A total of 35 FFPE tissue samples from 2012 to 2021 were obtained from the archive of the Institute of Veterinary Pathology. Specimens had originally been surgically excised from privately owned pet dogs and submitted for individual diagnostic and therapeutic purposes. Dog owners had given their consent and the work was ethically approved by the State Office for Health and Social Affairs, Berlin (StN 010/23). The tumors (AGASAC: n = 15, HGA: n = 10, CCH: n = 5 early stage, and n = 5 late stage) were selected based on histopathological diagnosis on hematoxylin and eosin stained slides by a board-certified veterinary pathologist. Tumor-adjacent tissue was removed to obtain largely homogeneous tumor cell masses. Further information on the individual tissue samples is provided in Supplementary Table S1.

Five 10 μm FFPE scrolls were prepared from entire cross sections, collected in sterile centrifuge tubes and stored at −80°C. Total RNA was extracted using the PureLink™ FFPE Total RNA Isolation Kit (Thermo Fisher) according to the manufacturer’s guidelines. Total RNA concentrations were measured with the NanoDrop™ 2000c spectrophotometer (Thermo Fisher) and quality was determined using the Agilent 5200 Fragment Analyzer (Agilent Technologies, Inc., Santa Clara, United States) employing the DNF-471F33 - SS Total RNA 15 nt - FFPE Illumina DV200 method mode with a range of smear analysis from 200 to 20,000 nt. Only samples with a total RNA quality number (RQN) (45) of > 4 and a DV₂₀₀ (percentage of RNA fragments over 200 nt in length) (46) of > 64.5% were chosen. For RNA quality classification, the Illumina® recommendations were applied, which denote a DV₂₀₀ of > 70% as high and a DV₂₀₀ of 50–70% as medium quality. For data on RNA quality, see Supplementary Table S2. Total RNA was treated with DNase I. Both the nCounter® and QuantSeq 3′ analyses were employed on the total RNA from the same isolation batches.

The RNA (150–250 ng) was hybridized to the nCounter® Canine IO Panel XT CodeSets (NanoString Technologies, Inc., Seattle, WA, United States), including probes representing 780 pre-selected genes and 20 “housekeeper” genes. A 30 probe Panel Plus (Supplementary Table S3) was added to the hybridization of the HGA and AGASAC samples following the manufacturer’s hybridization protocols (manual IDs: MAN-10023-11, MAN-10056-06). This Panel Plus included genes of further interest for this entity comparison. Hybridized samples were loaded onto the nCounter® MAX Analysis System’s Prep Station (NanoString) for purification and immobilization on sample cartridges, transferred to the Digital Analyzer for data collection and analyzed following the manufacturer’s user manual (manual ID: MAN-C0035-08).

Following the workflow described in the manufacturer’s recommendations (manual IDs: MAN-C0019-08, MAN-C0011-04), the reporter library files (RLF) and reporter code count (RCC) files were imported into the nSolver™ 4.0 Analysis Software (NanoString). Quality control and normalization followed default settings. Differential gene expression (DGE) analysis was implemented with the R 3.3.2-based Advanced Analysis 2.0 plug-in (version 2.0.134) with the recommended statistical settings. The raw, normalized, and DGE data were exported and read into R / Python Jupyter notebooks for further bioinformatic analysis and correlation calculations.

The RNA was sequenced with QuantSeq 3′ (Lexogen GmbH, Vienna, Austria) at Lexogen Services. DNase I treated total RNA (500–1,000 ng from HGA and AGASAC; 375 ng from the early and late CCH stages) was processed with the QuantSeq 3’ mRNA-Seq FWD Library Preparation Kit (Lexogen) according to the manufacturer’s guidelines (user guide: 015UG009V0251) using the low-quality RNA protocol. Quality of the libraries was determined with the Agilent 5300 Fragment Analyzer (DNF-474-33 - HS NGS Fragment 1-6000 bp method mode). The samples were pooled in equimolar ratios. The library pool was quantified using a Qubit™ dsDNA HS assay kit (Thermo Fisher) and sequenced utilizing an Illumina® NextSeq™ 500 system with a SR75 High Output Kit. In total, 0.011 sequencing lanes were used per sample, resulting in an output of 4,025,165 to 7,548,231 trimmed reads for the HGA and AGASAC samples with an average input read length of 63.41 to 71.6 nt and 5,426,266 to 7,783,141 trimmed reads for the CCH samples with an average input length of 66.21 to 71.65 nt.

