children |
line wrap: on
line diff
--- a/rglasso_cox.xml Wed Apr 29 12:07:11 2015 -0400 +++ b/rglasso_cox.xml Mon May 04 22:47:29 2015 -0400 @@ -7,223 +7,9 @@ <requirement type="package" version="2.14">glmnet_lars_2_14</requirement> </requirements> <command interpreter="python"> - rgToolFactory.py --script_path "$runme" --interpreter "Rscript" --tool_name "rglasso" + rgToolFactory.py --script_path "$runme" --interpreter "Rscript" --tool_name "rglasso" --output_dir "$html_file.files_path" --output_html "$html_file" --make_HTML "yes" </command> - <inputs> - <param name="title" type="text" value="lasso test" size="80" label="Title for job outputs" help="Typing a short, meaningful text here will help remind you (and explain to others) what the outputs represent"> - <sanitizer invalid_char=""> - <valid initial="string.letters,string.digits"><add value="_" /> </valid> - </sanitizer> - </param> - <param name="input1" type="data" format="tabular" label="Select an input tabular text file from your history. Rows represent samples; Columns are measured phenotypes" - multiple='False' optional="False" help="Tabular text data with samples as rows, phenotypes as columns with a header row of column identifiers" /> - <param name="xvar_cols" label="Select columns containing numeric variables to use as predictor (x) variables" type="data_column" data_ref="input1" numerical="False" - multiple="True" use_header_names="True" force_select="True" /> - <param name="force_xvar_cols" label="Select numeric columns containing variables ALWAYS included as predictors in cross validation" type="data_column" data_ref="input1" numerical="False" - multiple="True" use_header_names="True" force_select="False"/> - <conditional name="model"> - <param name="fam" type="select" label="GLM Link function for models" - help="Binary dependant variables will automatically be set to Binomial no matter what this is set to"> - <option value="gaussian" selected="true">Gaussian - continuous dependent (y)</option> - <option value="binomial">Binomial dependent variables</option> - <option value="poisson">Poisson (eg counts)</option> - <option value="cox">Cox models - require special setup for y variables - see below</option> - </param> - <when value="gaussian"> - <param name="yvar_cols" label="Select numeric columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="False" - multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> - <param name="output_full" type="hidden" value='F' /> - <param name="output_pred" type="hidden" value='F' /> - <param name="cox_id" label="Select column containing a unique sample identifier" - help = "Only really needed for output sample specific predicted values downstream." - type="data_column" data_ref="input1" numerical="False" force_select="True" - multiple="False" use_header_names="True" /> - </when> - <when value="binomial"> - <param name="yvar_cols" label="Select numeric columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="False" - multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> - <param name="output_full" type="hidden" value='F' /> - <param name="output_pred" type="select" label="Create a tabular output with predicted values for each subject from the optimal model for (eg) NRI estimates" > - <option value="F" selected="true">No predicted value output file</option> - <option value="T">Create a predicted value output file</option> - </param> - <param name="cox_id" label="Select column containing a unique sample identifier" - help = "Only really needed for output sample specific predicted values downstream." - type="data_column" data_ref="input1" numerical="False" force_select="True" - multiple="False" use_header_names="True" /> - <param name="predict_at" type="hidden" value='' /> - - </when> - <when value="poisson"> - <param name="yvar_cols" label="Select columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="True" - multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> - <param name="output_full" type="hidden" value='F' /> - <param name="output_pred" type="hidden" value='F' /> - <param name="predict_at" type="hidden" value='' /> - <param name="cox_id" label="Select column containing a unique sample identifier" - help = "Optional. Only really needed for output sample specific predicted values downstream. Free - enjoy" - type="data_column" data_ref="input1" numerical="True" force_select="False" - multiple="False" use_header_names="True" /> - </when> - <when value="cox"> - <param name="cox_time" label="Select column containing time under observation for Cox regression" - type="data_column" data_ref="input1" numerical="True" force_select="True" - multiple="False" use_header_names="True" help = "This MUST contain a time period - eg continuous years or days to failure or right censoring"/> - <param name="cox_status" label="Select column containing status = 1 for outcome of interest at the end of the time under observation or 0 for right censoring" - type="data_column" data_ref="input1" numerical="True" force_select="True" - multiple="False" use_header_names="True" help = "This MUST contain 1 for subjects who had an event at that time or 0 for a right censored observation"/> - <param name="cox_id" label="Select column containing a unique sample identifier" - help = "Optional. Only really needed for output sample specific predicted values downstream. Free - enjoy" - type="data_column" data_ref="input1" numerical="False" force_select="False" - multiple="False" use_header_names="True" /> - <param name="output_full" type="select" label="Create a tabular output with coefficients for all predictors" > - <option value="F" selected="true">No full model output file</option> - <option value="T">Create a full model output file</option> - </param> - <param name="output_pred" type="select" label="Create a tabular output with predicted values for each subject from the optimal model for (eg) NRI estimates" > - <option value="F" selected="true">No predicted value output file</option> - <option value="T">Create a predicted value output file</option> - </param> - <param name="predict_at" type="text" value='' label="Provide a comma separated list of times to make a prediction for each subject" - optional="True" help="Default (blank) will return predictions at 0%,25%,50%,75%,100% of the observed times which should be informative" /> - - </when> - </conditional> - <param name="optLambda" type="select" label="Value to use when reporting optimal model and coefficients" help="minLambda will have more predictors - 1SDLambda will be more parsimonious"> - <option value="lambda.1se" selected="true">Lambda + 1 SE of min MSE or AUC (fewer coefficients - more false negatives)</option> - <option value="lambda.min">Lambda at min MSE or max AUC (more coefficients - more false positives)</option> - </param> - <param name="logxform_cols" optional="True" label="Select numeric columns to be log transformed before use as predictors or dependent variables" type="data_column" - data_ref="input1" numerical="True" multiple="True" use_header_names="True" help = "The wisdom of doing this depends entirely on your predictors - eg can help diminish long-tailed outlier influence" - force_select="False"/> - <param name="do_standard" type="select" label="Standardise x vars" - help="If all measurements on same scale, may not be needed. Coefficients are always returned on the original scale."> - <option value="False" selected="true">No standardisation of predictors</option>l - <option value="True">Standardise predictors before model</option> - </param> - <param name="mdsplots" type="select" label="Generate MDS plots of samples in measurement space and measurements in sample space" > - <option value="False" selected="true">No MDS plots</option>l - <option value="True">Yes create MDS plots</option> - </param> - <param name="alpha" type="float" value="0.95" size="5" min="0.01" max="1.0" label="Alpha - see glmnet docs. 1 for pure lasso. 0.0 for pure ridge regression" - help="Default 0.95 allows lasso to cope better with expected predictor collinearity. Use (eg) 0.5 for hybrid regularised regression or (eg) 0.025 for ridge regression"/> - <param name="nfold" type="integer" value="10" size="5" label="Number of folds for internal cross validation" - help="Default of 10 is usually ok"/> - </inputs> - <outputs> - <data format="html" name="html_file" label="${title}.html"/> - <data format="tabular" name="model_file" label="${title}_modelres.xls"/> - <data format="tabular" name="output_full_file" label="${title}_full_cox_model.xls"> - <filter>model['output_full'] == 'T'</filter> - </data> - <data format="tabular" name="output_pred_file" label="${title}_predicted_from_model.xls"> - <filter>model['output_pred'] == 