MetaNorm is a normalization procedure for Nanostring nCounter datasets.

Please refer to our paper for more details.

Besides the R helper as well as the vignette we provide along with the package, we have also built a detailed online documentation where we guide you step-by-step on how to conduct meta analysis as well as normalizing a NanoString nCounter dataset.

• R>=4.0.2

For the stable version:

library(devtools)
install_github("jbarth216/MetaNorm",subdir="MetaNorm", ref="main")

or if you want to try out the development version:

install_github("Yuqiu-Yang/MetaNorm", subdir="MetaNorm", ref="main")

Other than making sure your R version is correct, there is no need for you to manually install other packages as they will be automatically installed when installing our package.

Along with the package, we have provided two example datasets meta_analysis_data.RData and normalization_data.RData. We will use them to quickly demonstrate how to get started with our package.

In the meta_analysis_data.RData, we provide positive probe data for the 13 collected datasets. To curate the data and generate empirical estimated coefficients based on this data simply run the following.

library(MetaNorm)
data("meta_analysis_data")
ds = curate_data(dataset=ds)
results = find_regression_coefs(df=ds)

The purpose of the meta-analysis is to provide posterior estimates to plug in to prior distributions of model parameters in the MetaNorm procedure.

The analysis is based on a complex Bayesian hierarchical model, similar to the ones used in MetaNorm and RCRnorm. The model is designed specifically for these datasets, and while not mean to be reproduced with other data, can certainly be a guide for similar analyses.

To run the Gibbs sampler, use the following code:

draw = meta_analysis(results$df, results$coeffs2, M=12000, n_keep=-1)

The result is a dataframe containing posterior sample draws of major variables. M specifies how many samples to draw while n_keep tells the program how many samples to keep. A negative number means to save all the samples.

To run several MCMC chains, simply repeat the previous code several times.

draw1 = meta_analysis(results$df, results$coeffs2, M=12000, n_keep=-1)
draw2 = meta_analysis(results$df, results$coeffs2, M=12000, n_keep=-1)

Once you plot the posterior draws, you will see something like this:

To carry out a normalization procedure for a NanoString nCounter dataset, you will need the MetaNorm function. We have also included a sample data used by RCRnorm for demonstration normalization_data.RData.

Please make sure that your datasets conform to our format. Once you load the dataset

data("normalization_data")

you will see the data is a list of 4 dataframes containing:

• '$pos_dat', a dataframe of positive probe read counts. Must have 6 rows and n columns (where n is the number of samples)
• '$neg_dat', a dataframe of negative probe read counts. Can have any number of rows (for varying negative probe amounts) and n columns
• '$hk_dat', a dataframe of housekeeping gene read counts (specified by researcher(s)). Can have any number of rows and n columns
• '$reg_dat', a dataframe of "regular" (i.e., all other) gene read counts. Can have any number of rows and n columns

Any deviation from this format will result in an error, or faulty results.

Once you have prepared your dataset, to perform MetaNorm simply run:

draw = MetaNorm(dat=dat, M=5000, all_draws=TRUE)

As with the meta_analysis function, M is the number of draws for each parameter, inclduing burn-in. The default of 15,000 is conservative, as MetaNorm can produce stable estimates often with only 2-3k draws total. all_draws=TRUE tells the algorithm to keep all the MCMC draws for all parameters (which is A LOT).

For most users, the only parameters of interest are the posterior means for kappa_hk and kappa_reg. Of course, you are welcome to compute the posterior summary statistics by yourself based on the MCMC draws. However, as the data structure is rather convoluted, we provide along with the MetaNorm function additional parameters to only return the parameters of interest. To enable this mode of inference, simply set all_draws=FALSE and provide some basic information on how many sample are for burn in and how to thin the MCMC draws.

output = MetaNorm(dat=dat, M=5000, all_draws=FALSE, burn_in=1000, thin=2)

By setting burn_in=1000, we are telling the algorithm to discard the first 1000 samples when computing the summary statistics. thin=2 tells the algorithm to only collect every other sample to reduce auto-correlation.

draws <- seq(burn_in+1, M, by=thin)

kappa_reg_stat <- apply(draw$kappa_reg[draws,,],c(2,3),mean)
kappa_hk_stat <- apply(draw$kappa_hk[draws,,],c(2,3),mean)

Currently, due to the design of the algorithm, even if we are only returning the posterior means, we still store all the MCMC sample draws internally. This would consume lots of memory space which would be a serious issue if we are working with a large dataset. To alleviate the issue, we provide a function: MetaNorm_LowMemory that will only keep the most recent MCMC draws for all parameters except for global parameters. Along with the posterior samples for global parameters, we will also compute the summary statistics for kappa_hk and kappa_reg. To use this function, run the following code

output = MetaNorm_LowMemory(dat=dat,  M=5000, burn_in = 1000, thin=2)

The output is a list containing two items: GlobalParameters_Draws (all draws for mu_a, mu_b, sig2_a, sig2_b) and Posterior_estimates (list of kappa_hk, kappa_reg posterior estimates).

The other two parameters are

• mm (default is 3) The radius (in standard deviations) for parameters with uniform priors
• seed (default is 0) Random seed, for reproducibility. If running multiple chains to check convergence/stability, make sure that each chain has a different seed.

For more details, please refer to our online documentation

The artical is currently published at Bioinformatics. To cite

@article{10.1093/bioinformatics/btae024,
    author = {Barth, Jackson and Yang, Yuqiu and Xiao, Guanghua and Wang, Xinlei},
    title = {MetaNorm: incorporating meta-analytic priors into normalization of NanoString nCounter data},
    journal = {Bioinformatics},
    volume = {40},
    number = {1},
    pages = {btae024},
    year = {2024},
    month = {01},
    issn = {1367-4811},
    doi = {10.1093/bioinformatics/btae024},
    url = {https://doi.org/10.1093/bioinformatics/btae024},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/40/1/btae024/56466495/btae024.pdf},
}

If you have any questions or suggestions, feel free to contact us 😃

• Dr. Xinlei (Sherry) Wang faculty page
• Dr. Jackson Barth faculty page
• Yuqiu Yang
• Dr. Guanghua (Andy) Xiao lab

• SMU M3 high-performance computing (HPC) cluster