Prognostic Molecular Classification of Appendiceal Mucinous Neoplasms Treated with Cytoreductive Surgery and Hyperthermic Intraperitoneal Chemotherapy

Abstract

Background.

Appendiceal mucinous neoplasm (AMN) with peritoneal metastasis is a rare but deadly disease with few prognostic or therapy-predictive biomarkers to guide treatment decisions. Here, we investigated the prognostic and biological attributes of gene expression-based AMN molecular subtypes.

Methods.

AMN specimens (n = 138) derived from a population-based subseries of patients treated at our institution with cytoreductive surgery and hyperthermic intraperitoneal chemotherapy (CRS/HIPEC) between 05/2000 and 05/2013 were analyzed for gene expression using a custom-designed NanoString 148-gene panel. Signed non-negative matrix factorization (sNMF) was used to define a gene signature capable of delineating robustly-classified AMN molecular subtypes. The sNMF class assignments were evaluated by topology learning, reverse-graph embedding and cross-cohort performance analysis.

Results.

Three molecular subtypes of AMN were discerned by the expression patterns of 17 genes with roles in cancer progression or anti-tumor immunity. Tumor subtype assignments were confirmed by topology learning. AMN subtypes were termed immune-enriched (IE), oncogeneenriched (OE) and mixed (M) as evidenced by their gene expression patterns, and exhibited significantly different post-treatment survival outcomes. Genes with specialized immune functions, including markers of T-cells, natural killer cells, B-cells, and cytolytic activity showed increased expression in the low-risk IE subtype, while genes implicated in the promotion of cancer growth and progression were more highly expressed in the high-risk OE subtype. In multivariate analysis, the subtypes demonstrated independent prediction power for post-treatment survival.

Conclusions.

Our findings suggest a greater role for the immune system in AMN than previously recognized. AMN subtypes may have clinical utility for predicting CRS/ HIPEC treatment outcomes.

While rare, appendiceal cancer is associated with considerable mortality due to the late stage at diagnosis, and the remote likelihood of being found by screening colonoscopic examinations.¹ Although the annual incidence of AMN in the United States is increasing,² it is estimated to range from only 450³ up to 5000² cases, which limits population studies. AMN all too frequently presents with peritoneal dissemination. Cytoreductive surgery (CRS) represents the primary therapeutic modality and, when found with peritoneal surface disease, is frequently combined with hyperthermic intraperitoneal chemotherapy (HIPEC).^4,5 Despite uniform treatment with CRS/HIPEC, patients demonstrate widely disparate prognoses, even after adjustment for known clinical and demographical factors.^6–11 This suggests that, as in other cancers, AMN heterogeneity at the molecular level may explain differences in tumor aggressiveness and response to CRS/HIPEC treatment.

There is a paucity of data on the molecular genetics of AMN. While this clearly relates to the rarity of AMN specimens for study, another contributing factor is the low cellularity and high mucin content characteristic of AMN, which presents technical challenges associated with obtaining samples with high cell purity and adequate nucleic acid integrity. By optimizing protocols for the latter, we recently enabled expression profiling analysis of appendiceal adenocarcinoma using oligonucleotide microarray technologies.^9,10 In that work, we identified a number of statistically significant prognostic gene-survival associations and discovered two potential molecular subtypes of AMN that could be discriminated by a 139-gene cassette. Moreover, the subtypes exhibited significantly different survival rates independent of known clinical and demographical risk factors. In the current study, we leveraged the NanoString nCounter^® system, a highly sensitive diagnostics platform for accurate RNA quantitation, to confirm the existence of prognostic molecular subtypes in an independent population-based cohort of 138 AMN cases treated by CRS/HIPEC at the Wake Forest Baptist Comprehensive Cancer Center. Thus, we extend our previous findings toward the establishment of a molecular prognostic assay to support treatment decisions for AMN patients.

