Package 'ipflasso'

Title: Integrative Lasso with Penalty Factors
Description: The core of the package is cvr2.ipflasso(), an extension of glmnet to be used when the (large) set of available predictors is partitioned into several modalities which potentially differ with respect to their information content in terms of prediction. For example, in biomedical applications patient outcome such as survival time or response to therapy may have to be predicted based on, say, mRNA data, miRNA data, methylation data, CNV data, clinical data, etc. The clinical predictors are on average often much more important for outcome prediction than the mRNA data. The ipflasso method takes this problem into account by using different penalty parameters for predictors from different modalities. The ratio between the different penalty parameters can be chosen from a set of optional candidates by cross-validation or alternatively generated from the input data.
Authors: Anne-Laure Boulesteix, Mathias Fuchs, Gerhard Schulze
Maintainer: Anne-Laure Boulesteix <[email protected]>
License: GPL
Version: 1.1
Built: 2024-11-08 06:19:47 UTC
Source: CRAN

Help Index

Cross-validated integrative lasso with adaptive penalty factors

Description

Runs cvr.ipflasso applying different data based penalty factors to predictors from different blocks.

Usage

cvr.adaptive.ipflasso(X, Y, family, type.measure, standardize = TRUE,
                                  alpha, type.step1, blocks, nfolds, ncv)

Arguments

X

a (nxp) matrix of predictors with observations in rows and predictors in columns.
Y

n-vector giving the value of the response (either continuous, numeric-binary 0/1, or Surv object).
family

should be "gaussian" for continuous Y, "binomial" for binary Y, "cox" for Y of type Surv.
type.measure

the accuracy/error measure computed in cross-validation. If not specified, type.measure is "class" (classification error) if family="binomial", "mse" (mean squared error) if family="gaussian" and partial likelihood if family="cox". If family="binomial", one may specify type.measure="auc" (area under the ROC curve).
standardize

whether the predictors should be standardized or not. Default is TRUE.
alpha

the elastic net mixing parameter for step 1: alpha=1 yields the L1 penalty (Lasso), alpha=0 yields the L2 penalty (Ridge).
type.step1

whether the models of step 1 should be run on the whole data set X (type.step1="comb") or separately for each block (type.step1="sep").
blocks

a list of length M of the format list(block1=...,block2=..., where the dots should be replaced by the indices of the predictors included in this block. The blocks should form a partition of 1:p.
nfolds

the number of folds of the CV procedure.
ncv

the number of repetitions of the CV. Not to be confused with nfolds. For example, if one repeats 50 times 5-fold-CV (i.e. considers 50 random partitions into 5 folds in turn and averages the results), nfolds equals 5 and ncv equals 50.

Details

The penalty factors are the inverse arithmetic means of the absolute model coefficients per block, generated in a first step of the function. The user can choose to determine these coefficients by running a Lasso model (alpha=1) or a Ridge model (alpha=0) either on the whole data set (type.step1="comb") or seperately for each block (type.step1="sep"). If type.step1 is ommited, it will be set to "sep" for Lasso and to "comb" for Ridge. If a Lasso model in step 1 returns any zero coefficient mean, the corresponding block will be excluded from the input date set X and step 2 will be run with the remaining blocks. If all model coefficient means are zero, step 2 will not be performed.

Value

A list with the following arguments:

coeff

the matrix of coefficients with predictors corresponding to rows and lambda values corresponding to columns. The first row contains the intercept of the models (for all families other than "cox").

In the special case of separate step 1 Lasso models and all coefficient means equal to zero, the intercept is the average of the separate model intercepts per block.
ind.bestlambda

the index of the best lambda according to CV.
lambda

the lambda sequence.

In the special case of separate step 1 Lasso models and all coefficient means equal to zero, it is the lambda sequence with the highest lambda value among the lambda sequences of all blocks.
cvm

the CV estimate of the measure specified by type.measure for each candidate lambda value.

In the special case of separate step 1 Lasso models and all coefficient means equal to zero, cmv is the average of the separate model cvms per block.
nzero

the number of non-zero coefficients in the selected model.

