| Title: | Multi-Class Discriminant Analysis using Binary Predictors |
|---|---|
| Description: | Implements functions for multi-class discriminant analysis using binary predictors, for corresponding variable selection, and for dichotomizing continuous data. |
| Authors: | Sebastian Gibb and Korbinian Strimmer. |
| Maintainer: | Korbinian Strimmer <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 1.0.4 |
| Built: | 2024-12-07 06:28:20 UTC |
| Source: | CRAN |
The "binda" package implements functions for multi-class discriminant analysis using binary predictors, for corresponding variable selection, and for dichotomizing continuous data.
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io/)
Gibb, S., and K. Strimmer. 2015. Differential protein expression and peak selection in mass spectrometry data by binary discriminant analysis. Bioinformatics 31:3156-3162. <DOI:10.1093/bioinformatics/btv334>
Website: https://strimmerlab.github.io/software/binda/
binda, binda.ranking, dichotomize, chances.
binda trains a diagonal multivariate Bernoulli model.
predict.binda performs corresponding class prediction.
binda(Xtrain, L, lambda.freqs, verbose=TRUE) ## S3 method for class 'binda' predict(object, Xtest, verbose=TRUE, ...)binda(Xtrain, L, lambda.freqs, verbose=TRUE) ## S3 method for class 'binda' predict(object, Xtest, verbose=TRUE, ...)
Xtrain |
A matrix containing the training data set. Note that the rows correspond to observations and the columns to variables. |
L |
A factor with the class labels of the training samples. |
lambda.freqs |
Shrinkage intensity for the frequencies. If not specified it is
estimated from the data. |
verbose |
Report shrinkage intensities (binda) and number of used features (predict.binda). |
object |
An |
Xtest |
A matrix containing the test data set. Note that the rows correspond to observations and the columns to variables. |
... |
Additional arguments for generic predict. |
For detailed description of binary discriminant analysis as implented in binda see Gibb and Strimmer (2015).
predict.binda predicts class probabilities for each test sample and returns
a list with two components:
class |
a factor with the most probable class assignment for each test sample, and |
posterior |
a matrix containing the respective class posterior probabilities. |
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io).
Gibb, S., and K. Strimmer. 2015. Differential protein expression and peak selection in mass spectrometry data by binary discriminant analysis. Bioinformatics 31:3156-3162. <DOI:10.1093/bioinformatics/btv334>
# load "binda" library library("binda") # training data set with labels Xtrain = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xtrain) = paste0("V", 1:ncol(Xtrain)) is.binaryMatrix(Xtrain) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) # learn predictor binda.fit = binda(Xtrain, L) # predict classes using new test data Xtest = matrix(c(1, 1, 0, 1, 1, 1, 1, 0, 0, 0, 1, 1), nrow=2, byrow=TRUE) colnames(Xtest) = paste0("V", 1:ncol(Xtest)) predict(binda.fit, Xtest)# load "binda" library library("binda") # training data set with labels Xtrain = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xtrain) = paste0("V", 1:ncol(Xtrain)) is.binaryMatrix(Xtrain) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) # learn predictor binda.fit = binda(Xtrain, L) # predict classes using new test data Xtest = matrix(c(1, 1, 0, 1, 1, 1, 1, 0, 0, 0, 1, 1), nrow=2, byrow=TRUE) colnames(Xtest) = paste0("V", 1:ncol(Xtest)) predict(binda.fit, Xtest)
binda.ranking determines a ranking of predictors by computing corresponding t-scores
between the group means and the pooled mean.
plot.binda.ranking provides a graphical visualization of the top ranking variables
binda.ranking(Xtrain, L, lambda.freqs, verbose=TRUE) ## S3 method for class 'binda.ranking' plot(x, top=40, arrow.col="blue", zeroaxis.col="red", ylab="Variables", main, ...)binda.ranking(Xtrain, L, lambda.freqs, verbose=TRUE) ## S3 method for class 'binda.ranking' plot(x, top=40, arrow.col="blue", zeroaxis.col="red", ylab="Variables", main, ...)
Xtrain |
A matrix containing the training data set. Note that the rows correspond to observations and the columns to variables. |
L |
A factor with the class labels of the training samples. |
lambda.freqs |
Shrinkage intensity for the class frequencies. If not specified it is
estimated from the data. |
verbose |
Print out some info while computing. |
x |
A "binda.ranking" object – this is produced by the binda.ranking() function. |
top |
The number of top-ranking variables shown in the plot (default: 40). |
arrow.col |
Color of the arrows in the plot (default is |
zeroaxis.col |
Color for the center zero axis (default is |
ylab |
Label written next to feature list (default is |
main |
Main title (if missing, |
... |
Other options passed on to generic plot(). |
The overall ranking of a feature is determined by computing a weighted sum of
the squared t-scores. This is approximately equivalent to the mutual information between the response and each variable. The same criterion is used in dichotomize. For precise details see Gibb and Strimmer (2015).
binda.ranking returns a matrix with the following columns:
idx |
original feature number |
score |
the score determining the overall ranking of a variable |
t |
for each group and feature the t-score of the class mean versus the pooled mean |
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io).