MATERIALS AND METHODS

Patients and Tumor Specimens

All tissues and clinical data utilized in this study were obtained for research use by patient consent under an institutional review board-approved protocol at Wake Forest Baptist Medical Center, Winston Salem, NC. Selection of AMN cases was facilitated by a prospectively maintained database of clinical and demographic information on >600 appendiceal cancer patients treated with CRS/HIPEC at Wake Forest Baptist Medical Center (WFBMC) from 1992 to present. The technique for CRS/HIPEC was performed in standard fashion that has been described in detail elsewhere.^9,10,12 All AMN tumor tissue specimens acquired at the time of surgery were grossly identified under the direction of a pathologist, processed as formalin-fixed paraffin-embedded (FFPE) tumor blocks and as snap frozen specimens, and maintained in a tumor tissue repository housed within our Cancer Center’s Tumor Tissue and Pathology Shared Resource. In the current study, we identified a population-based sub-series of 246 CRS/HIPEC-treated patients dating from the 13-year period May 2000 to May 2013 and annotated for disease progression or death with >0.5 year follow-up. Tissue processing and analysis were conducted by our Tumor Tissue and Pathology Shared Resource as described previously.¹⁰ Cellular content of tumor specimens was verified by a surgical pathologist. Localized regions of tumor-positive specimens with greatest cancer cellularity were identified and marked for punch biopsy. A minimum threshold of 10% cancer cellularity was observed for qualifying samples. The same procedure was followed for both low- and high-grade tumors. Not all available tumor specimens yielded adequate cellularity, and 51 cases were excluded on this basis.

AMN specimens were annotated for histological grade (high grade versus low grade), following the Bradley system.¹¹ Tumors with any of the following histological characteristics were classified as high grade: high-grade nuclear atypia to include prominent nucleoli, hyperchromatic nuclei, markedly irregular nuclear membrane or dense coarse chromatin; architectural complexity, such as cribriform pattern, papillary formation, or extensive nuclear stratification; high cellularity; frequent mitosis; and signet ring cells.

NanoString n-Counter Profiling

Frozen or FFPE-processed tumor samples corresponding to 195 AMN cases were homogenized using a Qiagen QiaShredder and total RNA was purified using Qiagen RNeasy Mini and RNeasy FFPE kits in our Cancer Genomics Shared Resource. Total RNA quality requirements for expression profiling on the NanoString platform included a minimum yield of 150 ng total RNA (quantified using an Eppendorf BioPhotometer) with ≥ 50% of the total mass corresponding to a length of 300 base pairs or greater as measured on an Agilent 2100 Bioanalyzer. Of note, traditional RIN value assessment is typically not applicable to FFPE-preserved tissues as a consequence of prevalent RNA degradation; thus it is not considered ideal for RNA quality control in NanoString n-Counter workflows. Among tumor samples with qualifying cancer cellularity (n = 195), 37 cases were excluded from further analysis owing to suboptimal RNA quality. Thus, 158 AMN specimens were carried forward for analysis on the n-Counter platform. The custom n-Counter code sets targeted 143 genes that, by microarray analysis, showed the most significant associations with patient overall and/or progression free survival (p < 0.001, Cox regression analysis) in our previously published studies.⁹ The code sets also targeted 5 housekeeping genes, 6 positive controls and 8 negative controls. The code sets were designed by NanoString to target each gene’s corresponding Affymetrix probe set “target sequence” for consistency with the discovery process described in our previous reports.⁹ RNA samples were profiled according to standard NanoString protocols on a NanoString n-Counter Analysis System operated by the Preclinical Genomic Pathology Laboratory at the University of North Carolina at Chapel Hill Line-berger Comprehensive Cancer Center. The resulting data were analyzed by our Cancer Genomics Shared Resource using the NanoString n-Solver software and normalized according to the geometric mean of the 5 housekeeping genes (ACTB, EEF1A1, GAPDH, RPL37A, and RPLP1). After normalization, samples with a content normalization factor > 6.0 (n = 20) were deemed potentially unreliable owing to poor NanoString detection, and excluded from further analysis. In total, tumor expression profiles from 138 patients (56% of the initial AMN population) were carried forward for further analysis.

Cohort Characteristics

Clinical and demographic characteristics of the 138 cases are described in Table 1. Most characteristics were proportionately similar between the low-grade and high-grade AMN cases. However, as expected, the prognosis after CRS/HIPEC treatment differed markedly between low and high grade. High-grade cases exhibited shorter intervals to death and disease progression. Preoperative chemotherapy was administered to 64 patients (46%) of the cohort; for one patient, the use of preoperative chemotherapy could not be confirmed.

TABLE 1.