In the special case of separate step 1 Lasso models and all coefficient means equal to zero, nzero is the sum of the non-zero coefficients of the separate models per block.
family

see arguments.
means.step1

the arithmetic means of the absolute model coefficients per block, returned by the first step of the function.
exc

the exclusion vector containing the indices of the block(s) to be excluded from X.

Author(s)

Gerhard Schulze ([email protected])

References

Schulze, Gerhard (2017): Clinical Outcome Prediction Based on Multi-Omics Data: Extension of IPF-LASSO. Masterarbeit, Ludwig-Maximilians-Universitaet Muenchen (Department of Statistics: Technical Reports) https://doi.org/10.5282/ubm/epub.59092

Examples

# load ipflasso library
library(ipflasso)

# generate dummy data
X<-matrix(rnorm(50*200),50,200)
Y<-rbinom(50,1,0.5)

cvr.adaptive.ipflasso(X=X,Y=Y,family="binomial",type.measure="class",standardize=FALSE,
                      alpha = 1,blocks=list(block1=1:50,block2=51:200),nfolds=5,ncv=10)

Repeating cv.glmnet

Description

the same as cv.glmnet but with several ncv repetitions of CV: cross-validation repeated ncv times (i.e. for ncv different random partitions)

Usage

cvr.glmnet(X, Y, family, standardize=TRUE,alpha=1, nfolds, ncv, type.measure,...)

Arguments

X

a (nxp) matrix of predictors with observations in rows and predictors in columns
Y

n-vector giving the value of the response (either continuous, numeric-binary 0/1, or Surv object)
family

should be "gaussian" for continuous Y, "binomial" for binary Y, "cox" for Y of type Surv
standardize

whether the predictors should be standardized or not. Default is TRUE.
alpha

the elastic net mixing parameter: alpha=1 yields the L1 penalty (lasso), alpha=0 yields the L2 penalty. Default is alpha=1 (lasso).
nfolds

the number of folds of CV procedure.
ncv

the number of repetitions of CV. Not to be confused with nfolds. For example, if one repeats 50 times 5-fold-CV (i.e. considers 50 random partitions into 5 folds in turn and averages the results), nfolds equals 5 and ncv equals 50.
type.measure

The accuracy/error measure computed in cross-validation. If not specified, type.measure is "class" (classification error) if family="binomial", "mse" (mean squared error) if family="gaussian" and partial likelihood if family="cox". If family="binomial", one may specify type.measure="auc" (area under the ROC curve).
...

Other arguments to be passed to the function cv.glmnet.

Value

A list with the following arguments:

coeff

the matrix of coefficients with predictors corresponding to rows and lambda values corresponding to columns. The first rows contains the intercept of the model (for all families other than "cox").
lambda

the lambda sequence
cvm

the CV estimate of the measure specified by type.measure for each candidate lambda value, averaged over the ncv runs of cv.glmnet

Author(s)

Anne-Laure Boulesteix (https://www.en.ibe.med.uni-muenchen.de/mitarbeiter/professoren/boulesteix/index.html)

References

Boulesteix AL, De Bin R, Jiang X, Fuchs M, 2017. IPF-lasso: integrative L1-penalized regression with penalty factors for prediction based on multi-omics data. Comput Math Methods Med 2017:7691937.

Examples

# load ipflasso library
library(ipflasso)

# generate dummy data
X<-matrix(rnorm(50*200),50,200)
Y<-rbinom(50,1,0.5)

cvr.glmnet(X=X,Y=Y,family="binomial",standardize=FALSE,nfolds=5,ncv=10,type.measure="class")

Cross-validated integrative lasso with fixed penalty factors

Description

Runs cvr.glmnet giving different penalty factors to predictors from different blocks.