Gibb, S., and K. Strimmer. 2015. Differential protein expression and peak selection in mass spectrometry data by binary discriminant analysis. Bioinformatics 31:3156-3162. <DOI:10.1093/bioinformatics/btv334>
binda, predict.binda, dichotomize.
# load "binda" library library("binda") # training data set with labels Xtrain = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xtrain) = paste0("V", 1:ncol(Xtrain)) is.binaryMatrix(Xtrain) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) # ranking variables br = binda.ranking(Xtrain, L) br # idx score t.Control t.Treatment #V2 2 4.000000 -2.000000 2.000000 #V4 4 4.000000 -2.000000 2.000000 #V5 5 4.000000 2.000000 -2.000000 #V6 6 4.000000 2.000000 -2.000000 #V3 3 1.333333 -1.154701 1.154701 #V1 1 0.000000 0.000000 0.000000 #attr(,"class") #[1] "binda.ranking" #attr(,"cl.count") #[1] 2 # show plot plot(br) # result: variable V1 is irrelevant for distinguishing the two groups# load "binda" library library("binda") # training data set with labels Xtrain = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xtrain) = paste0("V", 1:ncol(Xtrain)) is.binaryMatrix(Xtrain) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) # ranking variables br = binda.ranking(Xtrain, L) br # idx score t.Control t.Treatment #V2 2 4.000000 -2.000000 2.000000 #V4 4 4.000000 -2.000000 2.000000 #V5 5 4.000000 2.000000 -2.000000 #V6 6 4.000000 2.000000 -2.000000 #V3 3 1.333333 -1.154701 1.154701 #V1 1 0.000000 0.000000 0.000000 #attr(,"class") #[1] "binda.ranking" #attr(,"cl.count") #[1] 2 # show plot plot(br) # result: variable V1 is irrelevant for distinguishing the two groups
chances estimates Bernoulli parameters (=chances) from a binary matrix and associated class labels.
chances(X, L, lambda.freqs, verbose=TRUE)chances(X, L, lambda.freqs, verbose=TRUE)
X |
data matrix (columns correspond to variables, rows to samples). |
L |
factor containing the class labels, one for each sample (row). |
lambda.freqs |
shrinkage parameter for class frequencies (if not specified it is estimated). |
verbose |
report shrinkage intensity and other information. |
The class-specific chances are estimated using the empirical means over the 0s and 1s in each class. For estimating the pooled mean the class-specific means are weighted using the
estimated class frequencies. Class frequencies are estimated using freqs.shrink.
chances returns a list with the following components:
samples: the samples in each class,
regularization: the shrinkage intensity used to estimate the class frequencies,
freqs: the estimated class frequencies,
means: the estimated chances (parameters of Bernoulli distribution, expectations of 1s) for each variable conditional on class, as well as the marginal changes (pooled means).
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io).
# load binda library library("binda") # example binary matrix with 6 variables (in columns) and 4 samples (in rows) Xb = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xb) = paste0("V", 1:ncol(Xb)) # Test for binary matrix is.binaryMatrix(Xb) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) chances(Xb, L)# load binda library library("binda") # example binary matrix with 6 variables (in columns) and 4 samples (in rows) Xb = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) colnames(Xb) = paste0("V", 1:ncol(Xb)) # Test for binary matrix is.binaryMatrix(Xb) # TRUE L = factor(c("Treatment", "Treatment", "Control", "Control") ) chances(Xb, L)
dichotomize converts a matrix containing continous measurements into a binary matrix.
optimizeThreshold determines optimal thresholds for dichotomization.
dichotomize(X, thresh) optimizeThreshold(X, L, lambda.freqs, verbose=FALSE)dichotomize(X, thresh) optimizeThreshold(X, L, lambda.freqs, verbose=FALSE)
X |
data matrix (columns correspond to variables, rows to samples). |
thresh |
vector of thresholds, one for each variable (column). |
L |
factor containing the class labels, one for each sample (row). |
lambda.freqs |
shrinkage parameter for class frequencies (if not specified it is estimated). |
verbose |
report shrinkage intensity and other information. |
dichotomize assigns 0 if a matrix entry is lower than given column-specific threshold, otherwise it assigns 1.
optimizeThreshold uses (approximate) mutual information to
determine the optimal thresholds. Specifically, the thresholds are chosen to maximize the
mutual information between response and each variable. The same criterion is also used in
binda.ranking. For detailed description of the dichotomization procedure see Gibb and Strimmer (2015).
