Package {IntegMultiReg}

Title:	Integrative Bayesian Multiple Regression for Multi-Platform Biomarkers
Version:	0.1.0
Description:	A Bayesian framework that integrates several regression models to identify a parsimonious set of biomarkers shared across disparate data platforms (for example genomic, transcriptomic and proteomic assays). Subjects are partitioned into subgroups defined by their pattern of platform availability, so that no subject with partially missing platform data is excluded, and information is borrowed across subgroups through a Markov random field prior on the variable-selection indicators together with non-local (product moment) priors on the regression effects. The methodology was introduced for time-to-event outcomes by Chekouo, Stingo, Doecke and Do (2017) <doi:10.1111/biom.12587>; this package additionally supports continuous (Gaussian) and binary (probit) outcomes. Posterior inference is carried out by a Markov chain Monte Carlo sampler implemented in C for computational efficiency.
License:	GPL (≥ 3)
Encoding:	UTF-8
LazyData:	true
Depends:	R (≥ 4.0.0)
Imports:	graphics, grDevices, stats, utils
Suggests:	knitr, rmarkdown, testthat (≥ 3.0.0)
SystemRequirements:	GSL (>= 2.0)
VignetteBuilder:	knitr
Config/testthat/edition:	3
NeedsCompilation:	yes
Config/roxygen2/version:	8.0.0
RoxygenNote:	7.3.2
Packaged:	2026-08-05 01:44:33 UTC; tchekouo
Author:	Sinian Zhang [aut], Jianfeng Wang [aut], Thierry Chekouo [aut, cre]
Maintainer:	Thierry Chekouo <tchekouo@umn.edu>
Repository:	CRAN
Date/Publication:	2026-08-09 08:20:07 UTC

IntegMultiReg: Integrative Bayesian Multiple Regression for Multi-Platform Biomarkers

Description

The IntegMultiReg package implements the integrative multi-regression (IMR) approach of Chekouo et al. (2017) and extends it from time-to-event outcomes to continuous (Gaussian) and binary (probit) outcomes. Subjects are partitioned into availability subgroups according to which platforms they have measured, one regression model is built per subgroup, and information is shared across availability subgroups through a Markov random field prior on the variable-selection indicators combined with non-local priors on the regression coefficients.

Details

The main entry points are:

imr

Fit the model by MCMC and return an object of class "imr".

predict.imr

Predict outcomes for new subjects.

cv_imr

Assess predictive performance by cross-validation.

Standard methods print.imr, summary.imr, coef.imr, plot.imr and predict.imr are provided for the fitted object, together with the stand-alone displays plot_top_features and plot_subgroup_sizes.

Jianfeng Wang initiated the R interface that calls the C implementation.

Author(s)

Maintainer: Thierry Chekouo tchekouo@umn.edu (ORCID)

Authors:

• Sinian Zhang zhan9381@umn.edu

• Jianfeng Wang wan01188@umn.edu

References

Chekouo T, Stingo FC, Doecke JD, Do K-A (2017). "A Bayesian Integrative Approach for Multi-Platform Genomic Data: A Kidney Cancer Case Study." Biometrics, 73(2), 615–624. doi:10.1111/biom.12587

Marginal Posterior Inclusion Probabilities of an IMR Fit

Description

Extracts the posterior mean variable-selection probabilities (the marginal posterior inclusion probabilities, mPIP) of a fitted model.

Usage

## S3 method for class 'imr'
coef(object, ...)

coef_imr(object, ...)

Arguments

Value

A named list with one matrix per platform. Rows are the subgroups containing that platform (labelled by their availability bitstrings) and columns are the platform features.

See Also

[imr()]

Cross-Validated Predictive Accuracy of an IMR Fit

Description

'cv_imr()' evaluates predictive accuracy of a model fitted with [imr()] using repeated K-fold splits and Bayesian model averaging over the fitted MCMC samples. The sampler is not re-run inside each training fold; fit a separate object with 'method = "BMS"' if a no-borrowing comparison is needed. The accuracy measure depends on the outcome type: the concordance index (C-index) for right-censored outcomes, the area under the ROC curve (AUC) for binary outcomes, and the mean squared error (MSE) for continuous outcomes.

Usage

cv_imr(
  object,
  k = 5,
  rounds = 2,
  method = NULL,
  max_models = 100,
  verbose = FALSE
)

Arguments

Value

A list with two numeric matrices, each of dimension 'rounds' x '(n_subgroups + 1)', whose last column corresponds to all subjects pooled and whose remaining columns are named by the availability subgroup bitstrings:

The accuracy metric ('"C-index"', '"AUC"' or '"MSE"') is recorded in the '"metric"' attribute of the returned list.