Usage

cvr.ipflasso(X, Y, family, type.measure, standardize=TRUE, alpha=1, blocks, pf, nfolds,
  ncv)

Arguments

X

a (nxp) matrix of predictors with observations in rows and predictors in columns
Y

n-vector giving the value of the response (either continuous, numeric-binary 0/1, or Surv object)
family

should be "gaussian" for continuous Y, "binomial" for binary Y, "cox" for Y of type Surv
type.measure

The accuracy/error measure computed in cross-validation. If not specified, type.measure is "class" (classification error) if family="binomial", "mse" (mean squared error) if family="gaussian" and partial likelihood if family="cox". If family="binomial", one may specify type.measure="auc" (area under the ROC curve).
standardize

whether the predictors should be standardized or not. Default is TRUE.
alpha

the elastic net mixing parameter: alpha=1 yields the L1 penalty (lasso), alpha=0 yields the L2 penalty. Default is alpha=1 (lasso).
blocks

a list of length M the format list(block1=...,block2=..., where the dots should be replaced by the indices of the predictors included in this block. The blocks should form a partition of 1:p.
pf

a vector of length equal to the number of blocks M. Each entry contains the penalty factor to be applied to the predictors of the corresponding block. Example: if pf=c(1,2), the penalty applied to the predictors of the 2nd block is twice as large as the penalty applied to the predictors of the first block.
nfolds

the number of folds of CV procedure.
ncv

the number of repetitions of CV. Not to be confused with nfolds. For example, if one repeats 50 times 5-fold-CV (i.e. considers 50 random partitions into 5 folds in turn and averages the results), nfolds equals 5 and ncv equals 50.

Value

A list with the following arguments:

coeff

the matrix of coefficients with predictors corresponding to rows and lambda values corresponding to columns. The first rows contains the intercept of the model (for all families other than "cox").

ind.bestlambda

the index of the best lambda according to CV.
lambda

the lambda sequence.
cvm

the CV estimate of the measure specified by type.measure for each candidate lambda value.
nzero

the number of non-zero coefficients in the selected model.
family

See arguments.

Author(s)

Anne-Laure Boulesteix (https://www.en.ibe.med.uni-muenchen.de/mitarbeiter/professoren/boulesteix/index.html)

References

Boulesteix AL, De Bin R, Jiang X, Fuchs M, 2017. IPF-lasso: integrative L1-penalized regression with penalty factors for prediction based on multi-omics data. Comput Math Methods Med 2017:7691937.

Examples

# load ipflasso library
library(ipflasso)

# generate dummy data
X<-matrix(rnorm(50*200),50,200)
Y<-rbinom(50,1,0.5)

cvr.ipflasso(X=X,Y=Y,family="binomial",standardize=FALSE,
            blocks=list(block1=1:50,block2=51:200), 
            pf=c(1,2),nfolds=5,ncv=10,type.measure="class")

Cross-validated integrative lasso with cross-validated penalty factors

Description

Runs cvr.glmnet giving different penalty factors to predictors from different blocks and chooses the penalty factors by cross-validation from the list pflist of candidates.

Usage

cvr2.ipflasso(X, Y, family, type.measure, standardize=TRUE, 
              alpha=1, blocks, pflist, nfolds, ncv, 
              nzeromax = +Inf, plot=FALSE)

Arguments

X

a (nxp) matrix of predictors with observations in rows and predictors in columns
Y

n-vector giving the value of the response (either continuous, numeric-binary 0/1, or Surv object)
family

should be "gaussian" for continuous Y, "binomial" for binary Y, "cox" for Y of type Surv
type.measure

The accuracy/error measure computed in cross-validation. If not specified, type.measure is "class" (classification error) if family="binomial", "mse" (mean squared error) if family="gaussian" and partial likelihood if family="cox". If family="binomial", one may specify type.measure="auc" (area under the ROC curve).
standardize

whether the predictors should be standardized or not. Default is TRUE.
alpha

the elastic net mixing parameter: alpha=1 yields the L1 penalty (lasso), alpha=0 yields the L2 penalty. Default is alpha=1 (lasso).
blocks

a list of length M the format list(block1=...,block2=..., where the dots should be replaced by the indices of the predictors included in this block. The blocks should form a partition of 1:p.
pflist

a list of candidate penalty factors (see the argument pf of the function cvr.ipflasso) of the format weightslist=list(c(1,1),c(1,2),c(2,1),...).
nfolds

the number of folds of CV procedure.
ncv

the number of repetitions of CV. Not to be confused with nfolds. For example, if one repeats 50 times 5-fold-CV (i.e. considers 50 random partitions into 5 folds in turn and averages the results), nfolds equals 5 and ncv equals 50.
nzeromax

the maximal number of predictors allowed in the final model. Default is +Inf, i.e. the best model is selected based on CV without restriction.
plot

If plot=TRUE, the function outputs plots of CV errors and number of included predictors for each block.