The FASTQ sequencing files were first pre-processed (adapter trimmed, filtered) using Cutadapt (47) and subsequently aligned to the NCBI Reference Sequence (RefSeq) assembly for the dog (Canis lupus familiaris) CanFam3.1 (48) (GCF_000002285.3) with the STAR aligner (49). Read counts per gene were generated with two rounds of featureCounts (50) to mitigate the effect of too short 3′ UTR annotations. Reads that remained unassigned in the first round were subjected to the second round of featureCounts on an adjusted annotation with a 3′ extension by 2 kilobases. Multi-mapped reads were retained and all counts distributed evenly across all mapping locations. For comparison with nCounter® counts, the raw gene counts were normalized with the edgeR package (51–53). DGE analyses were performed with R using DESeq2 (54). The workflow is fully accessible in figshare (see data availability, DOI: 10.6084/m9.figshare.25768587). The significance thresholds were set at a log₂ (fold change) (log₂FC) of ≤ −1 or ≥ 1 and an adjusted p-value (p_adj) of ≤ 0.05.

As the nCounter® Canine IO Panel probes were designed using CamFam3.1, the same reference genome for QuantSeq 3′ read alignment was used for proper gene annotation comparison, despite the dog reference genome having been updated since. This same annotation offers a more robust background for comparison, as genes have been either dropped or added in the more recent genome.

Two biological comparisons were chosen to test whether data sets stemming from an intra- or an inter-tumor contrast might be more prone to errors, possibly reflected in a weaker correlation of results from different measurements. For the inter-tumor, entity-contrasting comparison, the HGA (n = 10) was compared to the AGASAC (n = 15 for gene correlation, n = 11 for DGE) as baseline (from now on “HGA versus AGASAC”) and for the intra-tumor, stage-dependent comparison, the early CCH stage (n = 5) was compared to the late CCH stage (n = 5) as baseline (from now on “CCH early versus late”).

For correlation calculations of gene abundance, nCounter® read counts were normalized with nSolver™, while QuantSeq 3′ read counts were normalized using three different methods based on the full set of known genes (Figure 1). nCounter® reads were normalized in nSolver™ in a two-step process with positive control and “housekeeping” gene normalization. At the level of QuantSeq 3′ data normalization, three different commonly applied normalization methods were compared. These were TMM (Trimmed Means of M-values); the method used by edgeR, CPM (counts per million), and RLE (relative log expression); the method used by DESeq2. All three methods were implemented within the edgeR package.

Different correlation coefficients were calculated for each the correlation of counts and the differentially expressed genes (DEGs) based on their log₂FC values. The Pearson correlation coefficient indicated the correlation of counts and Spearman’s rank correlation coefficient the correlation of gene rank. To mitigate the effect of outliers, the Pearson correlation was also calculated on log-transformed counts with a pseudocount of 1 (from now on “Pearson-log”) for the sample-wise correlation calculation.

Count correlations were considered on a sample-wise (with Pearson-log and Spearman) and gene-wise (with Pearson and Spearman) level. The sample-wise level evaluated the correlation of counts of all genes per matched FFPE sample in both analysis methods used. The gene-wise level viewed the correlation of counts of an individual gene across the analyzed sample pairs.

