Package 'R.ComDim'

Description

Adds or overwrites batch and/or metadata information for one or more blocks of an existing MultiBlock object.

Usage

AddMetadata(MB, block = NULL, batches = NULL, metadata = NULL)

Arguments

Value

The updated MultiBlock.

See Also

MultiBlock, FilterSamplesMultiBlock

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 samples, 50 variables
b2 <- matrix(rnorm(800), 10, 80) # 10 samples, 80 variables
batch_b1 <- rep(1, 10)
meta_b1 <- data.frame(condition = rep(c("A", "B"), 5))
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
# Add batch information to block 'b1':
mb <- AddMetadata(mb, block = "b1", batches = batch_b1)
# Add (or overwrite) metadata for block 'b1':
mb <- AddMetadata(mb, block = "b1", metadata = meta_b1)

Description

Return the block names of a MultiBlock object.

Usage

blockNames(x, ...)

## S4 method for signature 'MultiBlock'
blockNames(x, slot = "Data")

Arguments

Value

A character vector with the block names of the requested slot.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
blockNames(mb) # c("b1", "b2")
blockNames(mb, "Data") # same

Description

Set the block names of a MultiBlock object. Renames the Data and Variables slots, and updates Batch and Metadata names to stay consistent.

Usage

blockNames(x) <- value

## S4 replacement method for signature 'MultiBlock'
blockNames(x) <- value

Arguments

Value

The updated MultiBlock object.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
blockNames(mb) <- c("block1", "block2")
blockNames(mb) # c("block1", "block2")

Description

Extends any matrix decomposition method used for exploratory purposes to the multi-block field. The user provides a function (FUN) to compute the scores from the salience-weighted concatenated blocks; these scores are then used to derive the global scores, local scores, and loadings following the traditional ComDim-PCA framework.

Usage

ComDim_Exploratory(
  MB = MB,
  ndim = NULL,
  FUN = FUN,
  normalise = FALSE,
  threshold = 1e-10,
  loquace = FALSE,
  method = "FUN",
  ...
)

Arguments

Value

A ComDim object. Slots for supervised analysis (R2Y, Q2, DQ2, VIP, VIP.block, PLS.model, cv, Prediction) are empty. The populated slots are:

Method

The label supplied via the method argument.

ndim

Number of Common Dimensions extracted.

Q.scores

Global scores matrix ($n \times ndim$). Column names are CC1, CC2, etc.; row names are sample names. Each column $\mathbf{q}_a$ is a unit-norm consensus score derived from the dominant left direction of FUN applied to the salience-weighted concatenated blocks $\mathbf{W} = [\sqrt{\lambda_1}\mathbf{X}_1 \mid \cdots \mid \sqrt{\lambda_B}\mathbf{X}_B]$.

T.scores

Named list of block-specific local scores matrices ($n \times ndim$ each). For block $b$ and component $a$: local loading $\mathbf{p}_{ba} = \mathbf{X}_b'\mathbf{q}_a$ and local score $\mathbf{t}_{ba} = \mathbf{X}_b\,\mathbf{p}_{ba}(\mathbf{p}_{ba}'\mathbf{p}_{ba})^{-1}$.

P.loadings

Global loadings matrix ($p_{tot} \times ndim$). Column $a$ is $\mathbf{P}_a = \mathbf{X}'\mathbf{q}_a$, where $\mathbf{X}$ is the mean-centred (and optionally normalised) concatenated blocks.

Saliences

Block salience (weight) matrix ($ntable \times ndim$, row names = block names). Entry $(b,a)$ is $\lambda_{ba} = \mathbf{q}_a'\mathbf{X}_b\mathbf{X}_b'\mathbf{q}_a$, the variance of block $b$ captured by global score $a$.

R2X

Proportion of multi-block variance captured by each component (named vector, length $ndim$). Let $\mathbf{s}_a$ be the score vector returned by FUN for component $a$ (before unit-normalisation to obtain $\mathbf{q}_a$), so that $sv_a = \|\mathbf{s}_a\|$; then

$R2X_a = sv_a^4 \big/ \sum_k sv_k^4 = Singular_a^2 \big/ \sum_k Singular_k^2.$

Singular

Squared L2 norms of the FUN score vectors: $Singular_a = sv_a^2 = \|\mathbf{s}_a\|^2$, used to derive R2X.

Mean

List with MeanMB: named list of column-mean vectors per block, used for mean-centring.

Norm

List with NormMB: Frobenius norms used for block normalisation (all ones when normalise = FALSE).

variable.block

Character vector (length $p_{tot}$) indicating the block name of each row in P.loadings.

runtime

Total computation time in seconds.

References

Jouan-Rimbaud Bouveresse D, Rutledge DN (2024). A synthetic review of some recent extensions of ComDim. Journal of Chemometrics, 38(5), e3454. doi:10.1002/cem.3454

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 samples, 50 variables
b2 <- matrix(rnorm(800), 10, 80) # 10 samples, 80 variables
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))

if (requireNamespace("ica", quietly = TRUE)) {
  fun.ICA <- function(W, ndim, ...) {
    # W is the concatenated MB.
    # ndim is the number of components.
    result <- ica::ica(W, ndim)
    # The function must return the source estimates (analogous to PCA scores).
    return(result$S)
  }
  resultsICA <- ComDim_Exploratory(mb,
    ndim = 2,
    FUN = fun.ICA,
    method = "ICA"
  )
}

Description

Finding common dimensions in multi-block datasets using OPLS. Also known as ConsensusOPLS (ComDim-OPLS) for multiblock structures: orthogonal components uncorrelated with Y are extracted from all blocks simultaneously before the predictive components are computed.

Usage

ComDim_OPLS(
  MB = MB,
  y = y,
  ndim = 1,
  nort = 1,
  method = c("OPLS-DA", "OPLS-R"),
  decisionRule = c("fixed", "max")[2],
  normalise = FALSE,
  loquace = FALSE,
  cv.k = 7
)

Arguments

Details

This function is a wrapper around ConsensusOPLS. The core kernel-OPLS extraction is delegated to that package; all ComDim output slots (local scores, loadings, VIP, sensitivity, confusion matrix, etc.) are computed from the returned model objects.

Value

A ComDim object. All slots are populated. Key slots:

Method

"OPLS-DA" or "OPLS-R".

ndim

Number of predictive Common Dimensions.

Q.scores

Predictive global scores matrix ($n \times ndim$).

T.scores

Named list of block-specific predictive local scores.

P.loadings

Global predictive loadings.

Saliences

Predictive block salience matrix ($ntable \times ndim$).

Orthogonal

List with orthogonal component outputs: nort, Q.scores, T.scores, P.loadings.ort, Saliences.ort.

R2X

Named vector (length $ndim + nort$) of X-variance fractions.

R2Y

Named vector (length $ndim + nort$) of Y-variance fractions.

