Differential expression methods for NanoString nCounter Gene Expression

Methods used for raw data (RCC files):

Differential expression between two groups of samples is calculated using the Generalized Linear Model (GLM) for count data that was developed by the NanoString Biostatistics team (as originally implemented in nSolver Advanced Analysis software). The model assumes a negative binomial distribution and utilizes the raw data and estimates of noise and dispersion across all samples in the experiment to calculate fold-change and p-value for each gene. 

Additional details of this model (the Fast method) are available on page 50 of the nSolver Advanced Analysis 2.0 User Manual. Adjustment of p-values to account for multiple testing is performed using the Benjamini-Hochberg false discovery rate (FDR) method.

Frequently Asked Questions:

Q: Can fold change be calculated directly from the normalized data? 

A: Fold change is estimated from the statistical model developed by NanoString and not from direct comparison of means across sample groups. The statistical model uses information from the raw data, and incorporates noise and dispersion metrics from across the entire dataset into the fold-change calculation. 

Q: How is it possible for the fold change to be so large when the difference in mean expression between the sample groups is relatively small?

A: In some cases, depending on the power of the study and where a given gene is expressed in the dynamic range of the assay, the difference in fold change between NanoString’s Generalized Linear Model and their legacy approach (t-test, assuming continuous distribution) can be relatively large.

Methods used for normalized counts (CSV file):

The limma R library(1) was used to calculate fold changes and p-values and perform optional covariate correction. 

1. Ritchie, M. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res, 43(7) (2015).