The interpretation of the correlation strengths was based on previous publications for application in the medical field (55, 56): ≥ 0.8 = very strong, ≥ 0.6 to < 0.8 = moderately strong, ≥ 0.3 to < 0.6 = fair, and < 0.3 = poor.

For the “HGA versus AGASAC” comparison, all 25 samples were used for the sample- and gene-wise count correlations, as a higher sample size generally improves data reliability. For DEG correlation, only the 21 primary tumor samples were included, thereby decreasing some biological variability in the tissue background within the sample cohort. All 5 CCH samples in either the early or late stage were used in both count and DEG correlations.

Different procedures for the calculation of the false discovery rate (FDR) used for p-value adjustment during DGE calculation are routinely used. The nSolver™ user manual recommends using the Benjamini-Yekutieli (BY) procedure (57). On the other hand, DESeq2 routinely employs the Benjamini-Hochberg (BH) procedure (58). Thus, to allow for better comparisons, the BY method was employed on all data shown here, as it is the more conservative and discriminating of the two (59–61).

To begin, the genes included on the nCounter® Canine IO Panel were identified which are not annotated in CanFam3.1 and thus could not be routinely found in the QuantSeq 3′ data after alignment to this genome assembly. The genes were considered “missing” or without overlap between the two methods. It was consequently impossible to calculate correlation coefficients for these genes.

In the “HGA versus AGASAC” comparison, out of 830 genes included on the nCounter® panel with Panel Plus, 821 genes (98.9%) were found in the QuantSeq 3′ data. The following nine genes (1.1%) were missing in the QuantSeq 3′ data: FCAR, HAVCR2, IFNA7, IGHG, IGHM, IL2, MIF, TRAC, and TRBC.

Similarly, in the “CCH early versus late” comparison, out of 800 genes included on the nCounter® panel, ten genes (1.25%) were missing in the QuantSeq 3′ data: CCR2, FCAR, IFNA7, IGHG, IGHM, MIF, TRAC, TRBC, TRGC2, and TRGC3. Thus, 790 genes (98.75%) genes overlapped.

To determine the correlations on a count level, counts of all genes per matched FFPE sample analyzed with both nCounter® and QuantSeq 3′ (sample-wise level) were investigated. Overall, the sample-wise count correlations exhibited a narrow distribution of correlations and all correlation medians on this level were very strong in both biological comparisons (> 0.83).

For the “HGA versus AGASAC” comparison, data obtained with TMM provided a very strong count correlation (Pearson-log median: 0.87). All normalization methods, however, similarly yielded very strong count correlations (Pearson-log medians: > 0.84). All gene rank correlations were identically very strong (Spearman medians: 0.86) among all normalization methods (Figure 2A).

In the “CCH early versus late” comparison, the utilization of TMM also provided a very strong count correlation (Pearson-log median: 0.85). The data from all normalization methods, however, likewise yielded very strong correlations (Pearson-log medians: > 0.83). All gene rank correlations were equally very strong (Spearman medians: 0.84) among all normalization methods (Figure 2B).

For the sample-wise count correlation calculations in both biological comparisons, the application of TMM and CPM (medians: 0.87 and 0.85, respectively) both slightly improved the Pearson-log correlations compared to the correlations with no normalization (medians: 0.86 and 0.84, respectively). However, the application of RLE slightly decreased the correlations (medians: 0.85 and 0.84, respectively). No improvement was seen after implementing the three normalization methods on the Spearman correlations; here all coefficients were identical (medians: 0.86 and 0.84, respectively).

In summary, all sample-wise count correlations were very strong with TMM, CPM, and RLE. Both the Pearson-log and Spearman correlations were slightly stronger in the entity-contrasting comparison (“HGA versus AGASAC”) than in the stage-dependent comparison (“CCH early versus late”).

Next, the correlations on a count level based on the counts of an individual gene across FFPE sample pairs analyzed with both nCounter® and QuantSeq 3′ (gene-wise level) were examined. Altogether, for the gene-wise count correlations, there was a much wider distribution of correlations.