Q2

Cross-validated Q2 per class/response (when cv.k >= 2; otherwise training-set fit).

DQ2

(OPLS-DA only) Cross-validated discriminant Q2 per class.

VIP

Global total VIP (named vector, length $p_{tot}$).

VIP.block

Named list (one data.frame per block) with columns p, o, tot.

PLS.model

KOPLS regression objects: W, B, B0, Y.

cv

Cross-validation results when cv.k >= 2: k, Ypred, Q2, DQ2.

Prediction

Training-set predictions: Y.pred; for OPLS-DA also decisionRule, trueClass, predClass, Sensitivity, Specificity, confusionMatrix.

Mean

List with MeanMB and MeanY.

Norm

List with NormMB, FrobNorms, RVweights.

variable.block

Block membership of each variable.

runtime

Total computation time in seconds.

References

Boccard J, Rutledge DN (2013). A consensus OPLS-DA strategy for multiblock Omics data fusion. Analytica Chimica Acta, 769, 30–39. doi:10.1016/j.aca.2013.01.022

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
y <- rep(c("A", "B"), 5)
results <- ComDim_OPLS(mb, y, ndim = 1, nort = 1, method = "OPLS-DA")

Description

Finding common dimensions in multi-block datasets.

Usage

ComDim_PCA(
  MB = MB,
  ndim = NULL,
  normalise = FALSE,
  threshold = 1e-10,
  loquace = FALSE,
  CompMethod = "Normal",
  Partitions = 1
)

Arguments

Value

A ComDim object. Slots for supervised analysis (R2Y, Q2, DQ2, VIP, VIP.block, PLS.model, cv, Prediction) are empty. The populated slots are:

Method

"PCA".

ndim

Number of Common Dimensions extracted.

Q.scores

Global scores matrix ($n \times ndim$). Column names are CC1, CC2, etc.; row names are sample names. Each column $\mathbf{q}_a$ is a unit-norm consensus score, the dominant left singular vector of the salience-weighted concatenated blocks $\mathbf{W} = [\sqrt{\lambda_1}\mathbf{X}_1 \mid \cdots \mid \sqrt{\lambda_B}\mathbf{X}_B]$.

T.scores

Named list of block-specific local scores matrices ($n \times ndim$ each). For block $b$ and component $a$: local loading $\mathbf{p}_{ba} = \mathbf{X}_b'\mathbf{q}_a$ and local score $\mathbf{t}_{ba} = \mathbf{X}_b\,\mathbf{p}_{ba}(\mathbf{p}_{ba}'\mathbf{p}_{ba})^{-1}$.

P.loadings

Global loadings matrix ($p_{tot} \times ndim$). Column $a$ is $\mathbf{P}_a = \mathbf{X}'\mathbf{q}_a$, where $\mathbf{X}$ is the mean-centred (and optionally normalised) concatenated blocks.

Saliences

Block salience (weight) matrix ($ntable \times ndim$, row names = block names). Entry $(b,a)$ is $\lambda_{ba} = \mathbf{q}_a'\mathbf{X}_b\mathbf{X}_b'\mathbf{q}_a$, the variance of block $b$ captured by global score $a$.

R2X

Proportion of multi-block inertia captured by each component (named vector, length $ndim$). Let $d_a$ be the leading singular value of $\mathbf{W}$ for component $a$ (stored as $Singular_a = d_a^2$); then

$R2X_a = Singular_a^2 \big/ \sum_k Singular_k^2 = d_a^4 \big/ \sum_k d_k^4.$

Singular

Squared leading singular values of $\mathbf{W}$, one per component: $Singular_a = d_a^2$.

Mean

List with MeanMB: named list of column-mean vectors per block, used for mean-centring.

Norm

List with NormMB: Frobenius norms used for block normalisation (all ones when normalise = FALSE).

variable.block

Character vector (length $p_{tot}$) indicating the block name of each row in P.loadings.

runtime

Total computation time in seconds.

References

Jouan-Rimbaud Bouveresse D, Rutledge DN (2024). A synthetic review of some recent extensions of ComDim. Journal of Chemometrics, 38(5), e3454. doi:10.1002/cem.3454

Original MATLAB implementation: https://github.com/DNRutledge/ComDim/

Examples

# Example 1: two data blocks.
b1 <- matrix(rnorm(500), 10, 50) # 10 samples, 50 variables
b2 <- matrix(rnorm(800), 10, 80) # 10 samples, 80 variables
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
results <- ComDim_PCA(mb, 2)

# Example 2: two data blocks, each with different replicate number
b1 <- matrix(rnorm(500), 10, 50)
batch_b1 <- rep(1, 10)
b2 <- matrix(rnorm(2400), 30, 80)
batch_b2 <- c(rep(1, 10), rep(2, 10), rep(3, 10))
mb <- MultiBlock(
  Samples = list(
    b1 = paste0("samples_", 1:10),
    b2 = rep(paste0("samples_", 1:10), 3)
  ),
  Data = list(b1 = b1, b2 = b2),
  Batch = list(b1 = batch_b1, b2 = batch_b2),
  ignore.size = TRUE
)
rw <- SplitRW(mb)
results <- ComDim_PCA(rw, 2)

Description

Finding common dimensions in multi-block datasets using PLS.

Usage

ComDim_PLS(
  MB = MB,
  y = y,
  ndim = NULL,
  method = c("PLS-DA", "PLS-R"),
  decisionRule = c("fixed", "max")[2],
  normalise = FALSE,
  scale.y = FALSE,
  threshold = 1e-10,
  loquace = FALSE,
  CompMethod = "Normal",
  Partitions = 1,
  cv.k = 7
)

Arguments

Value

A ComDim object. All slots are populated. Key slots:

Method

"PLS-DA" or "PLS-R".

ndim

Number of Common Dimensions extracted.

Q.scores

Global consensus PLS scores ($n \times ndim$). Each column $\mathbf{q}_a$ (unit-norm) is the dominant left singular vector from the NIPALS PLS applied to the salience-weighted concatenated blocks.

T.scores

Named list of block-specific local scores ($n \times ndim$ each). Local loading $\mathbf{p}_{ba} = \mathbf{X}_b'\mathbf{q}_a$; local score $\mathbf{t}_{ba} = \mathbf{X}_b\,\mathbf{p}_{ba}(\mathbf{p}_{ba}'\mathbf{p}_{ba})^{-1}$.

P.loadings

Global loadings ($p_{tot} \times ndim$): $\mathbf{P} = \mathbf{X}'\mathbf{Q}$, where $\mathbf{X}$ is the mean-centred (and optionally normalised) concatenated blocks.

Saliences

Block salience matrix ($ntable \times ndim$): $\lambda_{ba} = \mathbf{q}_a'\mathbf{X}_b\mathbf{X}_b'\mathbf{q}_a$.