See Also

[imr()], [predict.imr()]

Examples

data("simIMR", package = "IntegMultiReg")
fit <- imr(
  platform_data_list = simIMR$platforms, outcome = simIMR$outcome,
  cov = simIMR$covariates, type_outcome = "binary",
  nu = c(-4, -3, -4), sample_mcmc = c(200, 100), ssize = 5, seed = 1
)
cv <- cv_imr(fit, k = 5, rounds = 2)
cv$total_cindex

Fit the Integrative Bayesian Multi-Platform Regression Model

Description

'imr()' fits the integrative multi-regression (IMR) model of Chekouo et al. (2017) by Markov chain Monte Carlo (MCMC). It identifies biomarkers associated with a time-to-event, binary or continuous outcome while borrowing information across all subjects, regardless of which platforms each subject has measured. Subjects are partitioned into availability subgroups defined by their pattern of platform availability; one regression is built per subgroup and information is shared across subgroups through a Markov random field (MRF) prior on the variable-selection indicators together with non-local priors on the regression coefficients. The sampler is implemented in C for efficiency.

The arguments are grouped by the orthogonal aspect of the analysis that each one controls: the *data* ('platform_data_list', 'outcome', 'cov'), the *likelihood* ('type_outcome'), the *model* ('method'), subgroup *filtering* ('ssize'), the *priors* ('nu', 'hh', 'h0', 'sig_alpha_psi', 'thet_alph_bet'), the *computation* ('sample_mcmc', 'seed') and the *output* ('verbose').

Usage

imr(
  platform_data_list,
  outcome,
  cov = NULL,
  type_outcome = c("right.censored", "binary", "continuous"),
  method = c("IMR", "BMS"),
  ssize = 30,
  nu = rep(-3, length(platform_data_list)),
  hh = 0.087,
  h0 = 10000,
  sig_alpha_psi = c(0.001, 0.001),
  thet_alph_bet = c(40, 10),
  sample_mcmc = c(2000, 1000),
  seed = NULL,
  verbose = FALSE
)

Arguments

Details

All feature and covariate data are standardized internally (mean 0, standard deviation 1); the centring and scaling factors are stored in the returned object so that [predict.imr()] can apply the same transformation to new subjects. For right-censored outcomes the latent log-survival times of censored subjects are imputed within the sampler; for binary outcomes a probit data-augmentation latent variable is sampled.

Value

An object of class '"imr"': a list with components

References

Chekouo T, Stingo FC, Doecke JD, Do K-A (2017). "A Bayesian Integrative Approach for Multi-Platform Genomic Data: A Kidney Cancer Case Study." Biometrics, 73(2), 615–624. doi:10.1111/biom.12587

See Also

[predict.imr()], [cv_imr()], [summary.imr()], [plot.imr()]

Examples

data("simIMR", package = "IntegMultiReg")
fit <- imr(
  platform_data_list = simIMR$platforms,
  outcome = simIMR$outcome,
  cov = simIMR$covariates,
  type_outcome = "binary",
  nu = c(-4, -3, -4),
  sample_mcmc = c(200, 100),
  ssize = 5,
  seed = 1
)
fit

Reduced TCGA-KIRC Multi-Platform Survival Example

Description

A reduced real-data example derived from public UCSC Xena TCGA kidney renal clear cell carcinoma (KIRC) data, not from controlled-access TCGA/GDC files. The data set mirrors the multi-platform structure of Chekouo, Stingo, Doecke and Do (2017): mRNA expression, miRNA expression and DNA methylation are observed on overlapping but not identical sets of patients, with clinical covariates and right-censored survival outcome.

Usage

kircIMR

Format

A named list with the components:

platforms

A list of three data frames, 'mrna', 'mirna' and 'methylation'. Each has an 'id' column containing package-internal patient labels followed by screened molecular features.

covariates

Clinical covariates with an 'id' column: age at initial diagnosis, female sex indicator, pathologic stage and histologic grade.

outcome

Alias of 'outcome.survival', for convenience.

outcome.survival

Data frame with 'id', observed survival/follow-up time in days and event indicator 'status' (1 = deceased, 0 = censored).

feature_screening

Per-platform feature-ranking information and a short description of the screening procedure used to create the reduced panel.

platform_availability

Patient-level logical matrix indicating which of the three platforms are available.

subgroup_sizes

Counts of the non-empty availability subgroups induced by the reduced data.

model_subgroup_sizes

The availability subgroups that would be modelled under the default real-data illustration threshold 'ssize = 30'.

paper_alignment

Metadata describing how the reduced example maps to the Biometrics KIRC case study.

source

Source URLs and reference metadata.