'T'</filter> - </data> - </outputs> - <tests> - <test> - <param name='input1' value='cox_test.xls' ftype='tabular' /> - <param name='treatment_name' value='case' /> - <param name='title' value='Cox glmnet test' /> - <param name='nfold' value='10' /> - <param name='logxform_cols' value='' /> - <param name='alpha' value='0.95' /> - <param name='do_standard' value="True" /> - <param name='cox_time' value='1' /> - <param name='cox_status' value='2' /> - <param name='cox_id' value='1' /> - <param name='predict_at' value='' /> - <param name='fam' value='cox' /> - <param name='yvar_cols' value='' /> - <param name='xvar_cols' value='3,4,5' /> - <param name='force_xvar_cols' value='3' /> - <param name='output_full' value='F' /> - <param name='output_pred' value='F' /> - <output name='model_file' file='coxlassotest_modelres.xls'> - <assert_contents> - <has_text text="rhubarb" /> - <has_text text="TRUE" /> - <!-- 	 is XML escape code for tab --> - <!-- has_line line="regulator	partial_likelihood	forced_in	glmnet_model	best_lambda" / --> - <has_line line="regulator	partial_likelihood	forced_in	glmnet_model	best_lambda	lambdaChoice	alpha" /> - <has_n_columns n="7" /> - </assert_contents> - </output> - <output name='html_file' file='coxlassotest.html' compare='diff' lines_diff='16' /> - </test> -</tests> -<help> - -**Before you start** - -Please read the glmnet documentation @ glmnet_ - -This Galaxy wrapper merely exposes that code and the glmnet_ documentation is essential reading -before getting useful results here. - -**What it does** - -From documentation at glmnet_ :: - - Glmnet is a package that fits a generalized linear model via penalized maximum likelihood. - The regularization path is computed for the lasso or elasticnet penalty at a grid of values for the regularization parameter lambda. - The algorithm is extremely fast, and can exploit sparsity in the input matrix x. - It fits linear, logistic and multinomial, poisson, and Cox regression models. - A variety of predictions can be made from the fitted models. - -Internal cross validation is used to optimise the choice of lambda based on CV AUC for logistic (binomial outcome) models, or CV mse for gaussian. - -**Warning about the tyrany of dimensionality** - -Yes, this package will select 'optimal' models even when you (optimistically) supply more predictors than you have cases. -The model returned is unlikely to represent the only informative regularisation path through your data - if you run repeatedly with -exactly the same settings, you will probably see many different models being selected. -This is not a software bug - the real problem is that you just don't have enough information in your data. - -Sufficiently big jobs will take a while (eg each lasso regression with 20k features on 1k samples takes about 2-3 minutes on our aged cluster) - -**Input** - -Assuming you have more measurements than samples, you supply data as a tabular text file where each row is a sample and columns -are variables. You specify which columns are dependent (predictors) and which are observations for each sample. Each of multiple -dependent variable columns will be run and reported independently. Predictors can be forced in to the model. - -**Output** - -For each selected dependent regression variable, a brief report of the model coefficients predicted at the -'optimal' nfold CV value of lambda. - -**Predicted event probabilities for Cox and Logistic models** - -If you want to compare (eg) two competing clinical predictions, there's a companion generic NRI tool -for predicted event probabilities. Estimates dozens of measures of improvement in prediction. Currently only works for identical id subjects -but can probably be extended to independent sample predictions. - -Given a model, we can generate a predicted p (for status 1) in binomial or cox frameworks so models can be evaluated in terms of NRI. -Of course, estimates are likely substantially inflated over 'real world' performance by being estimated from the same sample - but you probably -already knew that since you were smart enough to reach this far down into the on screen help. The author salutes you, intrepid reader! - -It may seem an odd thing to do, but we can predict p for an event for each subject from our original data, given a parsimonious model. Doing -this for two separate models (eg, forcing in an additional known explanatory