R2X

Proportion of X variance captured by each component (named vector, length $ndim$). Let $\mathbf{t}_a$ be the NIPALS PLS X-score vector for component $a$ on the salience- weighted blocks; then

$R2X_a = \|\mathbf{t}_a\|^4 \big/ \sum_k \|\mathbf{t}_k\|^4.$

R2Y

Cumulative Y-variance explained (named vector, length $ndim$). $R2Y_a$ is the $R^2$ from an OLS regression of $\mathbf{Y}$ on the first $a$ global scores with an intercept:

$R2Y_a = 1 - RSS_a / TSS_Y,$

where $RSS_a$ is the residual SS when predicting $\mathbf{Y}$ from $[1, \mathbf{q}_1, \ldots, \mathbf{q}_a]$. Note: $R2Y_a$ is cumulative — it reflects the total Y-variance explained by the first $a$ components together, not the marginal contribution of component $a$ alone.

Q2

Predictive Q2 per response column (PLS-R) or per class (PLS-DA), named accordingly:

$Q2 = 1 - PRESS / TSS_Y,$

where $PRESS = \sum_i (\hat{y}_i - y_i)^2$ and $TSS_Y = \sum_i (y_i - \bar{y})^2$. When cv.k >= 2, $\hat{y}_i$ are out-of-sample cross-validated predictions; otherwise training-set predictions are used (i.e. Q2 = R2Y for the full model).

DQ2

(PLS-DA only) Discriminant Q2 per class. Only penalising residuals contribute to the sum:

$DQ2 = 1 - PRESSD / TSS_Y,$

where $PRESSD$ sums $\hat{y}_i^2$ for class-0 samples with $\hat{y}_i > 0$, and $(\hat{y}_i - 1)^2$ for class-1 samples with $\hat{y}_i < 1$. Same cross-validation logic as Q2.

Singular

Squared L2 norm of the NIPALS PLS score vector per component ($\|\mathbf{t}_a\|^2$), used to derive R2X.

VIP

Global VIP scores (named numeric vector, length $p_{tot}$) using the Wold formula:

$VIP_j = \sqrt{p_{tot} \cdot \frac{\sum_a s_a \tilde{w}_{j,a}^2}{\sum_a s_a}},$

where $s_a = \|\mathbf{t}_a\|^2 \|\mathbf{q}_a\|^2$, $\tilde{w}_{j,a} = w_{j,a} / \|\mathbf{w}_a\|$ is the L2-normalised $j$-th element of the $a$-th NIPALS weight vector, and $p_{tot}$ is the total number of variables.

VIP.block

Named list (one data.frame per block) with columns p (per-block predictive VIP computed with block size $p_b$ instead of $p_{tot}$) and tot (= p for PLS; included for consistency with OPLS output). Row names are variable names.

PLS.model

List with: W (NIPALS X weight matrix, $p_{tot} \times ndim$); B (regression coefficients, $p_{tot} \times ncol(Y)$, $\mathbf{B} = \mathbf{W}(\mathbf{P}'\mathbf{W})^{-1}\mathbf{Q}'$, in original Y units); B0 (intercept vector, length $ncol(Y)$, $\mathbf{B}_0 = \bar{\mathbf{y}} - \bar{\mathbf{x}}\mathbf{B}$); Y (original response matrix as supplied). Training-set Y predictions: $\hat{\mathbf{Y}} = \mathbf{X}\mathbf{B} + \mathbf{B}_0$.

cv

Cross-validation results when cv.k >= 2 (empty list otherwise): k (number of folds), fold (sample-to-fold vector), Ypred ($n \times ncol(Y)$ matrix of out-of-sample predictions), Q2 (CV Q2 per class/response), DQ2 (mean CV DQ2 across classes, PLS-DA only), DQ2.perclass (CV DQ2 per class, PLS-DA only).

Prediction

Training-set predictions: Y.pred ($n \times ncol(Y)$); for PLS-DA also decisionRule, trueClass (character vector), predClass (data.frame), Sensitivity and Specificity (named per class), confusionMatrix (named list of 2x2 matrices, one per class).

Mean

List with MeanMB (column means per block), MeanY (column means of Y before any scaling), and ScaleY (column SDs of Y; all ones when scale.y = FALSE).

Norm

List with NormMB: Frobenius norms for block normalisation.

variable.block

Character vector (length $p_{tot}$) mapping each row of P.loadings and each element of VIP to its block.

runtime

Total computation time in seconds.

References

Jouan-Rimbaud Bouveresse D, Rutledge DN (2024). A synthetic review of some recent extensions of ComDim. Journal of Chemometrics, 38(5), e3454. doi:10.1002/cem.3454

Examples

b1 <- matrix(rnorm(500), 10, 50)
batch_b1 <- rep(1, 10)
b2 <- matrix(rnorm(2400), 30, 80)
batch_b2 <- c(rep(1, 10), rep(2, 10), rep(3, 10))
mb <- MultiBlock(
  Samples = list(
    b1 = paste0("samples_", 1:10),
    b2 = rep(paste0("samples_", 1:10), 3)
  ),
  Data = list(b1 = b1, b2 = b2),
  Batch = list(b1 = batch_b1, b2 = batch_b2),
  ignore.size = TRUE
)
rw <- SplitRW(mb)
y <- scale(1:10, center = TRUE)
results.plsr <- ComDim_PLS(rw, y, 2, method = "PLS-R")
groups <- c(rep("A", 5), rep("B", 5))
results.plsda <- ComDim_PLS(rw, y = groups, 2, method = "PLS-DA")

Description

Extends any PLS-like method used for regression or discriminant purposes to the multi-block field. The user provides a function (FUN) that computes one predictive component from the salience-weighted concatenated blocks; global scores, local scores, and loadings are then derived following the traditional ComDim-PLS framework. Optionally, orthogonal components returned by FUN (e.g. from an O-PLS wrapper) are captured. VIP scores and k-fold cross-validation are also supported.

Usage

ComDim_y(
  MB = MB,
  y = y,
  ndim = NULL,
  FUN = FUN,
  nort = 0L,
  type = c("regression", "discriminant")[1],
  decisionRule = c("fixed", "max")[2],
  normalise = FALSE,
  scale.y = FALSE,
  threshold = 1e-10,
  loquace = FALSE,
  method = "FUN",
  cv.k = 7,
  ...
)

Arguments

Value

A ComDim object with the following slots:

Method

The label supplied via the method argument.

ndim

Number of predictive Common Dimensions extracted.

Q.scores

Global consensus scores matrix ($n \times ndim$). Each column $\mathbf{q}_a$ (unit-norm) is derived from the dominant left direction of FUN applied to the salience-weighted concatenated blocks.