Moderately strong (medians: > 0.63) correlations were calculated for the “HGA versus AGASAC” comparison (Figure 2C). This was observed on the level of counts (Pearson medians: > 0.67) and gene rank (Spearman medians: > 0.63). All normalization methods generated stronger correlations compared to no normalization (Pearson median: 0.67, Spearman median: 0.63). The application of RLE resulted in a slightly stronger correlation of counts (Pearson median: 0.732), compared to TMM (Pearson median: 0.726) and CPM (Pearson median: 0.731). With the use of TMM and CPM, almost identical gene rank correlations (Spearman median: 0.66) were calculated, compared to RLE (Spearman median: 0.65).

The “CCH early versus late” comparison also showed a much larger distribution of correlation values in all normalization methods (Figure 2D). This was observed on the level of counts (Pearson medians: > 0.76) and gene rank (Spearman medians: > 0.71). The application of all three normalization methods only slightly improved the correlations on count level (Pearson medians: > 0.77), compared to no normalization (Pearson median: 0.76). The application of CPM resulted in a slightly stronger correlation for counts (Pearson median: 0.79), compared to TMM and RLE (Pearson medians: 0.777 and 0.779, respectively).

Notably, both the Pearson and Spearman correlations were slightly stronger in the stage-dependent comparison (“CCH early versus late”) than in the entity-contrasting comparison (“HGA versus AGASAC”), contrary to the sample-wise results.

To identify the gene-wise correlations of gene counts depending on the respective gene expression level, the Pearson and Spearman correlation coefficients normalized with TMM were plotted against the mean gene expression data from either nCounter® or QuantSeq 3′ (Figure 3). Thresholds were drawn based on the gene distribution within the scatter plots. All gene counts above a respective threshold were grouped into an expression strength category and the correlation coefficients calculated for each of the respective genes. Specifically, the mean expression correlations for all, the top 90%, and 50% of genes were calculated. In both comparisons, the median correlation among all genes including the lowly expressed genes was weaker (“HGA versus AGASAC”: Pearson: 0.73, Spearman: 0.66/“CCH early versus late”: Pearson: 0.77, Spearman: 0.73) than among all genes excluding the lowly expressed genes (“HGA versus AGASAC”: Pearson: 0.77, Spearman: 0.69/“CCH early versus late”: Pearson: 0.8, Spearman: 0.75) and including only the highly expressed genes (“HGA versus AGASAC”: Pearson: 0.85, Spearman: 0.8/“CCH early versus late”: Pearson: 0.83, Spearman: 0.81). Summarizing, the correlation of genes including the highly expressed genes was strongest.

To ascertain the correlations of genes based on their differential gene expression in the two biological comparisons, it was first necessary to calculate the correlations of the log₂FC values from DESeq2 without log₂FC-shrinkage (log₂FC-shrinkage = shrinks estimated effect size toward zero, stronger for genes with little information, i.e., low average read counts that can likely yield artificially high log₂FC estimates) on the overlapping genes. For the “HGA versus AGASAC” comparison, both correlations were very strong (Pearson: 0.91, Spearman: 0.87) based only on the log₂FC values of the genes. For the “CCH early versus late” comparison, on the other hand, both correlations were only moderately strong (Pearson: 0.68, Spearman: 0.72).

In order to determine the overlap of differential gene expression direction, the log₂FC values from DESeq2 with shrinkage were collated into scatter plots. Therefore, the log₂FC of a given gene from the QuantSeq 3′ data was plotted against the log₂FC of the corresponding gene from the nCounter® data. The dots of the corresponding genes were colored according to the significance thresholds p_adj ≤ 0.05 and log₂FC ≤ −1 or ≥ 1 and the correspondence of the expression direction based on the log₂FC (≤ −1 = “down”; log₂FC ≥ 1 = “up”).

