Please visit the following website for more bioinformatics tools from Dr. Tao Wang’s lab: https://qbrc.swmed.edu/labs/wanglab
This work is initiated when Dr. Seongoh Park visited QBRC at UT Southwestern Medical Center in 2018.
This is a README file of the R package BayesianMIR. In our paper, we develop the Bayesian multiple instance regression model we call BMIR, applied to the multiple instance regression problem. For more details about the structure of data and the Bayesian modeling, we refer readers to our paper available here.
To install our package, you may simply execute the following codes:
# install.packages("devtools")
devtools::install_github("inmybrain/BayesianMIR", subdir = "BayesianMIR") # don't forget to specify subdir!If you come across a problem like this, please refer to this answer in that issue.
We give a toy example to apply the main function BMIR_sampler, which
performs random sampling from the joint posterior distribution.
We first generate simulated data under the bag-instance structure.
rm(list = ls())
nsample <- 200
ninst <- 5
nprime <- 1 # do not change
nfeature <- 5
set.seed(6)
npoint <- (ninst - nprime) * nfeature
bag <- list()
for(i in 1:nsample){
prime_inst <- matrix(runif(nfeature, -5, 5), ncol = nfeature)
nonprime_inst <- (-1)^rbinom(npoint, 1, 1/2) * runif(npoint, min = -10, max = 10)
nonprime_inst <- matrix(nonprime_inst, ncol = nfeature, byrow = F)
# nonprime_inst <- matrix(c(rnorm(npoint, -5, 1),
# rnorm(npoint, 5, 1)),
# nrow = 2, byrow = T)
# nonprime_inst <- nonprime_inst[cbind(sample(1:2, npoint, replace = T), 1:npoint)]
# nonprime_inst <- matrix(nonprime_inst, ncol = nfeature, byrow = F)
bag[[i]] <- as.data.frame(rbind(prime_inst, nonprime_inst))
}
beta_true <- rep(2, nfeature + 1)
label <- unlist(lapply(bag, function(feature){
beta_true[1] + as.matrix(feature[1,,drop = FALSE]) %*% beta_true[-1]
}))
label <- label + rnorm(nsample, mean = 0, sd = 1)To facilitate our analysis, we use Tidy_dataset function to tidy data
up. The first 100 samples are used for model estimation, and the others
will be test samples to validate the fitted model.
library(BayesianMIR)
#> Warning: replacing previous import 'ggplot2::margin' by 'randomForest::margin'
#> when loading 'BayesianMIR'
tidydata <- Tidy_dataset(label = label[1:100],
feature_inst = bag[1:100])
newtidydata <- Tidy_dataset(feature_inst = bag[-(1:100)])Applying MISummarize to the output of Tidy_dataset, we can get the
basic information about the dataset:
- the number of bags,
- the number of features,
- (summary of) the numbers of instances in bags (or bag sizes).
MISummarize(tidydata)
#> Number of bags : 100
#> Number of features : 5 (V1, V2, V3, V4, V5)
#> Number of instances in bags :
#> Min. 1st Qu. Median Mean 3rd Qu. Max.
#> 5 5 5 5 5 5A scatter plot for multiple instances is given below using
MIScatterPlot. Each dimension of covariates is plotted on the x-axis
along with bag-level responses (labels) on the y-axis. If you know which
instance(s) is(are) primary in each bag, then they would be
distinguished from non-primary instances.
MIScatterPlot(tidydata = tidydata,
bag_size = 5,
true_primary = lapply(1:tidydata$nsample, function(x) rep(c(T,F), c(1, ninst - 1)))
)
We can obtain the MCMC samples using BMIR_sampler. By default, the
first halves are discarded as the burn-in steps. You can thin the
samples by using nthin option and get multiple chains by using
nchain option. Please refer to the help page of the function to see
what it returns.
## BMIR model fitting
ntotal <- 20000
BMIR_fit <- BMIR_sampler(ntotal = ntotal, tidydata = tidydata)
#> =============================================================
#> Bayesian Multiple Instance Regression
#> Elapsed time for chain1=0.027 mins: MCMC sampling is done!Using the fitted model, you can specify the estimated status of being a
primary instance on the scatter plot provided by MIScatterPlot. You
can check how many overlaps are between the estimated and the truth.