T.scores

Named list of block-specific local scores ($n \times ndim$ each). Local loading $\mathbf{p}_{ba} = \tilde{\mathbf{X}}_b'\mathbf{q}_a$ (computed on the ort-deflated block when nort > 0); local score $\mathbf{t}_{ba} = \tilde{\mathbf{X}}_b\,\mathbf{p}_{ba}(\mathbf{p}_{ba}'\mathbf{p}_{ba})^{-1}$.

P.loadings

Global loadings ($p_{tot} \times ndim$): $\mathbf{P} = \tilde{\mathbf{X}}'\mathbf{Q}$, where $\tilde{\mathbf{X}}$ is the (optionally ort-deflated) mean-centred concatenated blocks.

Saliences

Block salience matrix ($ntable \times ndim$): $\lambda_{ba} = \mathbf{q}_a'\tilde{\mathbf{X}}_b\tilde{\mathbf{X}}_b'\mathbf{q}_a$.

R2X

Proportion of X variance captured by each predictive component (named vector, length $ndim$). Let $\mathbf{t}_a$ be the X-score vector returned by FUN for component $a$:

$R2X_a = \|\mathbf{t}_a\|^4 \big/ \sum_k \|\mathbf{t}_k\|^4.$

When nort > 0, the denominator also includes the orthogonal $\|\mathbf{t}_{ort,k}\|^4$ terms, and the orthogonal R2X fractions are stored separately in Orthogonal$R2X.

R2Y

Cumulative Y-variance explained (named vector, length $ndim$):

$R2Y_a = 1 - RSS_a / TSS_Y,$

where $RSS_a$ is the residual SS from an OLS regression of $\mathbf{Y}$ on $[1, \mathbf{q}_1, \ldots, \mathbf{q}_a]$. Note: $R2Y_a$ is cumulative – the total Y-variance explained by the first $a$ components together, not the marginal contribution of component $a$ alone.

Q2

Predictive Q2 per response column (regression) or per class (discriminant), named accordingly:

$Q2 = 1 - PRESS / TSS_Y,$

where $PRESS = \sum_i (\hat{y}_i - y_i)^2$. When cv.k >= 2 and nort = 0: cross-validated (out-of- sample) predictions are used; otherwise training-set predictions. CV is automatically skipped when nort > 0.

DQ2

(Discriminant mode only) Discriminant Q2 per class, using only penalising residuals:

$DQ2 = 1 - PRESSD / TSS_Y,$

where $PRESSD$ sums $\hat{y}_i^2$ for class-0 samples with $\hat{y}_i > 0$, and $(\hat{y}_i - 1)^2$ for class-1 samples with $\hat{y}_i < 1$. Same cross-validation logic as Q2.

Singular

Squared L2 norm of the FUN X-score vector per component ($\|\mathbf{t}_a\|^2$), used to derive R2X.

VIP

Global total VIP (named vector, length $p_{tot}$): concatenation of VIP.block[[b]]$tot across blocks. When nort = 0, uses the Wold formula; when nort = 1, tot combines predictive and orthogonal VIPs (see VIP.block).

VIP.block

Named list (one data.frame per block). When nort = 0: columns p and tot (= p), using the Wold formula:

$VIPp_j = \sqrt{p_b \cdot \frac{\sum_a s_a \tilde{w}_{j,a}^2}{\sum_a s_a}},$

where $s_a = \|\mathbf{t}_a\|^2\|\mathbf{q}_a\|^2$ and $\tilde{w}_{j,a} = w_{j,a}/\|\mathbf{w}_a\|$ is the L2-normalised $j$-th element of the $a$-th weight vector. When nort = 1: columns p (Wold, same as above), o (orthogonal VIP, loadings-based: $VIPo_j = \sqrt{p_b \cdot \sum_a s_{oa}\tilde{P}_{o,j,a}^2 / \sum_a s_{oa}}$, where $s_{oa} = \|\mathbf{q}_{ort}[,a]\|^2$ and $\tilde{\mathbf{P}}_o$ is the column-L2-normalised block-slice of the ort loadings), and tot ($VIPtot_j = \sqrt{(VIPp_j^2 + VIPo_j^2)/2}$). Row names are variable names.

PLS.model

List with: W (X weight matrix collected from FUN, $p_{tot} \times ndim$); B (regression coefficients, $\mathbf{B} = \mathbf{W}(\mathbf{P}'\mathbf{W})^{-1}\mathbf{Q}'$, in original Y units); B0 (intercept, $\mathbf{B}_0 = \bar{\mathbf{y}} - \overline{\tilde{\mathbf{x}}}\mathbf{B}$); Y (original response matrix as supplied). Training-set predictions: $\hat{\mathbf{Y}} = \tilde{\mathbf{X}}\mathbf{B} + \mathbf{B}_0$.

cv

Cross-validation results when cv.k >= 2 and nort = 0 (empty list otherwise): k, fold (sample-to-fold vector), Ypred ($n \times ncol(Y)$ out-of-sample predictions), Q2 (CV Q2 per class/response), DQ2 (mean CV DQ2, discriminant only), DQ2.perclass (CV DQ2 per class, discriminant only).

Orthogonal

When nort > 0: list with nort, Q.scores (global ort scores, $n \times nort$, unit-norm), T.scores (block ort local scores, $n \times nort$ each), P.loadings.ort (ort loadings, $p_{tot} \times nort$), Saliences.ort ($ntable \times nort$), and R2X (orthogonal X-variance fractions, $R2X_{ort,a} = \|\mathbf{t}_{ort,a}\|^4 / total$). Empty list when nort = 0.

Prediction

Training-set predictions: Y.pred ($n \times ncol(Y)$); for discriminant analysis also decisionRule, trueClass, predClass (data.frame), Sensitivity and Specificity (per class), confusionMatrix (named list of 2x2 matrices).

Mean

List with MeanMB (column means per block), MeanY (column means of Y), and ScaleY (column SDs of Y; all ones when scale.y = FALSE).

Norm

List with NormMB: Frobenius norms for block normalisation.

variable.block

Character vector (length $p_{tot}$) mapping each row of P.loadings and each element of VIP to its block.

runtime

Total computation time in seconds.