The scatter plot for the “HGA versus AGASAC” comparison showed a total of 599 genes (Figure 4A). One hundred and twenty genes (79 “down”; 41 “up”) had the same log₂FC direction (red dots). One hundred and twelve genes were predicted only in one of the methods: 62 in nCounter® (orange dots): 16 “down”; 46 “up” and 50 in QuantSeq 3′ (blue dots): 42 “down”; 8 “up.” No genes were classified in opposite directions. Three hundred and sixty-seven genes fell under the significance thresholds.

For the “CCH early versus late” comparison, a total of 658 genes were displayed in the scatter plot (Figure 4C). However, with a p-value threshold based on the BY adjustment, no genes were recognized by nCounter® and 89 genes (15 “down”; 74 “up”) only by QuantSeq 3′. The other 569 genes fell under the significance thresholds.

The lists of DEGs with their corresponding log₂FC values and p_adj from both methods calculated by DESeq2 with shrinkage (scatter plot groups) from both biological comparisons are provided in Supplementary Table S4.

Differential gene expression (DGE) analysis performed using the BY p-value adjustment and log₂FC with shrinkage calculated a total of 182 significantly differentially expressed genes (DEGs) for nCounter® and 170 significantly DEGs for QuantSeq 3′ in the “HGA versus AGASAC” comparison. To identify the overlap of the significantly DEGs in the “HGA versus AGASAC” comparison, a Venn diagram was created using the Matplotlib Python library (62). An overlap of 120 significantly DEGs was found in the “HGA versus AGASAC” comparison (Figure 4B), corresponding to a Jaccard index of 0.52.

In the “CCH early versus late” comparison, no significantly DEGs were calculated for nCounter® and a total of 89 significantly DEGs for QuantSeq 3′ were computed. As there were no significantly DEGs in the nCounter® data, the Venn diagram showed no overlap, matching a Jaccard index of 0 (Figure 4D).

Here, we present the first systematic comparison between data resulting from the technical platforms QuantSeq 3′ and nCounter® using the Canine IO Panel on canine archival FFPE tissue. Previous correlation studies had contrasted data from QuantSeq 3′ or nCounter® to other mRNA quantification methods utilizing FFPE tissue. These had, for example, used different human FFPE tissues to directly compare nCounter® gene selections to other RNA-Seq methods or RT-PCR. For instance, a comparison of immune gene expression in 27 clear cell renal cell carcinoma samples between the nCounter® Pan Cancer Immune Profiling Panel and the Oncomine™ Immune Response Research Assay on FFPE specimens had revealed a moderately strong correlation (Spearman: 0.73) of fold changes for 248 shared genes. On a gene-wise level, 226 of these genes had shown positive correlations and 16 had negative correlations. The mean Pearson correlation for all genes had been fair at 0.45 (range: 0.98 to −0.25) (63). Comparing 20 genes in oral carcinoma samples using custom nCounter® CodeSets and RT-PCR, a moderately strong overall correlation (Pearson: 0.59) had been calculated (22) for gene expression levels in 19 FFPE sample pairs. In contrast, the correlation results of our study on canine tissues were stronger on both count and log₂FC levels. This could be due to the improved compatibility of nCounter® and QuantSeq 3′ for FFPE material compared to the Oncomine™ assay and RT-PCR that had been employed in the previous studies. Considering RNA quality, the DV₂₀₀ values of the total RNA in the samples chosen for this study on canine tissues were predominantly of high quality (n = 31 with DV₂₀₀ > 70%) with some samples with medium quality (n = 4 with DV₂₀₀ 50–70%). One previous study (63) had not provided RNA quality data. The other study (22) had employed only the RNA integrity number (RIN) and disclosed the use of strongly degraded RNA (mean RIN: 2.3, range 1.5–2.5). However, the DV₂₀₀ has been shown to be a superior criterion to evaluate RNA quality compared to the widely used RIN, especially when employing partially degraded RNA from FFPE samples (46). Thus, somewhat superior RNA quality may have contributed to the slightly better correlations in our comparison.