Details

The full TCGA source files are not distributed with the package. Instead, this object stores a small screened feature panel suitable for examples and vignette analyses. Patient identifiers are package-internal labels ('KIRC001', 'KIRC002', ...) rather than TCGA barcodes; no barcode mapping is distributed with the package. Users should not attempt participant re-identification or linkage to external resources. The generating script is provided in the source repository under 'data-raw/make_kircIMR.R'.

The source data are public UCSC Xena TCGA-KIRC sampleMap files: mRNA expression ('HiSeqV2'), miRNA expression ('miRNA_HiSeq_gene'), DNA methylation ('HumanMethylation450') and the KIRC clinical matrix. Molecular features are screened with univariable Cox models after basic filtering, and missing molecular values in the retained panel are median-imputed within feature. The full Biometrics analysis used 776 mRNA, 91 miRNA and 729 DNA methylation features; this package example retains 50, 30 and 50 features, respectively, to keep examples lightweight. The package does not distribute controlled-access TCGA/GDC data or the full Biometrics/Wiley supplementary data files.

References

Chekouo T, Stingo FC, Doecke JD, Do K-A (2017). "A Bayesian Integrative Approach for Multi-Platform Genomic Data: A Kidney Cancer Case Study." Biometrics, 73(2), 615–624. doi:10.1111/biom.12587

The Cancer Genome Atlas Research Network; National Cancer Institute Genomic Data Commons; UCSC Xena.

Examples

data("kircIMR", package = "IntegMultiReg")
sapply(kircIMR$platforms, dim)
kircIMR$subgroup_sizes
kircIMR$model_subgroup_sizes

Plot Method for IMR Fits

Description

Visualizes a fitted '"imr"' object. Three plot types are available:

'"selection"'

Heatmap of the marginal posterior inclusion probabilities (mPIP), one panel per platform, with features on the horizontal axis, availability subgroups on the vertical axis and a small intensity legend showing that darker values are closer to 1.

'"theta"'

Heatmap of the posterior mean MRF interaction parameters between availability subgroups, one panel per platform.

'"trace"'

Trace plot of the log-posterior across MCMC iterations.

See [plot_top_features()] and [plot_subgroup_sizes()] for two further ready made displays.

Usage

## S3 method for class 'imr'
plot(
  x,
  type = c("selection", "theta", "trace"),
  platform = NULL,
  base_cex = 1,
  cex_axis = NULL,
  cex_lab = NULL,
  cex_main = NULL,
  col = NULL,
  palette = NULL,
  legend = TRUE,
  legend_width = 0.28,
  mar = NULL,
  mgp = NULL,
  ...
)

plot_imr(x, ...)

Arguments

Value

'NULL', invisibly; called for the side effect of producing a plot.

See Also

[imr()], [plot_top_features()], [plot_subgroup_sizes()]

Plot the Availability Subgroup Sizes of an IMR Fit

Description

Draws a bar chart of the number of subjects in each modelled availability subgroup (the non-empty regions of the Venn diagram), giving a quick picture of how the sample is distributed across subgroups.

Usage

plot_subgroup_sizes(
  object,
  base_cex = 1,
  cex_axis = NULL,
  cex_lab = NULL,
  cex_main = NULL,
  cex_values = NULL,
  col = NULL,
  palette = "subgroup",
  show_values = TRUE,
  ylim = NULL,
  mar = NULL,
  mgp = NULL,
  ...
)

Arguments

Value

Invisibly, the named integer vector of subgroup sizes.

See Also

[imr()], [plot.imr()], [plot_top_features()]

Examples

data("simIMR", package = "IntegMultiReg")
fit <- imr(
  platform_data_list = simIMR$platforms, outcome = simIMR$outcome,
  cov = simIMR$covariates, type_outcome = "binary",
  nu = c(-4, -3, -4), sample_mcmc = c(200, 100), ssize = 5, seed = 1
)
plot_subgroup_sizes(fit)

Plot the Top Selected Features of an IMR Fit

Description

Draws a horizontal bar chart of the features with the highest marginal posterior inclusion probability (mPIP). For each platform-feature pair, the plotted score is the highest mPIP attained among the availability subgroups containing that platform; the displayed bars are the largest such scores across all platforms. This complements the per-platform heatmap of 'plot(fit, type = "selection")'.

Usage

plot_top_features(
  object,
  top = 10,
  base_cex = 1,
  cex_names = NULL,
  cex_axis = NULL,
  cex_lab = NULL,
  cex_main = NULL,
  cex_legend = NULL,
  col = NULL,
  palette = "platform",
  reference = 0.5,
  show_source = TRUE,
  legend = TRUE,
  xlim = c(0, 1),
  mar = NULL,
  mgp = NULL,
  ...
)

Arguments

Value

Invisibly, a data frame of the displayed features with columns 'platform', 'feature', 'mpip' and 'subgroup', where 'mpip' is the maximum subgroup mPIP for that platform-feature pair and 'subgroup' is the subgroup where the maximum is attained. Rows are ordered by decreasing mPIP.