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 samples, 50 variables
b2 <- matrix(rnorm(800), 10, 80) # 10 samples, 80 variables
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))

## Example 1: ComDim-PLS (regression) ---------------------------------------
# Single-step NIPALS PLS wrapper (one predictive component per call).
# Note: 'tx' is used instead of 't' to avoid shadowing base::t().
fun.PLS <- function(W, y, ndim, ...) {
  output <- list()
  w <- t(W) %*% y / as.numeric(t(y) %*% y) # X weight (u = y, 1 step)
  w <- w / sqrt(sum(w^2)) # L2 normalise
  tx <- W %*% w # X score
  p <- t(W) %*% tx / as.numeric(t(tx) %*% tx) # X loading
  q <- t(y) %*% tx / as.numeric(t(tx) %*% tx) # Y loading
  u <- y %*% q / as.numeric(t(q) %*% q) # Y score
  output$scores <- as.vector(tx)
  output$P <- as.vector(p)
  output$W <- as.vector(w)
  output$Q <- as.vector(q)
  output$U <- as.vector(u)
  return(output)
}

y <- c(1, 1, 1, 1, 1, 5, 5, 5, 10, 10)
resultsPLS <- ComDim_y(mb,
  y = y, ndim = 2,
  type = "regression",
  FUN = fun.PLS,
  method = "PLS",
  cv.k = 0
)

## Example 2: ComDim-OPLS-DA (discriminant, nort = 1) ----------------------
# Thin wrapper around OPLS_NIPALS_DNR(), the package's NIPALS OPLS engine.
# All inputs (W, y, and any extra args such as 'threshold') are forwarded
# directly via '...'. Use this pattern when nort > 0; for nort = 0 the
# simpler PLS wrapper in Example 1 is sufficient (no orthoscores needed).
fun.OPLS <- function(W, y, ndim, ...) {
  res <- OPLS_NIPALS_DNR(W = W, y = y, ...)
  list(
    scores      = as.vector(res$t_pred),
    P           = as.vector(res$p),
    W           = as.vector(res$w_pred),
    Q           = as.vector(res$q),
    U           = as.vector(res$u),
    orthoscores = matrix(res$t_ort, ncol = 1)
  )
}

groups <- c(rep("A", 5), rep("B", 5))
resultsOPLS <- ComDim_y(mb,
  y = groups, ndim = 1,
  nort = 1,
  type = "discriminant",
  FUN = fun.OPLS,
  method = "OPLS-DA",
  cv.k = 0
)

## Example 3 (not run): ComDim-OPLS-DA via ropls ---------------------------
# Wrapping ropls::opls is also possible. Key points:
#   - Use orthoI = 1 (fixed) instead of NA so the output is predictable.
#   - Always return output$orthoscores; ComDim_y ignores it in phases
#     where ort has already been removed.
#   - Expand the single ropls Q loading to match the ncol(y_dummy) width.

if (requireNamespace("ropls", quietly = TRUE)) {
fun.OPLSDA.ropls <- function(W, y, ndim, ...) {
  output <- list()
  # Convert dummy matrix to ropls-compatible -1/+1 vector
  Y <- c(-1, 1)[apply(y, 1, function(x) match(1, x))]
  result <- tryCatch(
    ropls::opls(
      x = W, y = Y, predI = 1, orthoI = 1,
      fig.pdfC = "none", info.txtC = "none"
    ),
    error = function(e) {
      ropls::opls(
        x = W, y = Y, predI = 1, orthoI = 0,
        fig.pdfC = "none", info.txtC = "none"
      )
    }
  )
  output$scores <- result@scoreMN[, 1]
  output$P <- result@loadingMN[, 1]
  output$W <- result@weightMN[, 1]
  output$U <- result@uMN[, 1]
  # Expand the single ropls Q loading to match the 2-column dummy matrix:
  # loadings for class1 and class2 are antisymmetric in binary PLS-DA.
  output$Q <- c(-result@cMN[, 1], result@cMN[, 1])
  output$y <- result@suppLs$yModelMN # internal y (for scaling detection)
  # Orthogonal scores (used during the ort pre-loop when nort > 0)
  if (!is.null(result@orthoScoreMN) && ncol(result@orthoScoreMN) > 0) {
    output$orthoscores <- result@orthoScoreMN   # n x k matrix; col jj used for jj-th ort
  } else {
    output$orthoscores <- matrix(0, nrow = nrow(W), ncol = 1)
  }
  return(output)
}

b1_r <- matrix(rnorm(8 * 30), 8, 30)
b2_r <- matrix(rnorm(8 * 20), 8, 20)
mb_r <- MultiBlock(Data = list(b1 = b1_r, b2 = b2_r))
resultsOPLSDA <- ComDim_y(mb_r,
  y = c(rep("NI", 4), rep("OFF", 4)),
  ndim = 1,
  nort = 1,
  type = "discriminant",
  FUN = fun.OPLSDA.ropls,
  method = "OPLS-DA(ropls)",
  cv.k = 0
)
}

Description

Object of the type MultiBlock, to use as input for ComDim analyses.

The output of a ComDim analysis.

Slots

Samples

vector with the sample names. If this data is not available, the slot will be filled with integers.

Data

A list with the data-blocks.

Variables

A character vector with the variable names. If this data is not available, the slot will be filled with integers.

Batch

A list with the vectors with the batch information for each data-block. Optional.

Metadata

A list with the samples metadata.

Method

The algorithm used in the core of the ComDim analysis (ex. PCA, PLS,...)

ndim

The number of components.

Q.scores

The Global scores.

T.scores

The Local scores.

P.loadings

The Loadings

Saliences

The Saliences

R2X

The explained variance of the MultiBlock.

R2Y

For regression or discriminant models, the explained variance of the Y-block.

Q2

For regression or discriminant models, the predicted variance of the Y-block.

DQ2

For discriminant models, the predicted discriminant variance of the Y-block.

Singular

The singular values.

Mean

The mean values for each variable in the MultiBlock.

Norm

The norm values for each variable in the MultiBlock.

PLS.model

For ComDim analyses using PLS as the core algorithm, contains the W, B, B0 and Y matrices.

cv

For ComDim_KOPLS, it contains the index of the samples used during the cross-validation.

Prediction

A list with the predicted Y, the decision rule used for sample classification, the sensitivity, the specificity, and a confusion matrix.

Metadata

A list with the per-sample metadata for each block.

variable.block

A vector with the same length as the P.loadings, indicating the block each variable belongs to.

runtime

The used running time.

Description

Splits data into several blocks, allowing variables to appear in more than one block simultaneously. Each variable is duplicated into every block to which it is assigned according to the metadata mapping table.

Usage

ExpandMultiBlock(data = NULL, metadata = NULL, minblock = 0, loquace = TRUE)

Arguments

Details

For each row in metadata that matches a variable in data, the variable's values are copied into the corresponding block column. Column names in the resulting expanded matrix are formed as <block>.<variable>. Variables with all-NA values after expansion are removed. If no matches exist between data column names and metadata, NULL is returned with a warning. If minblock filtering removes all blocks, NULL is returned.

Value

A MultiBlock object whose blocks are defined by the first column of metadata, or NULL if no valid blocks could be constructed.

See Also

MultiBlock, ProcessMultiBlock

Examples

data(mouse_ds)
lipidsMB <- ExpandMultiBlock(data = lipids, metadata = metadata_lipids,
  minblock = 0, loquace = FALSE)

Description

Extracellular metabolites in growth medium

Usage

data(mouse_ds)

Format

An object of class matrix (inherits from array) with 12 rows and 298 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

doi:10.15252/msb.202010141

Examples

data(mouse_ds)
# MB <- MultiBlock(Data = list(RNAseq3 = RNAseq3,
#   lipids = lipids, intra = intra, extra = extra))

Description

Retain a subset of samples in a MultiBlock object.