Other studies performed on cancer cell lines (18) or peripheral blood (64) had shown that the correlations on count level are dependent on the gene expression level. Generally, weaker correlations had been observed for genes with low expression levels and stronger correlations had been calculated for genes with high expression levels. This is in line with our study’s results: both nCounter® and QuantSeq 3′ exhibited the same expression-level dependent correlation phenomenon, with stronger correlations for higher expressed genes.

The strong correlations observed in this study raise much hope for retrospective studies using samples from veterinary pathology archives, as routinely stored FFPE tissues seem principally accessible for transcriptome analyses using both nCounter® and QuantSeq 3′. In comparison to FFPE material, fresh, or fresh-frozen tissue is much more difficult to obtain in veterinary and comparative pathology for logistical, ethical, potentially infectious, and legal reasons. Additionally, the cost of RNA-Seq has greatly decreased since its first introduction and in combination with the easy storage of FFPE material provides a cost-effective research and diagnostic tool for veterinary researchers and clinicians. Especially in favor of Russell’s and Burch’s 3 R (reduce, refine, replace) principles to improve ethical concerns when using experimental animals, it is imperative to obtain the highest possible benefit out of animal tissues. Obviously, routine archiving of FFPE tissues for many decades has proven to be a fortunate circumstance.

Confounding factors when using FFPE material, however, include the type, buffering state, and concentration of formalin used, duration of fixation, tissue processing, and storage conditions, which all may impact RNA quality (1, 6, 65–68). In the case of this study, in which all tissues stemmed from routinely fixed, processed, and stored samples as part of the institutional diagnostic service, such limitations seemed to have had little, if any impact, on the results based on the RNA quality (Supplementary Table S2).

Still, multiple other aspects likely have an influence on correlations of data from different transcriptome analysis methodologies and might also account for the differences in DEGs detected in both methods in this study. To begin, the differing detection methods may result in slight discrepancies in the identification of transcripts and bioinformatic alignment of sequences after RNA-Seq may not be as specific as the probes used in nCounter®. Furthermore, by focusing on a selection of approximately 800 genes, the nCounter® panel offers the possibility to detect even weakly expressed transcripts, which may be overlooked compared to highly expressed genes, depending on the depth of sequencing in the case of QuantSeq 3′. However, pre-selected gene panels have some inherent limitations due to their focused design compared to genome-wide approaches, such as RNA-Seq. As expression analysis is restricted to specific genes of interest, changes in other genes not included on the panel remain undetected. Thus, potential novel biomarkers, broader biological processes, or unforeseen gene interactions can remain unidentified (69). Incomplete interpretation of underlying mechanisms or the oversight of key regulatory genes involved in a disease phenotype can result. Rather, panels are intended for focused research questions or validation of known targets. The choice of genes selected for inclusion is dependent on current knowledge and thus introduces a bias, restricting the scope of discovery. The gene set enrichment analysis tools available for RNA-Seq are of little statistical value for panels, as the background gene set is limited and biased by the panel’s focus. Thus, pre-selection violates the statistical assumption of unbiased, comprehensive, and genome-wide gene expression data (70).