measurement to the new model) allows comparison of the two models -predicted status for each subject, or the same model in independent populations to see how badly it does - -**Attributions** - -glmnet_ is the R package exposed by this Galaxy tool. - -Galaxy_ (that's what you are using right now!) for gluing everything together - -Otherwise, all code and documentation comprising this tool was written by Ross Lazarus and is -licensed to you under the LGPL_ like other rgenetics artefacts - -.. _LGPL: http://www.gnu.org/copyleft/lesser.html -.. _glmnet: http://web.stanford.edu/~hastie/glmnet/glmnet_alpha.html -.. _Galaxy: http://getgalaxy.org - - -</help> - <configfiles> <configfile name="runme"> <![CDATA[ @@ -236,27 +22,27 @@ message=function(x) {print.noquote(paste(x,sep=''))} -ross.cv.glmnet = function (x, y, weights, offset = NULL, lambda = NULL, type.measure = c("mse", - "deviance", "class", "auc", "mae"), nfolds = 10, foldid, - grouped = TRUE, keep = FALSE, parallel = FALSE, ...) +ross.cv.glmnet = function (x, y, weights, offset = NULL, lambda = NULL, type.measure = c("mse", + "deviance", "class", "auc", "mae"), nfolds = 10, foldid, + grouped = TRUE, keep = FALSE, parallel = FALSE, ...) { - if (missing(type.measure)) + if (missing(type.measure)) type.measure = "default" else type.measure = match.arg(type.measure) - if (!is.null(lambda) && length(lambda) < 2) + if (!is.null(lambda) && length(lambda) < 2) stop("Need more than one value of lambda for cv.glmnet") N = nrow(x) - if (missing(weights)) + if (missing(weights)) weights = rep(1, N) else weights = as.double(weights) y = drop(y) glmnet.call = match.call(expand.dots = TRUE) - sel = match(c("type.measure", "nfolds", "foldid", "grouped", + sel = match(c("type.measure", "nfolds", "foldid", "grouped", "keep"), names(glmnet.call), F) - if (any(sel)) + if (any(sel)) glmnet.call = glmnet.call[-sel] glmnet.call[[1]] = as.name("glmnet") - glmnet.object = glmnet(x, y, weights = weights, offset = offset, + glmnet.object = glmnet(x, y, weights = weights, offset = offset, lambda = lambda, ...) glmnet.object\$call = glmnet.call is.offset = glmnet.object\$offset @@ -266,56 +52,56 @@ nz = sapply(nz, function(x) sapply(x, length)) nz = ceiling(apply(nz, 1, median)) } - else nz = sapply(predict(glmnet.object, type = "nonzero"), + else nz = sapply(predict(glmnet.object, type = "nonzero"), length) - if (missing(foldid)) + if (missing(foldid)) foldid = sample(rep(seq(nfolds), length = N)) else nfolds = max(foldid) - if (nfolds < 3) + if (nfolds < 3) stop("nfolds must be bigger than 3; nfolds=10 recommended") outlist = as.list(seq(nfolds)) if (parallel && require(foreach)) { - outlist = foreach(i = seq(nfolds), .packages = c("glmnet")) %dopar% + outlist = foreach(i = seq(nfolds), .packages = c("glmnet")) %dopar% { sel = foldid == i - if (is.matrix(y)) + if (is.matrix(y)) y_sub = y[!sel, ] else y_sub = y[!sel] - if (is.offset) + if (is.offset) offset_sub = as.matrix(offset)[!sel, ] else offset_sub = NULL - glmnet(x[!sel, , drop = FALSE], y_sub, lambda = lambda, - offset = offset_sub, weights = weights[!sel], + glmnet(x[!sel, , drop = FALSE], y_sub, lambda = lambda, + offset = offset_sub, weights = weights[!sel], ...) } } else { for (i in seq(nfolds)) { sel = foldid == i - if (is.matrix(y)) + if (is.matrix(y)) y_sub = y[!sel, ] else y_sub = y[!sel] - if (is.offset) + if (is.offset) offset_sub = as.matrix(offset)[!sel, ] else offset_sub = NULL - outlist[[i]] = glmnet(x[!sel, , drop = FALSE], - y_sub, lambda = lambda, offset = offset_sub, + outlist[[i]] = glmnet(x[!sel, , drop = FALSE], + y_sub, lambda = lambda, offset = offset_sub, weights = weights[!sel], ...) } } fun = paste("cv", class(glmnet.object)[[1]], sep = ".") - cvstuff = do.call(fun, list(outlist, lambda, x, y, weights, + cvstuff = do.call(fun, list(outlist, lambda, x, y, weights, offset, foldid, type.measure, grouped, keep)) cvm = cvstuff\$cvm cvsd = cvstuff\$cvsd cvname = cvstuff\$name - out = list(lambda = lambda, cvm = cvm, cvsd = cvsd, cvup = cvm + + out = list(lambda = lambda, cvm = cvm, cvsd = cvsd, cvup = cvm + cvsd, cvlo = cvm - cvsd, nzero = nz, name = cvname, glmnet.fit = glmnet.object) - if (keep) + if (keep) out = c(out, list(fit.preval = cvstuff\$fit.preval, foldid = foldid)) - lamin = if (type.measure == "auc") + lamin = if (type.measure == "auc") getmin(lambda, -cvm, cvsd) else