Usage

FilterSamplesMultiBlock(MB, samples = sampleNames(MB))

Arguments

Value

The MultiBlock object restricted to the requested samples. The Batch and Metadata slots are also subsetted and reordered to match.

See Also

MultiBlock, AddMetadata, ProcessMultiBlock

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
rownames(b1) <- rownames(b2) <- paste0("sample_", 1:10)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
mb <- FilterSamplesMultiBlock(mb, samples = paste0("sample_", 1:5))

Description

GCMS data on cell differentiation

Usage

data(dataset3)

Format

An object of class matrix (inherits from array) with 36 rows and 15 columns.

Author(s)

Cabrero et al. (2019) Scientific data 6:256 (doi:10.1038/s41597-019-0202-7)

Source

Metabolights

Examples

data(dataset3)
mb_d3 <- MultiBlock(Data = list(gcms = gcms))

Description

Intracellular metabolites in growth medium

Usage

data(mouse_ds)

Format

An object of class matrix (inherits from array) with 12 rows and 230 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

doi:10.15252/msb.202010141

Examples

data(mouse_ds)
# MB <- MultiBlock(Data = list(RNAseq3 = RNAseq3,
#   lipids = lipids, intra = intra, extra = extra))

Description

Metadata for the metabolites in growth medium

Usage

data(mouse_ds)

Format

An object of class data.frame with 687 rows and 2 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

doi:10.15252/msb.202010141

Examples

data(mouse_ds)
#' intraMB <- ExpandMB(data = intra, metadata = KEGG_table_metabolites,
#'   minblock = 10, loquace = FALSE)

Description

LCMS data on cell differentiation

Usage

data(dataset3)

Format

An object of class matrix (inherits from array) with 36 rows and 44 columns.

Author(s)

Cabrero et al. (2019) Scientific data 6:256 (doi:10.1038/s41597-019-0202-7)

Source

Metabolights

Examples

data(dataset3)
#' mb_d3 <- MultiBlock(Data = list(lcms = lcms))

Description

Lipid profiles of cell extracts

Usage

data(mouse_ds)

Format

An object of class matrix (inherits from array) with 12 rows and 437 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

doi:10.15252/msb.202010141

Examples

data(mouse_ds)
# MB <- MultiBlock(Data = list(RNAseq3 = RNAseq3,
#   lipids = lipids, intra = intra, extra = extra))

Description

Creates a long (tidy) data frame with the local P-loadings from a ComDim model, suitable for use with ggplot2.

Usage

MakeComDimLoadingsTable(model, blocks = NULL, dim = NULL, dim.ort = NULL)

Arguments

Value

A long data frame with one row per variable–component combination, containing the following columns:

variable.id

Variable name (factor).

variable.id.number

Position of the variable across all blocks (factor).

block.id

Integer index of the block.

block.name

Name of the block (factor).

dim

Component number.

value

Loading value (from P.loadings or Orthogonal$P.loadings.ort).

See Also

MakeComDimScoresTable, ComDim_PCA

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 rows and 50 columns
b2 <- matrix(rnorm(800), 10, 80) # 10 rows and 80 columns
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
model <- ComDim_PCA(mb, ndim = 2)
tbl <- MakeComDimLoadingsTable(model)

Description

Creates a long (tidy) data frame with the global and/or local scores from a ComDim model, suitable for use with ggplot2.

Usage

MakeComDimScoresTable(
  model,
  blocks = NULL,
  dim = NULL,
  dim.ort = NULL,
  include = c("Q.scores", "T.scores", "Q.scores.ort", "T.scores.ort")[1:2]
)

Arguments

Value

A long data frame with one row per sample–component–score-type combination, containing the following columns:

sample.id

Sample name (factor).

sample.id.number

Integer position of the sample (factor).

block.id

Block index, or "Global" / "Global.ort" for Q scores.

block.name

Block name, or "Global" / "Global.ort" for Q scores (factor).

dim

Component number.

scores.type

One of "Global", "Global.ort", "Local", or "Local.ort" (factor).

scores.type.dim

Concatenation of scores type and component number, e.g. "Q.scores1" (factor).

value

Score value.

See Also

MakeComDimLoadingsTable, ComDim_PCA

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 rows and 50 columns
b2 <- matrix(rnorm(800), 10, 80) # 10 rows and 80 columns
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
model <- ComDim_PCA(mb, ndim = 2)
tbl <- MakeComDimScoresTable(model)

Description

Metadata for the lipid profiles of cell extracts

Usage

data(mouse_ds)

Format

An object of class data.frame with 437 rows and 2 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

doi:10.15252/msb.202010141

Examples

data(mouse_ds)
#' lipidsMB <- ExpandMB(data = lipids, metadata = metadata_lipids,
#'   minblock = 0, loquace = FALSE)

Description

Metadata for the RNAseq data of cell extracts

Usage

data(mouse_ds)

Format

An object of class data.frame with 16890 rows and 2 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

SEQOUT

Examples

data(mouse_ds)
#' RNAseqMB <- ExpandMB(data = RNAseq3, metadata = metadata_RNAseq3,
#'   minblock = 500, loquace = FALSE)

Description

miRNAseq data on cell differentiation

Usage

data(dataset3)

Format

An object of class matrix (inherits from array) with 36 rows and 469 columns.

Author(s)

Cabrero et al. (2019) Scientific data 6:256 (doi:10.1038/s41597-019-0202-7)

Source

NCBI

Examples

data(dataset3)
#' mb_d3 <- MultiBlock(Data = list(mirnaseq = mirnaseq))

Description

Converts a MultiAssayExperiment into a MultiBlock. Samples are first intersected across all experiments so that only common samples are retained.

Usage

MultiAssayExperiment2MultiBlock(se, colData_samplenames = NULL, Batch = NULL)

Arguments

Details

Columns (samples) are intersected across all experiments via intersectColumns before conversion. Each experiment in the MultiAssayExperiment becomes one block in the MultiBlock (rows = samples, columns = features). Sample names are taken from the colData row names of se. Metadata from colData is stored only for the first block; subsequent blocks share the same sample order but carry no additional metadata. If Batch is specified, the corresponding column is extracted from colData, removed from the metadata, and stored as the Batch slot of the MultiBlock.

Value

A MultiBlock object with one block per experiment in se.

See Also

MultiBlock, MultiBlock2MultiAssayExperiment

Examples

if (requireNamespace("MultiAssayExperiment", quietly = TRUE)) {
library(MultiAssayExperiment)
mae <- MultiAssayExperiment(
  experiments = ExperimentList(
    block1 = matrix(rnorm(50), nrow = 5, dimnames = list(paste0("s", 1:5), paste0("v", 1:10))),
    block2 = matrix(rnorm(30), nrow = 5, dimnames = list(paste0("s", 1:5), paste0("w", 1:6)))
  )
)
mb <- MultiAssayExperiment2MultiBlock(mae)
}

Description

Creates a MultiBlock object from a named list of data blocks.