Value

A list with the following arguments:

coeff

the matrix of coefficients obtained with the best combination of penalty factors, with covariates corresponding to rows and lambda values corresponding to columns. The first row contains the intercept of the model.
ind.bestlambda

the index of the best lambda as selected by CV for the best combination of penalty factors.
bestlambda

the best lambda as selected by CV for the best combination of penalty factors.
ind.bestpf

the index of the best penalty factor selected by CV from the list of candidates pflist.
cvm

the CV error for each candidate lambda value, averaged over the ncv runs of cv.glmnet.
a

a list of length length(pflist) containing the outputs of the function cvr.ipflasso for all candidate penalty factors from pflist.
family

See arguments.

Author(s)

Anne-Laure Boulesteix (https://www.en.ibe.med.uni-muenchen.de/mitarbeiter/professoren/boulesteix/index.html)

References

Boulesteix AL, De Bin R, Jiang X, Fuchs M, 2017. IPF-lasso: integrative L1-penalized regression with penalty factors for prediction based on multi-omics data. Comput Math Methods Med 2017:7691937.

Examples

# load ipflasso library
library(ipflasso)

# generate dummy data
X<-matrix(rnorm(50*200),50,200)
Y<-rbinom(50,1,0.5)

cvr2.ipflasso(X=X,Y=Y,family="binomial",type.measure="class",standardize=FALSE,
              blocks=list(block1=1:50,block2=51:200),
              pflist=list(c(1,1),c(1,2),c(2,1)),nfolds=5,ncv=10)

Using an IPF-lasso model for prediction of new observations

Description

Derives predictions for new observations from a model fitted by the functions cvr.ipflasso or cvr2.ipflasso.

Usage

ipflasso.predict(object, Xtest)

Arguments

object

the output of either cvr.ipflasso (if the user chooses the penalty factor himself) or cvr2.ipflasso (if the user cross-validates the penalty factor).
Xtest

a ntest x p matrix containing the values of the predictors for the test data. It should have the same number of columns as the matrix X used to obtain the model result.

Value

A list with the following arguments:

linpredtest

a ntest-vector giving the value of the linear predictor for the test observations
classtest

a ntest-vector with values 0 or 1 giving the predicted class for the test observations (for binary Y).
probabilitiestest

a ntest-vector giving the predicted probability of Y=1 for the test observations (for binary Y).

Author(s)

Anne-Laure Boulesteix (https://www.en.ibe.med.uni-muenchen.de/mitarbeiter/professoren/boulesteix/index.html)

References

Boulesteix AL, De Bin R, Jiang X, Fuchs M, 2017. IPF-lasso: integrative L1-penalized regression with penalty factors for prediction based on multi-omics data. Comput Math Methods Med 2017:7691937.

Examples

# load ipflasso library
library(ipflasso)

# generate dummy data
X<-matrix(rnorm(50*200),50,200)
Xtest<-matrix(rnorm(20*200),20,200)
Y<-rbinom(50,1,0.5)

# fitting the IPF-lasso model
model1<-cvr.ipflasso(X=X,Y=Y,family="binomial",standardize=FALSE,
                    blocks=list(block1=1:50,block2=51:200),
                    pf=c(1,2),nfolds=5,ncv=10,type.measure="class")

# making predictions from Xtest
ipflasso.predict(object=model1,Xtest=Xtest)

Area under the curve (AUC)

Description

computes the area under the ROC curve (AUC) for the marker 'linpred' and the binary status 'Y'.

Usage

my.auc(linpred, Y)

Arguments

linpred

n-vector giving the value of the marker.
Y

n-vector giving the binary status, coded as 0/1.

Value

the area under the curve