MIScatterPlot(tidydata = tidydata,
bag_size = 5,
true_primary = lapply(1:tidydata$nsample, function(x) rep(c(T,F), c(1, ninst - 1))),
pred_primary = lapply(split(BMIR_fit$pip[,1], tidydata$membership), function(x) rank(-x, ties.method = "min") <= 1)
)
By slightly modifying ggs_density function from the package ggmcmc,
we can show one of the Bayesian inferences that BMIR does provide: the
highest posterior density intervals of parameters.
# install.packages("ggmcmc")
library("ggmcmc")
ggs_density <- function (D, ncol, family = NA, rug = FALSE, hpd = FALSE, greek = FALSE)
## - ncol is added!
## - ci -> ggmcmc::ci
## - [Low, High] interval is commented
{
if (!is.na(family)) {
D <- get_family(D, family = family)
}
if (attributes(D)$nChains <= 1) {
f <- ggplot(D, aes(x = value))
}
else {
f <- ggplot(D, aes(x = value, colour = as.factor(Chain),
fill = as.factor(Chain)))
}
f <- f + geom_density(alpha = 0.3) + scale_fill_discrete(name = "Chain") +
scale_colour_discrete(name = "Chain")
if (!greek) {
f <- f + facet_wrap(~Parameter, ncol = ncol, scales = "free")
}
else {
f <- f + facet_wrap(~Parameter, ncol = ncol, scales = "free",
labeller = label_parsed)
}
if (rug)
f <- f + geom_rug(alpha = 0.1)
if (hpd) {
ciD <- ggmcmc::ci(D)
f <- f + geom_segment(data = ciD, size = 2, color = "blue", inherit.aes = FALSE,
aes(x = low, xend = high, y = 0, yend = 0))
# +geom_segment(
# data = ciD,
# size = 1,
# inherit.aes = FALSE,
# aes(
# x = Low,
# xend = High,
# y = 0,
# yend = 0
# )
# )
}
return(f)
}
ggs_mcmc <- ggmcmc::ggs(BMIR_fit$mcmclist)
ggs_mcmc$Parameter <- factor(ggs_mcmc$Parameter, labels = c(paste0("coef", 1:(nfeature + 1)), "sig2_error"))
ggs_density(ggs_mcmc %>%
filter(Iteration > max(Iteration) / 2),
ncol = 2,
hpd = TRUE) +
geom_vline(data = data.frame(Parameter = c(paste0("coef", 1:(nfeature + 1)), "sig2_error"),
true_val = c(rep(2, 1 + nfeature), 1)),
aes(xintercept = true_val), color = "red") +
labs(x = "Value", y = "Density") +
theme(axis.text.y = element_blank(),
axis.ticks.y = element_blank())
When new bags are given, BMIR can both predict labels and identify
primary instances using predict.BMIR. By default, predict.BMIR
depends on randomForest function from the package randomForest,
which helps identifying primary instances in new bags. If you specify
k (the number of primary instances in each new bag) larger than 1,
then the selected primary instances will be aggregated by their average.
pred_fit <- predict.BMIR(BMIRchain = BMIR_fit$mcmclist$Chain1,
pip = BMIR_fit$pip[,1],
tidydata = tidydata,
newtidydata = newtidydata,
k = 1)Let us see how prediction works.
ggplot(data = data.frame(pred = pred_fit$newtidydata$label,
true = label[-(1:100)]),
mapping = aes(x = pred, y = true)) +
geom_point() + geom_abline(intercept = 0, slope = 1, color = "red")
To cite this package, please use this bibtex format:
@article{Park:2020,
author = {Seongoh Park and Xinlei Wang and Johan Lim and Guanghua Xiao and Tianshi Lu and Tao Wang},
title ={Bayesian multiple instance regression for modeling immunogenic neoantigens},
journal = {Statistical Methods in Medical Research},
volume = {29},
number = {10},
pages = {3032-3047},
year = {2020},
doi = {10.1177/0962280220914321},
note ={PMID: 32401701},
URL = {https://doi.org/10.1177/0962280220914321},
eprint = {https://doi.org/10.1177/0962280220914321}
}We are happy to troubleshoot any issue with the package;
-
please contact to the maintainer by seongohpark6@gmail.com, or
-
please open an issue in the github repository.