Furthermore, different normalization methods may have an effect. This study employed the nSolver™-embedded normalization method for the nCounter® data and compared three common normalization methods (TMM, CPM, RLE) for the QuantSeq 3′ data. All methods performed similarly, with only minimal differences that seemingly did not influence the strength of correlations. Extending beyond these normalization methods, there is however a multitude of other approaches, which were developed to correct biases or weaknesses, such as accounting for gene length, refinement for FFPE RNA-Seq data, or microarray input (71, 72). For nCounter® data, further methods have likewise been developed for normalization and differential expression analysis apart from nSolver™. As the three normalization methods employed in this study provided little differences in correlations, it was not deemed necessary to test the impact of alternate normalization strategies on count correlations. Instead, it seems that DESeq2 and edgeR can reliably be used for count normalizations on QuantSeq 3′ data and that the software package one uses for normalization will likely have very little impact on the data. Similarly, other packages or approaches for differential gene expression (DGE) analysis are available (73). Again, in this study, the impact that these alternate DGE calculations may have on DEG correlations was not investigated. Overall, the analysis of DEGs is based on different approaches, i.e., statistical calculations, available software, and pathway lists for enrichment analyses, when using transcriptomic platforms. The nCounter® reads are usually fed into the nSolver™ software with the Advanced Analysis plug-in and pathway analyses are conducted with the program’s inbuilt modules. Annotations are assigned to most genes, which are provided in annotation files. The annotations for the Canine IO Panel are based on two previously developed human panels: the immune response categories originate from the PanCancer Immune Profiling Panel and the functional annotations are based on the IO 360 Panel. These annotations in turn stem from the Gene Ontology (GO) (74, 75), Kyoto Encyclopedia of Genes and Genomes (KEGG) (76), and Reactome (77, 78) databases. If an investigator has another research focus in mind, a gene curation team is able to annotate genes included on a panel with suitable terms. A custom annotation route, using an annotation engine, assigns annotations to customized gene add-ons (Panel Plus) or CodeSets directly from a database based on context (NanoString curation team, personal communication). Additionally, based on a specific research query, each investigator has the ability to manually change and adjust the annotations in an annotation file. This can lead to fluctuations and different statistical outcomes and should be transparently declared in consequent publications. With a multitude of gene annotation databases available for pathway analysis, such as GO, KEGG, Reactome, and Molecular Signatures Database (MSigDB) (79, 80), the nCounter® annotations add further nomenclature for gene grouping categories and pathways, which are not always transferable and/or interchangeable when comparing terms. However, researches may be able to answer questions more target-oriented this way.

Noteworthy, in the stage-dependent comparison conducted in this study, 89 DEGs were detected using QuantSeq 3′, but none using nCounter®. As previously outlined, it is generally to be expected that the two methods will yield divergent results, partly due to the differing scope of potentially detectable transcripts (whole transcriptome versus focus on 800 genes) and partly, to a lesser extent, due to different underlying detection methods used. Still, when comparing the stages of CCH, it must be emphasized that there are only minor differences at the transcriptome level (36). This is also reflected by the low number of detected DEGs using QuantSeq 3′ in the stage-dependent contrary to the entity-contrasting comparison in this study. However, studies that specifically had examined individual transcripts utilizing quantitative PCR found differences between early and late stages of CCH, for example in expression of proinflammatory cytokines (34). Thus, further investigation is required in order to resolve this discrepancy. Larger group sizes than those used in this study (n = 5) might be necessary to detect a larger number of significantly DEGs. Additionally, to the presumably minor differences at transcriptome level, the number of DEGs is likely to be reduced by the implementation of the BY procedure, which has been demonstrated to be more conservative than other FDR calculation methods (57–61).

Major differences also lie in data processing. As with all large-scale transcriptome analysis systems, vast amounts of raw data are generated requiring further bioinformatic analysis. The QuantSeq 3′ system provides nucleotide sequences in FASTQ data files for alignment and normalization using standard bioinformatic methods. Prior bioinformatic know-how is thus indispensable. In comparison, the nCounter® system has the advantage of the specially developed nSolver™ software. The generated raw data can be imported into this software, normalized and analyzed by the user without any previous bioinformatic expertise.

This technical study was based on two comparisons between contrasting tumor entities, the first between early and late stages of the spontaneously regressing canine cutaneous histiocytoma (CCH) and the second between two common canine perianal tumors with very distinct biological behaviors. The widely overlapping data obtained from the two methodological approaches clearly allowed for several oncologically interesting implications, with mutual confirmations by the two methods used. These results and interpretations, however, are subject of separate publications already published (36) or in preparation.