Patient characteristics

	Total (IM = 138)	Low grade (N = 76)	High grade (N = 38)
Demographics
Age–year (onset*), Mean (SD)	52.0 (12.1)	53.1(13.2)	50.9 (9.0)
Male sex-no. (%)	53 (43.4%)	30 (42.9%)	15 (44.1%)
Female sex-no. (%)	69 (56.6%)	40 (57.1%)	19 (55.9%)
Race
African-no. (%)	15 (10.9%)	13 (17.1%)	1 (2.63%)
Caucasion-no. (%)	107 (77.5%)	54 (71.1%)	34 (89.5%)
Asian-no. (%)	2 (1.45%)	2 (2.63%)	0
Others-no. (%)	2 (1.45%)	2 (2.63%)	0
Ethnicity
Hispanic or Lartino-no. {%)	3 (2.17%)	3 (3.95%)	0
Others-no. (%)	136 (97.83%)	73 (96.05%)	48 (100%)
Overall Survival: no.	135	76	38
Events-no. (%)	65 (48.1%)	23 (30.3%)	32 (84.2%)
Followup (yr), Mean (SD)	2.95 (2.14)	3.87(2.21)	1.48 (1.01)
Time of event (yr), Mean (SD)	1.96 (1.68)	2.91 (2.21)	1.43 (1.00)
Progression-free survival: no.	80	53	19
Events-no. (%)	54 (67.5%)	33 (62.3%)	18 (94.7%)
Followup (yr), Mean (SD)	2.51 (2.16)	2.96(2.30)	1.03 (0.713)
Time of event (yr), mean (SD)	1.36(0.997)	1.63(1.13)	0.907 (0.497)
Surgical score R	126	71	34
R0/R1-no. (%)	50 (39.7%)	26 (36.6%)	11 (32.3%)
R2-no. (%)	76 (60.3%)	45 (63.4%)	23 (67.6%)
ECOG performance status	122	70	33
Grade 0-no. (%)	64 (52.5%)	33 (47.1%)	16 (48.5%)
Grade 1-no. (%)	46 (37.7%)	31 (44.3%)	11 (33.3%)
Grade 2-no. (%)	10 (8.20%)	5 (7.14%)	5 (15.2%)
Grade 3-no. (%)	2 (1.64%)	1 (1.43%)	1 (3.03%)
Preoperative chemotherapy-no. (%)	64 (46.4%)	21 (27.6%)	34 (89.5%)
Adjuvant chemotherapy-no. (%)	39 (28.2%)	18 (23.7%)	16 (42.1%)

Tumor Subtyping

Transcriptional heterogeneity among patients was explored using signed non-negative matrix factorization (sNMF)—an algorithm we previously developed ¹³ to more reliably discern subpopulations defined by differential gene expression. Briefly, the data are input as a z-normalized gene expression matrix A with genes as rows and samples as columns. Two non-negative matrices, A⁺ and A⁻, are derived from the original input as:

$$A_{ij}^{+}=\{\begin{array}{ll}{A}_{ij},\hfill & \text{if}{A}_{ij}>0\hfill \\ 0,\hfill & \text{otherwise}\hfill \end{array}\text{and}{A}_{ij}^{-}=\{\begin{array}{c}-A_{ij},\text{if}{A}_{ij}<0\\ 0,\text{otherwise}\end{array}$$

where row i indicates the i’th gene and column j indicates the j’th sample. Instead of approximating A~WH, the non-negative $\left(\begin{array}{l}{A}^{+}\hfill \\ A^{-}\hfill \end{array}\right)~\left(\begin{array}{l}{W}^{+}\hfill \\ W^{-}\hfill \end{array}\right)H$ is solved, with H representing the membership of samples to corresponding subtypes, and W⁺ and W⁻ representing the membership of upregulated and downregulated genes in these subtypes, respectively. The importance of each gene is further evaluated by its overall contribution to all subtypes using a summarized membership matrix W^′ in which $W_{ij}^{\prime }=W_{ij}^{+}+W_{ij}^{-}.$ After summarization, genes that are significantly upregulated in a subtype and significantly downregulated in another subtype are more likely to have a higher weight in W^′ than those that contribute to only one subtype, and thus are considered preferential genes. Before subtyping, genes were first screened according to their relevance in treatment outcomes. The significance of each gene for predicting survival endpoints after CRS/HIPEC treatment was examined using univariate Cox proportional hazards regression. False discovery rates (FDRs) ¹⁴ were calculated to adjust for multiple testing. Among the 143 endogenous genes of interest, 22 were significantly associated with survival (FDR\0.05, Supplementary Fig. 1) and thus were used as initial candidate genes for patient subtyping. Tumor samples were then subjected to unsupervised bi-clustering using sNMF and the expression data from the 22 genes. The sNMF algorithm identified 17 genes with high weight in W^′ and three AMN subtypes that were well differentiated by the gene expression profiles. Details are listed in Table 2.