getmin(lambda, cvm, cvsd) out = c(out, as.list(lamin)) @@ -342,7 +128,7 @@ mdsPlot = function(dm,myTitle,groups=NA,outpdfname,transpose=T) { - + samples = colnames(dm) mt = myTitle pcols=c('maroon') @@ -355,11 +141,11 @@ mydata = dm if (transpose==T) { - mydata = t(dm) + mydata = t(dm) } npred = ncol(mydata) - d = dist(mydata) - fit = cmdscale(d,eig=TRUE, k=min(10,npred-2)) + d = dist(mydata) + fit = cmdscale(d,eig=TRUE, k=min(10,npred-2)) xmds = fit\$points[,1] ymds = fit\$points[,2] pdf(outpdfname) @@ -388,7 +174,7 @@ formstring=paste("y ~",cn) form = as.formula(formstring) ok = complete.cases(x) - + if (sum(ok) < length(ok)) { x = x[ok,] yvec = yvec[ok] @@ -408,9 +194,6 @@ return(p1) } - - - getpredp_cox = function(x,time,status,id,predict_at) { cols = colnames(x) @@ -419,13 +202,13 @@ return(NA) } cn = paste(colnames(x), collapse = ' + ') - + formstring=paste("Surv(time, status) ~",cn) - + form = as.formula(formstring) - + ok = complete.cases(x) - + if (sum(ok) < length(ok)) { x = x[ok,] time = time[ok] @@ -444,7 +227,6 @@ } - dolasso_cox = function(x,y,debugOn=F,maxsteps=10000,nfold=10,xcolnames,ycolnames,optLambda='lambda.1se',out_full=F,out_full_file=NA, out_pred=F,out_pred_file=NA,cox_id=NA, descr='Cox test',do_standard=F,alpha=0.9,penalty,predict_at,mdsplots=F) { @@ -457,7 +239,7 @@ if (class(p) == "try-error") { print.noquote(paste('Unable to produce predictors in sample space mds plot',p)) - } + } outpdfname = 'cox_samples_in_x_space_MDS.pdf' p = try({mdsPlot(x,'samples in measurement space',groups=y,outpdfname=outpdfname,transpose=F) },T) if (class(p) == "try-error") @@ -468,7 +250,7 @@ if (is.na(predict_at)) { predict_at = quantile(y) } message(paste('@@@ Cox model will be predicted at times =',paste(predict_at,collapse=','))) do_standard = do_standard - standardize = do_standard + standardize = do_standard normalize = do_standard p = try({larsres = glmnet(x,y,family='cox',standardize=standardize,alpha=alpha,penalty.factor=penalty )},T) if (class(p) == "try-error") @@ -492,13 +274,13 @@ try( { pdf(outpdf) - plot(larsres,main='cox glmnet',label=T) + plot(larsres,main='cox glmnet',label=T) grid() dev.off() },T) - + larscv = NA - + p = try({larscv=ross.cv.glmnet(x,y,family=fam,type.measure='deviance',penalty=penalty)},T) if (class(p) == "try-error") { print.noquote(paste('Unable to cross validate your data',p)) @@ -588,23 +370,23 @@ sink() return(NA) } - + mt = paste('Glmnet fraction deviance for',target) outpdf = paste(target,'glmnetPath.pdf',sep='_') pdf(outpdf) plot(larsres,main=mt,label=T) grid() dev.off() - + outpdf = paste(target,'glmnetDeviance.pdf',sep='_') - + mt2 = paste('Glmnet lambda for',target) - + pdf(outpdf) plot(larsres,xvar="lambda",main=mt2,label=T) grid() dev.off() - + larscv = NA if (fam=="binomial") { tmain = paste(target,'AUC') @@ -621,13 +403,12 @@ sink() return(NA) } - - + pdf(outpdf) plot(larscv,main=tmain) grid() dev.off() - + lse = larscv\$cvhits.1se lmin = larscv\$cvhits.min tot = lse + lmin @@ -637,7 +418,7 @@ print.noquote(nzhits) out_nz_file = paste(target,'cross_validation_model_counts.xls',sep='_') write.table(nzhits,out_nz_file,quote=FALSE, sep="\t",row.names=F) - + ipenalty = c(0,penalty) if (optLambda == 'lambda.min') { best_lambda = larscv\$lambda.min @@ -747,7 +528,7 @@ corPlot=function(xdat=c(),main='main title',is_raw=T) { - library(pheatmap) + library(pheatmap) library(gplots) if (is_raw) { cxdat = cor(xdat,method="spearman",use="pairwise.complete.obs") @@ -762,17 +543,15 @@ } - - runTest = function(n=10) -{ +{ set.seed (NULL) Y = data.frame(y1=runif (n),y2=runif(n)) Xv <- runif(n*n) X <- matrix(Xv, nrow = n, ncol = n) - + mydf <- data.frame(Y, X) - + regres_out = dolasso_generic(predvars=X,depvars=Y,debugOn=T,p.cutoff = 0.05,maxsteps=10000,nfold=10, descr='randomdata',do_standard=do_standard,defaultFam="gaussian",alpha=0.05) return(regres_out) @@ -900,6 +679,218 @@ </configfile> </configfiles> + <inputs> + <param name="title" type="text" value="lasso test" size="80" label="Title for job outputs" help="Typing a short, meaningful text here will help remind you (and explain to others) what the outputs represent"> + <sanitizer invalid_char=""> + <valid initial="string.letters,string.digits"><add value="_" /> </valid> + </sanitizer> + </param> + <param name="input1" type="data" format="tabular" label="Select an input tabular text file from your history. Rows