Class frequencies are estimated using freqs.shrink.
dichotomize returns a binary matrix.
optimizeThreshold returns a vector containing the variable thresholds.
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io).
Gibb, S., and K. Strimmer. 2015. Differential protein expression and peak selection in mass spectrometry data by binary discriminant analysis. Bioinformatics 31:3156-3162. <DOI:10.1093/bioinformatics/btv334>
binda.ranking, freqs.shrink, mi.plugin, is.binaryMatrix.
# load binda library library("binda") # example data with 6 variables (in columns) and 4 samples (in rows) X = matrix(c(1, 1, 1, 1.75, 0.4, 0, 1, 1, 2, 2, 0.4, 0.09, 1, 0, 1, 1, 0.5, 0.1, 1, 0, 1, 0.5, 0.6, 0.1), nrow=4, byrow=TRUE) colnames(X) = paste0("V", 1:ncol(X)) # class labels L = factor(c("Treatment", "Treatment", "Control", "Control") ) rownames(X) = paste0(L, rep(1:2, times=2)) X # V1 V2 V3 V4 V5 V6 #Treatment1 1 1 1 1.75 0.4 0.00 #Treatment2 1 1 2 2.00 0.4 0.09 #Control1 1 0 1 1.00 0.5 0.10 #Control2 1 0 1 0.50 0.6 0.10 # find optimal thresholds (one for each variable) thr = optimizeThreshold(X, L) thr # V1 V2 V3 V4 V5 V6 #1.00 1.00 2.00 1.75 0.50 0.10 # convert into binary matrix # if value is lower than threshold -> 0 otherwise -> 1 Xb = dichotomize(X, thr) is.binaryMatrix(Xb) # TRUE Xb # V1 V2 V3 V4 V5 V6 #Treatment1 1 1 0 1 0 0 #Treatment2 1 1 1 1 0 0 #Control1 1 0 0 0 1 1 #Control2 1 0 0 0 1 1 #attr(,"thresh") # V1 V2 V3 V4 V5 V6 #1.00 1.00 2.00 1.75 0.50 0.10# load binda library library("binda") # example data with 6 variables (in columns) and 4 samples (in rows) X = matrix(c(1, 1, 1, 1.75, 0.4, 0, 1, 1, 2, 2, 0.4, 0.09, 1, 0, 1, 1, 0.5, 0.1, 1, 0, 1, 0.5, 0.6, 0.1), nrow=4, byrow=TRUE) colnames(X) = paste0("V", 1:ncol(X)) # class labels L = factor(c("Treatment", "Treatment", "Control", "Control") ) rownames(X) = paste0(L, rep(1:2, times=2)) X # V1 V2 V3 V4 V5 V6 #Treatment1 1 1 1 1.75 0.4 0.00 #Treatment2 1 1 2 2.00 0.4 0.09 #Control1 1 0 1 1.00 0.5 0.10 #Control2 1 0 1 0.50 0.6 0.10 # find optimal thresholds (one for each variable) thr = optimizeThreshold(X, L) thr # V1 V2 V3 V4 V5 V6 #1.00 1.00 2.00 1.75 0.50 0.10 # convert into binary matrix # if value is lower than threshold -> 0 otherwise -> 1 Xb = dichotomize(X, thr) is.binaryMatrix(Xb) # TRUE Xb # V1 V2 V3 V4 V5 V6 #Treatment1 1 1 0 1 0 0 #Treatment2 1 1 1 1 0 0 #Control1 1 0 0 0 1 1 #Control2 1 0 0 0 1 1 #attr(,"thresh") # V1 V2 V3 V4 V5 V6 #1.00 1.00 2.00 1.75 0.50 0.10
is.binaryMatrix tests whether m is a matrix and whether it contains only 0s and 1s.
Note that functions like binda.ranking and binda require a binary matrix as input.
is.binaryMatrix(m)is.binaryMatrix(m)
m |
a matrix. |
is.binaryMatrix returns either TRUE or FALSE.
Sebastian Gibb and Korbinian Strimmer (https://strimmerlab.github.io).
# load binda library library("binda") # test matrix m = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) # Test for binary matrix is.binaryMatrix(m) # TRUE# load binda library library("binda") # test matrix m = matrix(c(1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1), nrow=4, byrow=TRUE) # Test for binary matrix is.binaryMatrix(m) # TRUE