See Also

[imr()], [plot.imr()], [plot_subgroup_sizes()]

Examples

data("simIMR", package = "IntegMultiReg")
fit <- imr(
  platform_data_list = simIMR$platforms, outcome = simIMR$outcome,
  cov = simIMR$covariates, type_outcome = "binary",
  nu = c(-4, -3, -4), sample_mcmc = c(200, 100), ssize = 5, seed = 1
)
plot_top_features(fit, top = 8)

Predict Outcomes for New Subjects

Description

'predict()' method for objects of class '"imr"' produced by [imr()]. New subjects are matched to the availability subgroup models learned during training, missing platforms are handled automatically, and the test features are standardized with the training set's centring and scaling factors before Bayesian model averaging.

Usage

## S3 method for class 'imr'
predict(
  object,
  newdata,
  platform_names = NULL,
  covariates = NULL,
  max_models = 100,
  verbose = FALSE,
  ...
)

predict_imr(object, newdata, ...)

Arguments

Details

Predictions are only produced for subjects observed on at least one platform (and, when 'covariates' is supplied, with covariate data). For '"binary"' outcomes the returned 'predict' column is a probability obtained through the probit link ('pnorm'); for '"continuous"' and '"right.censored"' outcomes it is the predicted (latent) response.

Value

A named list with one data frame per active availability subgroup model. Each data frame has columns 'id' (subject identifier) and 'predict' (predicted value or probability, rounded to three decimals).

See Also

[imr()], [cv_imr()]

Examples

data("simIMR", package = "IntegMultiReg")
fit <- imr(
  platform_data_list = simIMR$platforms, outcome = simIMR$outcome,
  cov = simIMR$covariates, type_outcome = "binary",
  nu = c(-4, -3, -4), sample_mcmc = c(200, 100), ssize = 5, seed = 1
)
new_x <- simIMR$platforms[[1]][1:10, ]
new_z <- simIMR$platforms[[2]][1:7, ]
predict(fit, newdata = list(new_x, new_z), covariates = simIMR$covariates)

Print Method for IMR Fits

Description

Prints a compact overview of the fit, including a platform key ('P1', 'P2', ...) that decodes the availability-subgroup bitstrings. Optionally, it can also print the top-ranked features per platform, ranked by each feature's maximum mPIP over availability subgroups containing that platform.

Usage

## S3 method for class 'imr'
print(x, threshold = 0.5, rank = FALSE, top = 5, ...)

Arguments

Value

'x', invisibly.

Simulated Multi-Platform Example Data

Description

A small simulated data set illustrating the integrative multi-regression (IMR) setting: three molecular platforms measured on overlapping but partially missing sets of subjects, clinical covariates available for everyone, and three outcome types (continuous, binary and right-censored) driven by the same latent signal.

Usage

simIMR

Format

A named list with the components:

platforms

A list of three data frames, 'genomic' (20 features, subjects 1–240), 'proteomic' (10 features, subjects 1–180) and 'metabolomic' (8 features, subjects 121–300). Each has an 'id' column followed by the feature columns.

covariates

A data frame of three clinical covariates ('age', 'sex', 'stage') with an 'id' column, available for all 300 subjects.

outcome

The binary outcome data frame (alias of 'outcome.binary'), used as the canonical demonstration outcome.

outcome.binary

Data frame with 'id' and a 0/1 response 'y'.

outcome.continuous

Data frame with 'id' and a numeric response 'y'.

outcome.survival

Data frame with 'id', observed 'time' and event indicator 'status' (1 = event, 0 = censored).

truth

A list giving, per platform, the column indices of the features with non-zero effects used to generate the outcomes.

Details

The platforms are observed on the following subjects, which induces four non-empty availability subgroups (bitstrings over genomic/proteomic/ metabolomic): '011' (120 subjects), '111' (60), '101' (60) and '100' (60).

The generating script is provided in 'data-raw/make_simIMR.R'.

Examples

data("simIMR", package = "IntegMultiReg")
sapply(simIMR$platforms, dim)
str(simIMR$truth)

Summarize an IMR Fit

Description

Produces a per-platform summary of the selected features (those whose marginal posterior inclusion probability exceeds 'threshold' in at least one subgroup), ranked by their maximum inclusion probability.

Usage

## S3 method for class 'imr'
summary(object, threshold = 0.5, ...)

## S3 method for class 'summary.imr'
print(x, ...)

summary_imr(object, ...)

Arguments

Value

An object of class '"summary.imr"': a list with the run metadata and, for each platform, a data frame of selected features with their maximum mPIP and the subgroup achieving it.