Background NanoString’s GeoMx Digital Spatial Profiling (DSP) technology enables profiling of gene or protein expression from fresh or archival tissues. Specific regions of interest (ROIs) are identified via fluorescently labeled visualization markers. Within a given ROI, oligonucleotide tags from labeled, incubated antibodies can be released by area of interest (AOI)-specific exposure to UV light. With DSP, multiple AOIs can be collected within an individual tissue and/or within an individual patient. As with other technologies, technical variation that needs to be accounted before meaningful conclusions can be drawn.¹ Herein, we discuss technical considerations for normalizing and examining DSP data with multiple within-sample observations. We have two goals: 1) determine how different technical artifacts affect raw protein or RNA counts 2) provide guidelines for normalization strategies based on the biological questions of interest. To address these, we examine a recent melanoma dataset to quantify protein expression levels in tumor and stroma AOIs and to determine associations of specific proteins with clinical benefit (CB) from immunotherapy.

Methods Seventy-nine segmented ROIs containing matched tumor and stroma compartments were examined from eight patients at baseline (range: 4–12 ROIs). Five of these patients showed CB, defined as complete response, partial response, or remaining progression-free for 6 months. Following UV cleavage, liberated oligonucleotide tags were collected via microcapillary into a microtiter plate, and then processed using the nCounter Prep Station and Digital Analyzer as per manufacturer instructions.

Results Each AOI included 57 protein counts and six categories of control molecules/metrics (e.g., isotype molecules, AOI-specific cellularity). Before normalization, we examined controls and excluded those showing correlations with CB or segmentation type. We compared different normalization strategies including area and isotype normalization, upper quartile, and RUV.² For each strategy, we used linear and negative binominal mixed models to correlate protein expression with CB status, segmentation type, or their interaction. Findings consistent throughout many analysis combinations included higher MART1 expression in the CB group, lower PD-L1 and Ki-67 in the CB group, and lower HLA-DR expression in tumor segments of the CB group.

Conclusions ROIs can vary in size, cellularity, and staining, and normalization is important to account for technical differences when quantifying expression in spatial profiling studies. Normalization choices can affect outcome, and it is important to check whether proposed control proteins are in fact unassociated with the biological factors of interest. Mixed modeling approaches can be used to simultaneously model variation between ROIs within a sample and determine differences between sample groups.

Trial Registration ClinicalTrials. gov NCT02731729

Ethics Approval The study protocol and amendments were approved by the IRB of each participating institute. Written informed consent was obtained from all patients before conducting any study-related procedures.

References

1. Abbas-Aghababazadeh F, Li, and Fridley BL. Comparison of normalization approaches for gene expression studies completed with high-throughput sequencing,’ PLoS One, vol. 13, no. 10, p. e0206312, Oct. 2018, [Online]. Available: https://doi.org/10.1371/journal.pone.0206312.

2. Risso D, Ngai J, Speed TP, and Dudoit S. ‘Normalization of RNA-seq data using factor analysis of control genes or samples. Nat. Biotechnol 2014;32(9):pp. 896–902, doi: 10.1038/nbt.2931.