represent samples; Columns are measured phenotypes" + multiple='False' optional="False" help="Tabular text data with samples as rows, phenotypes as columns with a header row of column identifiers" /> + <param name="xvar_cols" label="Select columns containing numeric variables to use as predictor (x) variables" type="data_column" data_ref="input1" numerical="False" + multiple="True" use_header_names="True" force_select="True" /> + <param name="force_xvar_cols" label="Select numeric columns containing variables ALWAYS included as predictors in cross validation" type="data_column" data_ref="input1" numerical="False" + multiple="True" use_header_names="True" force_select="False"/> + <conditional name="model"> + <param name="fam" type="select" label="GLM Link function for models" + help="Binary dependant variables will automatically be set to Binomial no matter what this is set to"> + <option value="gaussian" selected="true">Gaussian - continuous dependent (y)</option> + <option value="binomial">Binomial dependent variables</option> + <option value="poisson">Poisson (eg counts)</option> + <option value="cox">Cox models - require special setup for y variables - see below</option> + </param> + <when value="gaussian"> + <param name="yvar_cols" label="Select numeric columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="False" + multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> + <param name="output_full" type="hidden" value='F' /> + <param name="output_pred" type="hidden" value='F' /> + <param name="cox_id" label="Select column containing a unique sample identifier" + help = "Only really needed for output sample specific predicted values downstream." + type="data_column" data_ref="input1" numerical="False" force_select="True" + multiple="False" use_header_names="True" /> + </when> + <when value="binomial"> + <param name="yvar_cols" label="Select numeric columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="False" + multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> + <param name="output_full" type="hidden" value='F' /> + <param name="output_pred" type="select" label="Create a tabular output with predicted values for each subject from the optimal model for (eg) NRI estimates" > + <option value="F" selected="true">No predicted value output file</option> + <option value="T">Create a predicted value output file</option> + </param> + <param name="cox_id" label="Select column containing a unique sample identifier" + help = "Only really needed for output sample specific predicted values downstream." + type="data_column" data_ref="input1" numerical="False" force_select="True" + multiple="False" use_header_names="True" /> + <param name="predict_at" type="hidden" value='' /> + + </when> + <when value="poisson"> + <param name="yvar_cols" label="Select columns containing variables to use as the dependent (y) in elasticnet" type="data_column" data_ref="input1" numerical="True" + multiple="True" use_header_names="True" help = "If multiple, each will be modelled against all the x variables and reported separately." force_select="True"/> + <param name="output_full" type="hidden" value='F' /> + <param name="output_pred" type="hidden" value='F' /> + <param name="predict_at" type="hidden" value='' /> + <param name="cox_id" label="Select column containing a unique sample identifier" + help = "Optional. Only really needed for output sample specific predicted values downstream. Free - enjoy" + type="data_column" data_ref="input1" numerical="True" force_select="False" + multiple="False" use_header_names="True" /> + </when> + <when value="cox"> + <param name="cox_time" label="Select column containing time under observation for Cox regression" + type="data_column" data_ref="input1" numerical="True" force_select="True" + multiple="False" use_header_names="True" help = "This MUST contain a time period - eg continuous years or days to failure or right censoring"/> + <param name="cox_status" label="Select column containing status = 1 for outcome of interest at the end of the time under observation or 0 for right censoring" + type="data_column" data_ref="input1" numerical="True" force_select="True" + multiple="False" use_header_names="True" help = "This MUST contain 1 for subjects who had an event at that time or 0 for a right censored observation"/> + <param name="cox_id" label="Select column containing a