TABLE 2.

Multivariate cox regression for overall survival analysis

	Hazard ratio	95% Confidence interval	P
Grade	6.6	2.8, 15	1.2e-05
ECOG score	1.4	0.89, 2.3	0.14
R score	2.7	1.3, 5.7	0.009
Adjuvant chemo	0.68	0.34, 1.3	0.27
Preoperative chemo	1.4	0.63, 3.0	0.42
Age	0.99	0.97, 1.0	0.64
Sex (M)	0.89	0.48, 1.7	0.72
OE subtype	2.5	1.0, 6.2	0.044
M subtype	1.9	0.77, 4.6	0.177

n = 98, number of events = 46 Likelihood ratio test: p = 2e-9

Bold texts indicate statistically significant risk factors with p-values less than 0.05

Topology Analysis

The topology of patient sample heterogeneity was further explored using manifold learning and reverse graph embedding.^15–17 The transcriptomic data of all cancer-related or housekeeping genes over 138 patients were used for principal graph learning. A latent-graph-preserved dimension reduction was performed using DDRTree¹⁵ and the topological heterogeneity among samples was constructed as a principal graph underlying the transcriptomic data. Samples were clustered according to the branches of the principal graph they were assigned to. Monocle 2¹⁷ was used for AMN sample topology analysis.

Survival Analysis

Both overall survival and progression-free survival were analyzed for clinicopathological risk factors as well as the discovered molecular subtypes. Kaplan–Meier estimator and log-rank tests were used for non-parametric survival analysis. Univariate and multivariate Cox proportional hazard models were further used to evaluate the statistical and clinical significance of candidate risk factors, including the molecular subtypes, cancer grade, residual tumor score (R), ECOG score (Eastern Cooperative Oncology Group Scale of Performance Status), and patients’ age and sex.

Robustness Analysis of the 17-Gene Signature

We examined the robustness of the 17-gene signature and the prognostic attributes of the AMN subtypes in an independent patient population of low- and high-grade AMN (n = 40) profiled on the Affymetrix U133A Gene-Chip system, which was previously published by our team.⁹ Briefly, the expression of the 17 signature genes discovered in this work were scaled and used to objectively classify patients into the corresponding AMN subtypes using the sNMF approach with the same parameter settings. The overall survival rates of patients partitioned into the AMN subtypes were compared to examine the prognostic power of the 17-gene signature.

RESULTS

AMN tumor gene expression profiles were generated on the NanoString gene panel and analyzed by the sNMF algorithm. A 17-gene signature, composed of 7 immune genes and 10 cancer-associated genes, was identified that could robustly classify the 138 patients into three distinct molecular subtypes, termed immune-enriched (IE, n = 40), oncogene-enriched (OE, n = 43), and mixed (M, n = 55) (Fig. 1). These subtypes showed significantly different gene expression patterns and post-treatment survival outcomes. Genes with specialized functions in lymphocyte physiology, including markers of T-cells (TRA, TRBC1) natural killer cells (KLRF1, KLRG1) and cytolytic activity (PRF1), exhibited elevated expression levels in association with the more favorable outcomes of IE, consistent with heightened effector cell infiltration and protective anti-tumor function in these tumors. By contrast, genes implicated in the promotion of cancer growth and progression (EPCAM, SPINK1, ESRP1, CLDN3, CLDN4, ELF3, GPX2) were more highly expressed in the poor outcome-associated OE group, consistent with a more aggressive tumor growth phenotype.

FIG. 1.

AMN molecular subtypes defined by sNMF. Three identified AMN molecular subtypes (immune enriched, oncogene enriched, and mixed) and 17 associated signature genes (7 immune genes and 10 oncogenes) displayed with clinical features (grade, ECOG score, and R score) are shown. Top: heatmap of the normalized gene expression patterns (rows) corresponding to tumor samples comprising the three molecular subtypes (columns). Bottom: patient clinical features (grade, ECOG score, and R score) corresponding to tumor samples in heatmap columns are shown

The inter-tumor heterogeneity in terms of gene expression patterns among AMN samples were further explored using reversed graph embedding. As demonstrated in Fig. 2, in a typical gene expression topology analysis, tumor heterogeneity was visualized from three perspectives: (1) local similarity: where transcriptionally similar samples were visualized as data points positioned close to each other; (2) global heterogeneity: where gradual changes in gene expression patterns across the whole cohort were visualized as trend lines (black lines) across data points (samples); and (3) hierarchical clusters: where major clusters were represented as data points around long branches, with minor clusters around short branches.

