Integrative Bayesian Multiple Regression for Multi-Platform Biomarkers

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) ; 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.

IntegMultiReg

Integrative Bayesian Multiple Regression for Multi-Platform Biomarkers.

IntegMultiReg implements the integrative multi-regression (IMR) model of Chekouo, Stingo, Doecke and Do (2017, Biometrics) and extends it from time-to-event outcomes to continuous (Gaussian) and binary (probit) outcomes.

Given several molecular platforms measured on overlapping but partially missing sets of subjects, IMR partitions subjects into the availability subgroups of a Venn diagram, fits one regression per subgroup, and shares information across availability subgroups through

• non-local (product moment) priors on the regression coefficients, and
• a Markov random field (MRF) prior on the variable-selection indicators,

so that no subject with partially observed platforms is discarded and the same biomarkers tend to be selected across availability subgroups.

Installation

The package contains C code that links against the GNU Scientific Library (GSL), which must be installed first:

• macOS: brew install gsl
• Debian/Ubuntu: sudo apt-get install libgsl-dev
• Windows: GSL is provided by Rtools.

Once available on CRAN, install the package with:

install.packages("IntegMultiReg")

Alternatively, install a local source tarball:

install.packages("IntegMultiReg_0.1.3.tar.gz", repos = NULL, type = "source")

The CRAN checking tools checkbashisms and qpdf are not runtime dependencies. Package users do not need them. Maintainers running R CMD check --as-cran locally can install them with brew install checkbashisms qpdf on macOS or sudo apt-get install devscripts qpdf on Debian/Ubuntu.

Quick start

library(IntegMultiReg)
data("simIMR")

fit <- imr(
  platform_data_list = simIMR$platforms,
  outcome            = simIMR$outcome.binary,
  cov                = simIMR$covariates,
  type_outcome       = "binary",
  nu                 = c(-4, -3, -4),
  sample_mcmc        = c(2000, 1000),
  ssize              = 30,
  seed               = 1
)

fit                       # short summary
summary(fit)              # selected biomarkers per platform
coef(fit)                 # per-platform mPIP matrices
plot(fit, type = "selection")
plot_top_features(fit)    # ranked biomarker bar chart
predict(fit, newdata = simIMR$platforms[1:2], covariates = simIMR$covariates)
cv_imr(fit)               # fold-split predictive assessment using fitted samples

Real-data example

kircIMR is a reduced public UCSC Xena TCGA-KIRC survival example aligned with the Biometrics kidney cancer case study: mRNA expression, miRNA expression, DNA methylation, clinical covariates and right-censored survival. It is derived from public UCSC Xena TCGA-KIRC sampleMap files, not from controlled-access TCGA/GDC files, and contains only a reduced Cox-screened feature panel.

The package replaces TCGA barcodes with package-internal IDs such as KIRC001 and does not distribute a barcode mapping. Users should not attempt participant re-identification or linkage to external resources.

data("kircIMR")
sapply(kircIMR$platforms, dim)
kircIMR$model_subgroup_sizes

kirc_fit <- imr(
  kircIMR$platforms,
  kircIMR$outcome.survival,
  cov = kircIMR$covariates,
  type_outcome = "right.censored",
  nu = c(-4, -3, -4),
  sample_mcmc = c(4000, 1000),
  ssize = 30,
  seed = 1
)

See the package vignette vignette("IntegMultiReg") for a complete walk-through.

Reference

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. https://doi.org/10.1111/biom.12587

When using kircIMR, please also acknowledge TCGA, the National Cancer Institute Genomic Data Commons, and UCSC Xena as the public data sources.

Coefficient and predictive uncertainty (0.1.3)

After fitting, posterior_draws(fit) adds conditional pMOM coefficient and variance draws to the retained selection models. summary(draws) and confint(draws) report coefficient intervals including point mass at zero for inactive molecular features. Clinical covariates remain always included.

# Use an adequately explored fit and inspect both stages of diagnostics.
draws <- posterior_draws(fit, seed = 2)
draws$diagnostics
confint(draws)
predict(draws, simIMR$platforms, covariates = simIMR$covariates,
        type = "mean")       # uncertainty in the conditional response mean
predict(draws, simIMR$platforms, covariates = simIMR$covariates,
        type = "response")   # uncertainty in a future outcome

These are approximate model-averaged intervals: selection weights retain the original Laplace approximation. Conditional split R-hat does not assess the original selection chain; increase simulation effort when diagnostics are poor. Coefficients use subgroup-standardized predictor scales. Binary probability intervals use type = "mean"; binary response intervals are discrete.

Survival migration from 0.1.2

Version 0.1.3 logs positive event and censoring times once, matching the original Biometrics supplementary code. Supply raw times, including positive times below one; refit previous survival models. The optional survival_scale = "identity" reproduces the historical raw-time implementation. CV partitioning is unchanged.

For log-time fits, predict(fit, ...) returns a log-time point prediction. predict(draws, ...) returns time-scale intervals and median point summaries. The time-scale posterior mean need not exist under the variance mixture; it is not estimated by averaging exponentiated draws. With type = "mean" the interval summarizes conditional mean time, whereas type = "response" includes future outcome variability. The censoring process for future observations is not modeled.