Usage

MultiBlock(
  Samples = NULL,
  Data,
  Variables = NULL,
  Batch = NULL,
  Metadata = NULL,
  ignore.names = FALSE,
  ignore.size = FALSE
)

Arguments

Value

A MultiBlock object.

References

Puig-Castellví F, Jouan-Rimbaud Bouveresse D, Mazéas L, Chapleur O, Rutledge DN (2021). Rearrangement of incomplete multi-omics datasets combined with ComDim for evaluating replicate cross-platform variability and batch influence. Chemometrics and Intelligent Laboratory Systems, 218, 104422. doi:10.1016/j.chemolab.2021.104422

Examples

b1 <- matrix(rnorm(500), 10, 50) # 10 samples, 50 variables
b2 <- matrix(rnorm(800), 10, 80) # 10 samples, 80 variables
# Minimal call: Samples and Variables are filled in automatically.
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))

# With explicit sample names (enables cross-block alignment):
rownames(b1) <- paste0("s", 1:10)
rownames(b2) <- paste0("s", 1:10)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))

# Blocks with different row counts (replicate design):
b3 <- matrix(rnorm(800), 30, 80)
batch_b3 <- c(rep(1, 10), rep(2, 10), rep(3, 10))
mb3 <- MultiBlock(
  Data = list(b3 = b3), Batch = list(b3 = batch_b3),
  ignore.names = TRUE, ignore.size = TRUE
)

Description

Combines the blocks of a MultiBlock into a single matrix by column-binding the selected blocks. Useful for computing summary statistics across the whole MultiBlock (e.g. maximum value).

Usage

MultiBlock2Matrix(MB = MB, blocks = NULL, vars = NULL)

Arguments

Details

Blocks are column-bound in the order given by blocks. Row names of the output matrix are the sample names of MB. Column names are the variable names; if the same variable name appears in more than one block, duplicates are disambiguated by appending .<block_name> to all occurrences of the repeated name, and a warning is issued.

Value

A numeric matrix with rows corresponding to samples and columns corresponding to the variables of the selected blocks concatenated in order.

See Also

MultiBlock, ProcessMultiBlock

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
# Combine all blocks into a matrix and compute a summary statistic:
mat <- MultiBlock2Matrix(mb)
max(mat)
# Combine only the first block, keeping variables 1-10:
mat <- MultiBlock2Matrix(mb, blocks = 1, vars = list(1:10))

Description

Converts a MultiBlock into a MultiAssayExperiment. Each block becomes one experiment; sample names, batch information, and metadata are carried over into the colData of the result.

Usage

MultiBlock2MultiAssayExperiment(MB, MSEmetadata = NULL)

Arguments

Details

Each block in MB is transposed (features x samples) and stored as a named matrix in the ExperimentList. Row names are set to the variable names and column names to the sample names of the MultiBlock. The colData is constructed from MB@Samples; any Metadata and Batch information present in the MultiBlock is appended as additional columns. A sampleMap is generated mapping every sample to every experiment using the same primary and column names.

Value

A MultiAssayExperiment object with one experiment per block in MB.

See Also

MultiBlock, MultiAssayExperiment2MultiBlock

Examples

if (requireNamespace("MultiAssayExperiment", quietly = TRUE)) {
library(MultiAssayExperiment)
b1 <- matrix(rnorm(50), 5, 10, dimnames = list(paste0("s", 1:5), paste0("v", 1:10)))
b2 <- matrix(rnorm(30), 5,  6, dimnames = list(paste0("s", 1:5), paste0("w", 1:6)))
mb <- MultiBlock(Data = list(block1 = b1, block2 = b2))
mae <- MultiBlock2MultiAssayExperiment(mb)
mae <- MultiBlock2MultiAssayExperiment(mb, MSEmetadata = list(study = "example"))
}

Description

Remove NA and Infinite values from a MultiBlock object in a single pass. Variables whose combined count of NA and Infinite values meets or exceeds minfrac * nrow are discarded first. Remaining Infinite values are then replaced according to inf.method; remaining NA values are imputed according to na.method.

Usage

NAInfRemoveMultiBlock(
  MB,
  blocks = NULL,
  minfrac = 0.5,
  na.method = c("none", "zero", "median", "discard", "fixed.value", "fixed.value.all",
    "fixed.noise", "random.noise", "QRILC")[8],
  inf.method = c("none", "fixed.noise", "random.noise")[3],
  constant = 0,
  factor.NA = 0.5,
  sd.noise = 0.3,
  tune.sigma = 1,
  showWarning = TRUE
)

Arguments

Value

The processed MultiBlock object.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
b2[c(2, 3, 5), c(1, 2, 3)] <- NA
b2[c(1, 4), c(4, 5)] <- Inf
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
mb <- NAInfRemoveMultiBlock(mb, na.method = "zero", inf.method = "fixed.noise")

Description

Return the number of columns (variables) in each block of a MultiBlock.

Usage

## S4 method for signature 'MultiBlock'
ncol(x)

Arguments

Value

A named integer vector with the number of columns per block.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
ncol(mb) # c(b1 = 50, b2 = 80)

Description

Normalize all blocks from a MultiBlock object. The ranknorm transform is based on that from the RNOmni package (doi:10.1111/biom.13214).

Usage

NormalizeMultiBlock(
  MB,
  blocks = NULL,
  method = c("none", "auto", "mean", "pareto", "norm", "geometric", "ranknorm")[5],
  infinite.as.NA = FALSE,
  constant = 0,
  offset = 3/8,
  showWarning = TRUE
)

Arguments

Value

The normalized MultiBlock object.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
b2[c(2, 3, 5), c(1, 2, 3)] <- NA
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
mb <- NormalizeMultiBlock(mb, method = "auto")

Description

Return the number of rows (samples) in each block of a MultiBlock. For a valid MultiBlock all blocks share the same number of rows, so the result is a named integer vector with one (identical) value per block.

Usage

## S4 method for signature 'MultiBlock'
nrow(x)

Arguments

Value

A named integer vector with the number of rows per block.

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
nrow(mb) # c(b1 = 10, b2 = 10)

Description

One NIPALS OPLS step on a (lambda-weighted) concatenated block matrix. Computes one predictive component and one orthogonal component in a single pass. This function is the recommended building block for constructing an OPLS-based FUN argument for ComDim_y() when nort > 0. For nort = 0 (plain PLS) a simpler PLS wrapper is sufficient.