FIG. 2.

Gene expression topology graphs of AMN patients. Transcriptomic similarities among AMN samples are shown represented by circles (AMN samples), distance (similarity of samples) and trend line trajectories (global heterogeneity among samples). Trend line segments represent distinct hierarchical clusters. a Patient samples (circles) are colored according to topology class assignments (left panel) or sNMF molecular subtype class assignments (right panel). b The changes of relative expression levels of selected genes (y-axis) along the trajectory, from OE toward IE (x-axis). EPCAM (OE gene member), left panel. PRF1 (IE gene member), right panel

This analysis (Fig. 2a, left) uncovered three dominating topological classes (represented by data points around the three long black lines) and two minor classes (the two short black lines). Among them, two (data points colored in orange and purple) represented the high-risk OE subtypes, one (gray data points) for the intermediate-risk M subtype, and two (green and blue data points) for the low-risk IE subtype. To examine whether the discovered molecular subtypes were consistent with the sNMF-based subtyping results, we labeled the sNMF-based subtyping results onto the learned gene expression topology graph (Fig. 2a, right). The results of the two subtyping approaches were highly consistent (Fig. 2a). The Jaccard similarities between the IE, OE, and M subtypes and the corresponding topological clusters were 0.85, 0.84, and 0.73, respectively.

We further examined whether the gene expression patterns in the three molecular subtypes discovered by the sNMF approach were consistent with the gene expression topology analysis. We used EPCAM and PRF1 as examples. In Fig. 2b, points represent the normalized gene expression levels in the corresponding samples along the trajectory shown in Fig. 2a, starting from the OE end toward the IE end. Moving along the branches from the high-risk OE cluster (left branch) toward the low-risk IE cluster, the expression of oncogenes (e.g., EPCAM) declined, while the immune activities (e.g., PRF1) increased.

These findings from the transcriptomic topology analysis indicated that the discovered subtypes were reproducible by different subtyping methods, that they represented the majority of the heterogeneity of the AMN population, and that the discovered gene expression patterns were consistent across the topological graph.

The overall survival and progression-free survival of the discovered subtypes (Fig. 3) differed significantly (logrank tests with p ≤ 1.3e-06 and p ≤ 2.0e-04, respectively). Three-year overall survival rates were 83% (95% CI 71.5%–96.5%), 55% (43%–72%), and 36% (24%–55%), and 3-year progression survival rates were 65.5% (51%–84%), 30% (17%–53%), and 14% (5.0%–41%) for IE, M, and OE subtypes, respectively. To test the robustness of the 17-gene signature for delineating prognostic AMN subtypes, we objectively applied the sNMF algorithm comprising the 17 genes to our previously reported 40-patient AMN cohort ⁹ profiled on the Affymetrix GeneChip platform. Three AMN subtypes were again discerned by the signature, and the differential survival rates of the resulting AMN subtypes (p ≤ 2e-05) closely mirrored those observed for the 138-patient cohort, thus demonstrating a robust performance of the signature across patient populations and RNA quantitation platforms.

FIG. 3.

Analysis of survival differences between AMN subtypes. The overall (left) and progression-free (middle) survival curves of the three identified subtypes (IE: immune enriched, OE: oncogene enriched, M: mixed) in the reported cohort as well as the overall survival curves (right) of a 40-patient AMN cohort profiled on the Affymetrix GeneChip platform were estimated by Kaplan–Meier estimator. Statistical metrics about these survival profiles including sample size, number of events, and median survival time with 95% confidence interval are summarized in the corresponding Tables beneath survival plots. Log-rank tests were used to evaluate the survival differences between AMN subtypes

The molecular subtypes demonstrated independent prediction power for post-treatment survival (Table 2). Using multivariate Cox regression, we adjusted the prediction power of the subtypes by removing effects of other clinical and demographical factors including cancer grade (low and high), performance status (ECOG score), residual tumor score (R score, grouped as R0/R1 and R2), preoperative chemotherapy, adjuvant chemotherapy and patients’ age and sex. The molecular subtypes remained a significant risk factor in the model (p ≤ 0.044, hazard ratio: 2.5) together with cancer grades (p ≤ 1.2e-05, hazard ratio: 6.6) and R scores (p ≤ 0.009, hazard ratio: 2.7).