unique sample identifier" + help = "Optional. Only really needed for output sample specific predicted values downstream. Free - enjoy" + type="data_column" data_ref="input1" numerical="False" force_select="False" + multiple="False" use_header_names="True" /> + <param name="output_full" type="select" label="Create a tabular output with coefficients for all predictors" > + <option value="F" selected="true">No full model output file</option> + <option value="T">Create a full model output file</option> + </param> + <param name="output_pred" type="select" label="Create a tabular output with predicted values for each subject from the optimal model for (eg) NRI estimates" > + <option value="F" selected="true">No predicted value output file</option> + <option value="T">Create a predicted value output file</option> + </param> + <param name="predict_at" type="text" value='' label="Provide a comma separated list of times to make a prediction for each subject" + optional="True" help="Default (blank) will return predictions at 0%,25%,50%,75%,100% of the observed times which should be informative" /> + + </when> + </conditional> + <param name="optLambda" type="select" label="Value to use when reporting optimal model and coefficients" help="minLambda will have more predictors - 1SDLambda will be more parsimonious"> + <option value="lambda.1se" selected="true">Lambda + 1 SE of min MSE or AUC (fewer coefficients - more false negatives)</option> + <option value="lambda.min">Lambda at min MSE or max AUC (more coefficients - more false positives)</option> + </param> + <param name="logxform_cols" optional="True" label="Select numeric columns to be log transformed before use as predictors or dependent variables" type="data_column" + data_ref="input1" numerical="True" multiple="True" use_header_names="True" help = "The wisdom of doing this depends entirely on your predictors - eg can help diminish long-tailed outlier influence" + force_select="False"/> + <param name="do_standard" type="select" label="Standardise x vars" + help="If all measurements on same scale, may not be needed. Coefficients are always returned on the original scale."> + <option value="False" selected="true">No standardisation of predictors</option>l + <option value="True">Standardise predictors before model</option> + </param> + <param name="mdsplots" type="select" label="Generate MDS plots of samples in measurement space and measurements in sample space" > + <option value="False" selected="true">No MDS plots</option>l + <option value="True">Yes create MDS plots</option> + </param> + <param name="alpha" type="float" value="0.95" size="5" min="0.01" max="1.0" label="Alpha - see glmnet docs. 1 for pure lasso. 0.0 for pure ridge regression" + help="Default 0.95 allows lasso to cope better with expected predictor collinearity. Use (eg) 0.5 for hybrid regularised regression or (eg) 0.025 for ridge regression"/> + <param name="nfold" type="integer" value="10" size="5" label="Number of folds for internal cross validation" + help="Default of 10 is usually ok"/> + </inputs> + <outputs> + <data format="html" name="html_file" label="${title}.html"/> + <data format="tabular" name="model_file" label="${title}_modelres.xls"/> + <data format="tabular" name="output_full_file" label="${title}_full_cox_model.xls"> + <filter>model['output_full'] == 'T'</filter> + </data> + <data format="tabular" name="output_pred_file" label="${title}_predicted_from_model.xls"> + <filter>model['output_pred'] == 'T'</filter> + </data> + </outputs> + <tests> + <test> + <param name='input1' value='cox_test.xls' ftype='tabular' /> + <param name='treatment_name' value='case' /> + <param name='title' value='Cox glmnet test' /> + <param name='nfold' value='10' /> + <param name='logxform_cols' value='' /> + <param name='alpha' value='0.95' /> + <param name='do_standard' value="True" /> + <param name='cox_time' value='1' /> + <param name='cox_status' value='2' /> + <param name='cox_id' value='1' /> + <param name='predict_at' value='' /> + <param name='fam' value='cox' /> + <param name='yvar_cols' value='' /> + <param name='xvar_cols' value='3,4,5' /> + <param name='force_xvar_cols' value='3' /> + <param name='output_full' value='F' /> + <param name='output_pred' value='F' /> + <output name='model_file' file='coxlassotest_modelres.xls'> + <assert_contents> + <has_text text="rhubarb" /> + <has_text text="TRUE" /> + <!