Usage

OPLS_NIPALS_DNR(W = W, y = y, threshold = 1e-10)

Arguments

Value

A named list:

t_pred

Predictive X-score (length n).

w_pred

Predictive X-weight (length p), L2-normalised.

p

X-loading (length p).

q

Y-loading (length q = ncol(y)).

u

Y-score (length n).

t_ort

Orthogonal X-score (length n).

w_ort

Orthogonal X-weight (length p), L2-normalised.

p_ort

Orthogonal X-loading (length p).

See Also

ComDim_y for the multi-block OPLS wrapper that uses this function.

Description

Projects a new MultiBlock dataset into an existing ComDim model. Works with models produced by ComDim_PCA, ComDim_PLS, ComDim_OPLS, ComDim_y, and ComDim_Exploratory. The projection type (PCA-like, PLS-like, OPLS-like) is determined from the model structure rather than the method string, so custom method labels from ComDim_y are handled automatically.

Usage

PredictMultiBlock(MB = MB, y, model = model, normalise = FALSE, loquace = TRUE)

Arguments

Value

The model ComDim object with updated slots:

• Q.scores — projected global scores (new samples x ndim).

• T.scores — projected local scores (per block).

• Orthogonal$Q.scores — projected global ort scores (if model has orthogonal components).

• Orthogonal$T.scores — projected local ort scores.

• Prediction$Y.pred — predicted Y (supervised models).

• Q2, DQ2, classification slots — when y is supplied for a supervised model.

Description

Apply a custom function to transform a MultiBlock and/or select variables or blocks. When multiple operations are supplied, the order of execution is: vars subsetting, then FUN, then FUN.SelectVars, then FUN.SelectBlocks.

Usage

ProcessMultiBlock(
  MB = MB,
  blocks = NULL,
  vars = NULL,
  FUN = NULL,
  FUN.SelectVars = NULL,
  FUN.SelectBlocks = NULL
)

Arguments

Value

The processed MultiBlock object, with data matrices transformed and/or blocks/variables removed according to the supplied arguments. Blocks not listed in blocks are left unchanged.

See Also

MultiBlock, MultiBlock2Matrix

Examples

b1 <- matrix(rnorm(500), 10, 50)
b2 <- matrix(rnorm(800), 10, 80)
mb <- MultiBlock(Data = list(b1 = b1, b2 = b2))
# Normalize each block to 0-100 range
BY100 <- function(x) 100 * (x - min(x)) / (max(x) - min(x))
mb <- ProcessMultiBlock(mb, FUN = BY100)
# Keep only variables with non-zero variance
mb <- ProcessMultiBlock(mb, FUN.SelectVars = function(x) apply(x, 2, var) > 0)
# Remove blocks where all values are below 0.5
mb <- ProcessMultiBlock(mb, FUN.SelectBlocks = function(x) max(x) >= 0.5)

Description

RNAseq data on cell differentiation

Usage

data(dataset3)

Format

An object of class matrix (inherits from array) with 36 rows and 12762 columns.

Author(s)

Cabrero et al. (2019) Scientific data 6:256 (doi:10.1038/s41597-019-0202-7)

Source

NCBI

Examples

data(dataset3)
#' mb_d3 <- MultiBlock(Data = list(rnaseq = rnaseq))

Description

RNAseq data of cell extracts

Usage

data(mouse_ds)

Format

An object of class matrix (inherits from array) with 12 rows and 9254 columns.

Author(s)

Radic Shechter et al. (2021) Molecular Systems Biology 17:e10141 (doi:10.15252/msb.202010141)

Source

SEQOUT

Examples

data(mouse_ds)
# MB <- MultiBlock(Data = list(RNAseq3 = RNAseq3,
#   lipids = lipids, intra = intra, extra = extra))

Description

Return the sample names of a MultiBlock object.

Usage

sampleNames(x)

## S4 method for signature 'MultiBlock'
sampleNames(x)

Arguments

Value

A vector with the sample names.

Examples

b1 <- matrix(rnorm(500), 10, 50)
rownames(b1) <- paste0("s", 1:10)
mb <- MultiBlock(Data = list(b1 = b1))
sampleNames(mb) # "s1" ... "s10"

Description

Set the sample names of a MultiBlock object.

Usage

sampleNames(x) <- value

## S4 replacement method for signature 'MultiBlock'
sampleNames(x) <- value

Arguments

Value

The updated MultiBlock object.

Examples

b1 <- matrix(rnorm(500), 10, 50)
mb <- MultiBlock(Data = list(b1 = b1))
sampleNames(mb) <- paste0("patient_", 1:10)
sampleNames(mb)

Description

Finds the important variables presenting a coordinated response across all specified replicate-blocks for a given ComDim component.

Usage

SelectFeaturesRW(
  RW = RW,
  results = results,
  ndim = NULL,
  blocks = NULL,
  threshold_cor = 1,
  threshold_cov = 1,
  mean.RW = TRUE,
  plots = "NO"
)

Arguments

Details

The function applies an S-plot approach to identify variables that are both strongly covarying and strongly correlated with the Q scores of the chosen component. For each block in blocks, covariance (s1) and correlation (s2) of every variable with the (pseudo-inverse-scaled) Q scores are computed. A variable is considered important in a block if its absolute covariance exceeds threshold_cov * sd(s1) and its absolute correlation exceeds threshold_cor * sd(s2). Only variables that satisfy both criteria in all specified blocks simultaneously are returned. The sign of the local P-loadings is used to separate variables into positive and negative groups.

Value

A named list with two elements:

$positive

Integer indices (named by variable name) of the important variables presenting a positive relationship with the Q scores (positive covariance, positive correlation, and positive local P-loading) across all specified blocks.

$negative

Integer indices (named by variable name) of the important variables presenting a negative relationship with the Q scores (negative covariance, negative correlation, and negative local P-loading) across all specified blocks.

When plots is not "NO", S-plots are also displayed as a side effect: each plot shows covariance on the x-axis and correlation on the y-axis, with selected variables highlighted in red.

See Also

SplitRW, ComDim_PCA

Examples

b1 <- matrix(rnorm(500), 10, 50)
batch_b1 <- rep(1, 10)
b2 <- matrix(rnorm(800), 30, 80)
batch_b2 <- c(rep(1, 10), rep(2, 10), rep(3, 10))
mb <- MultiBlock(
  Data = list(b1 = b1, b2 = b2),
  Batch = list(b1 = batch_b1, b2 = batch_b2),
  ignore.names = TRUE, ignore.size = TRUE
)
rw <- SplitRW(mb)
results <- ComDim_PCA(rw, 2)
# Identify important variables for component 1 across replicate blocks 2, 3, and 4
features <- SelectFeaturesRW(RW = rw, results = results, ndim = 1, blocks = c(2, 3, 4))
# Use stricter thresholds and display S-plots side by side
features <- SelectFeaturesRW(RW = rw, results = results, ndim = 1, blocks = c(2, 3, 4),
                             threshold_cor = 2, threshold_cov = 2, plots = "together")