DISCUSSION

Accurate patient prognosis plays an essential role in precision oncology.^18,19 In AMN, the predominant prognostic variables associated with CRS/HIPEC outcomes are histologic grade, completeness of cytoreduction (R score) and patient performance status (ECOG score). In this study, we confirm that molecular subtypes of AMN, defined by gene signatures reflective of effector immune cell infiltration and tumor oncogenic properties, carry significant additive prognostic value not provided by conventional prognostic markers. In our previous expression profiling study ⁹ which utilized Affymetrix oligonucleotide arrays, unbiased hierarchical clustering analysis initially delineated three AMN molecular subtypes based largely on the expression characteristics of the oncogene-enriched genes. The highly expressed subtype was associated with poor survival outcomes, while the subtypes with intermediate and low expression were associated with better survival outcomes. Further clustering of the tumors based on a 139-gene cassette yielded only two predominant subtypes that exhibited significant survival differences, with the intermediate tumors dispersed among the two subtypes. The re-assignment of the intermediate-expressing subtype to the two predominant subtypes owed to the nature of the clustering algorithm and the relatively small sample size investigated in that study. In the current study, however, the sNMF algorithm also identified a distinguishable intermediate-expressing subtype [i.e., equating with the mixed (M) subtype] with a survival rate intermediate to that of the OE and IE subtypes, confirming in this larger study the existence of a clinically relevant mixed expression subtype. Consistent with our previous findings, multivariate Cox regression analysis showed an independent relationship between the subtype classes and overall survival. The high-risk OE subtype exhibited a hazard ratio that was exceeded only by histologic grade, highlighting its clinical relevance as a significant prognostic biomarker in this setting. Furthermore, the prognostic attributes of the AMN molecular subtypes discerned by the 17-gene signature were found to be highly reproducible in our previously reported 40-patient AMN cohort. This indicates a robust level of portability of the signature across patient populations and RNA quantitation platforms (NanoString n-Counter and Affymetrix GeneChip microarray). Thus, our findings confirm the reproducible and independent significance of gene-survival associations in AMN.^9,10

In this work, we identified an optimal set of 17 genes for partitioning AMN cases into IE, OE, and M prognostic subtypes. The genes comprising our model suggest that AMN is cellularly heterogeneous, with a complex immune, stromal and cancer cell composition that underlies patient outcomes. The association between the low-risk IE subtype and the heightened expression of genes with specialized roles in effector immune cell function suggests a role for immune-mediated control of AMN progression in some patients. In these tumors, further investigation of the functional and transcriptional dynamics of the AMN tumor microenvironment would more fully elucidate the cellular complexity of AMN underlying its pathogenesis, progression and tumor-immune interactions.

Cohort size is always a critical factor in studies that investigate tumor heterogeneity, but even more so in rare diseases such as AMN. Our study cohort (n = 138) is the largest to date for AMN genomics profiling with clinical follow-up and has enabled the discovery of reliable molecular subtypes of AMN. The consistency observed between two independent clustering approaches, as well as the distinct prognostic survival differences observed, suggests that the major AMN subtypes definable by gene expression patterns were discovered in this analysis.

One limitation of this study was the relatively small sample size associated with high-grade AMN (n = 38). As demonstrated in Supplementary Fig. 2, the three molecular subtypes exhibited distinct survival profiles among the patients with low-grade disease (n = 76). However, in the context of high-grade disease, the survival function of the IE subtype (n = 6) could not be sufficiently established. Further analysis of subtypes in a larger population of high-grade AMN is warranted.

As in other tumor systems, the risk-related molecular subtypes described here hold promise for application in clinical decision support. To be implemented clinically, prognostic gene expression assays must be both reliable and facile. This takes into account consideration for both the RNA assay platform and the reliability of the sample processing parameters. The nCounter^® Analysis System requires as little as 25 ng of total RNA, requires less than 15 min hands-on operation, and generates prognostic outcomes within 24 h. Our findings support the feasibility of using the nCounter^® platform for the prognostic molecular subtyping of AMN patients. AMN surgical tissues are cellularly heterogeneous, marked by frequently low or dispersed tumor cellularity and diverse cellular and acellular stromal features. How this impacts the reliability of AMN molecular subtyping is a question of optimal sample processing methodology that warrants further investigation. Strategies to incorporate AMN subtyping into the clinical management of AMN is the focus of ongoing work and will require multi-institutional validation studies.

Whether the OE, IE and M subtypes benefit from different therapeutic approaches is an attractive hypothesis. Clearly, our current findings cannot address this issue; however, the OE subtype seems a good candidate for preoperative chemotherapy trials and the IE for preoperative immunotherapy studies.