-- 	 is XML escape code for tab --> + <!-- has_line line="regulator	partial_likelihood	forced_in	glmnet_model	best_lambda" / --> + <has_line line="regulator	partial_likelihood	forced_in	glmnet_model	best_lambda	lambdaChoice	alpha" /> + <has_n_columns n="7" /> + </assert_contents> + </output> + <output name='html_file' file='coxlassotest.html' compare='diff' lines_diff='16' /> + </test> +</tests> +<help> + +**Before you start** + +Please read the glmnet documentation @ glmnet_ + +This Galaxy wrapper merely exposes that code and the glmnet_ documentation is essential reading +before getting useful results here. + +**What it does** + +From documentation at glmnet_ :: + + Glmnet is a package that fits a generalized linear model via penalized maximum likelihood. + The regularization path is computed for the lasso or elasticnet penalty at a grid of values for the regularization parameter lambda. + The algorithm is extremely fast, and can exploit sparsity in the input matrix x. + It fits linear, logistic and multinomial, poisson, and Cox regression models. + A variety of predictions can be made from the fitted models. + +Internal cross validation is used to optimise the choice of lambda based on CV AUC for logistic (binomial outcome) models, or CV mse for gaussian. + +**Warning about the tyrany of dimensionality** + +Yes, this package will select 'optimal' models even when you (optimistically) supply more predictors than you have cases. +The model returned is unlikely to represent the only informative regularisation path through your data - if you run repeatedly with +exactly the same settings, you will probably see many different models being selected. +This is not a software bug - the real problem is that you just don't have enough information in your data. + +Sufficiently big jobs will take a while (eg each lasso regression with 20k features on 1k samples takes about 2-3 minutes on our aged cluster) + +**Input** + +Assuming you have more measurements than samples, you supply data as a tabular text file where each row is a sample and columns +are variables. You specify which columns are dependent (predictors) and which are observations for each sample. Each of multiple +dependent variable columns will be run and reported independently. Predictors can be forced in to the model. + +**Output** + +For each selected dependent regression variable, a brief report of the model coefficients predicted at the +'optimal' nfold CV value of lambda. + +**Predicted event probabilities for Cox and Logistic models** + +If you want to compare (eg) two competing clinical predictions, there's a companion generic NRI tool +for predicted event probabilities. Estimates dozens of measures of improvement in prediction. Currently only works for identical id subjects +but can probably be extended to independent sample predictions. + +Given a model, we can generate a predicted p (for status 1) in binomial or cox frameworks so models can be evaluated in terms of NRI. +Of course, estimates are likely substantially inflated over 'real world' performance by being estimated from the same sample - but you probably +already knew that since you were smart enough to reach this far down into the on screen help. The author salutes you, intrepid reader! + +It may seem an odd thing to do, but we can predict p for an event for each subject from our original data, given a parsimonious model. Doing +this for two separate models (eg, forcing in an additional known explanatory measurement to the new model) allows comparison of the two models +predicted status for each subject, or the same model in independent populations to see how badly it does + +**Attributions** + +glmnet_ is the R package exposed by this Galaxy tool. + +Galaxy_ (that's what you are using right now!) for gluing everything together + +Otherwise, all code and documentation comprising this tool was written by Ross Lazarus and is +licensed to you under the LGPL_ like other rgenetics artefacts + +.. _LGPL: http://www.gnu.org/copyleft/lesser.html +.. _glmnet: http://web.stanford.edu/~hastie/glmnet/glmnet_alpha.html +.. _Galaxy: http://getgalaxy.org +</help> + <citations> <citation type="bibtex"> @Article{Friedman2010, title = {Regularization Paths for Generalized Linear Models via Coordinate Descent}, @@ -917,6 +908,3 @@ </citation> </citations> </tool> - - -