Stromal-Based Signatures for the Classification of Gastric Cancer

Gastric cancer is the second leading cause of cancer-related death worldwide, with the highest prevalence in Asia (1, 2). Therapy for metastatic gastric cancer includes a combination of chemotherapies and targeted therapies for Her2 (ERBB2) and VEGFR2 (KDR) as monoclonal antibodies. The prevalence of Her2 overexpression is only ∼15% to 20%; thus, the majority of patients are dependent on chemotherapy and antiangiogenesis. Although cancer immunotherapy is in the early phases of development for gastric cancer, there are trials under way to explore this modality both as monotherapy and in combination with the anti-VEGFR2 antibody ramucirumab (www.clinicaltrials.gov and refs.3 and4). Both stromal therapies, antiangiogenesis and immunotherapy, are in need of reliable means to identify those patients that will receive optimal therapeutic benefit, and to spare unnecessary side effects for those where the chance of clinical benefit is low.

One of the goals of the human genome project is to characterize the genetic and epigenetic variation of cancers, enabling their use to generate new therapeutic hypotheses. In gastric cancer, there have been several large efforts published in recent years. The two largest come from The Cancer Genome Atlas (TCGA; ref. 5) and Asian Cancer Research Group (ACRG; ref. 6). Both have published RNA profile analyses that divide patients into 4 subgroups. In the present study, we generated stromal-specific RNA and IHC-based signatures that represent different stages of stromal activation in cancer. We then interrogated how this stromal approach compares with the prior tumor-centric approaches to come to a theoretical classification system to separate which patients might benefit most from antiangiogenesis and cancer immunotherapy.

Studies were conducted by Crown Bio, Shanghai, China. Subcutaneous primary human gastric models (HuPrime) were established and passaged in mice before use in efficacy studies as follows: small 2 to 4 mm in diameter fresh tumor fragments were implanted subcutaneously into BALBC/nude mice (6–8 weeks of age). When tumors reached an average volume of ∼150 mm³ (132–193 mm³), the animals were randomized by tumor volume into treatment groups (n = 7–8). For all statistical evaluations, the level of significance was set at P < 0.05.

The Ad-VEGF-A¹⁶⁴ flank model was performed as described in refs. 7 and 8. Samples for messenger RNA (mRNA) profiling studies were processed by Asuragen, Inc. using GeneChip Mouse Genome 430 2.0 Array (Affymetrix) as described previously (9). A summary of the image signal data, detection calls, and gene annotations for every gene interrogated on the arrays was generated using the Affymetrix Statistical Algorithm MAS 5.0 (GCOS v1.3) algorithm (scaling factor = 1,500) and quantile normalization across samples was applied. Log₂ transformation and mean-centered standardization was performed. Accession number in GEO is GSE76630.

Multiplexed fluorescent IHC was performed as previously described (10). Antibodies used were PECAM-1 (CD31) (AbCam ab28364 at 1:50) or (Invitrogen A11006 at 1:100); aquaporin-1 (AQP1) (Proteintech 20333-1-AP at 1:100 or 1:250); von Willebrand factor (vWF) (AbCam ab115771 at 1:50) or (Millipore MAB3442 at 1:250); and CD34 (Biolegend 119302 at 1:100) or (Novocastra clone QBEnd/10 NCL-L-END at 1:250); Smooth muscle actin (SMA; Sigma C6198 at 1:400); GLUT1 (Millipore 07-1401 at 1:500).

Multiplexed tissue imaging analysis of gastric tissue microarrays (TMA) was performed using the GE Multiomyx platform essentially as described in 11 (see Supplementary Methods for more elaborate descriptions of these methods). Briefly, four immunostains contributed to blood vessel area detection: CD31, CD34, vWF, and AQP1. Five clusters of vessel phenotypes were obtained: Early vessels were [CD34⁺/AQP⁻/vWF] and [CD34⁺/AQP1⁺/vWF⁻], and Late vessels were [CD34⁺/AQP1⁺/vWF⁺], [CD34−AQP1⁺vWF⁺], and [CD34⁻, AQP1⁻/vWF⁺]). For immune cell detection, CD3 (for T-cell detection) and CD163 (for macrophage detection) were stained. Four phenotypic classes of gastric tumors from the TMAs were derived according to threshold values for the percentage of early and late vessels (using a value of 55% of late vessels as cutoff, where values above are “late” and values below are “early”), percentage of macrophage area (using a value of 28% as a cutoff), and percentage of T-cell area (using a value of 3.4% as a cutoff).

Archived and de-identified FFPE tissues were provided by the Wood Hudson Cancer Research under a HIPAA waiver. The TMA was created from archived FFPE specimens. Clinical follow-up information was obtained from existing medical records in Tumor Registries. Archived surgical specimens were used to identify 1-mm cores from Donor Blocks for embedding into a Recipient Block of paraffin with receptacle holes prepared in a grid pattern using a Pathology Devices TMArrayer.

Tissues were imaged on a Marianas 575 system (Intelligent Imaging Innovations). DAB-stained flanks were reviewed and classified by pathologists using the classification established by Dr. Harold Dvorak's group (12). Quantifications of the vascular marker phenotypes were performed with an iCys research imaging cytometer (Thor labs; ref. 10). Triplicate gastric tumor TMAs and a control TMA with normal and tumor tissues were stained and quantified using two panels. The first panel included CD31, CD34, and AQP1 and the second had CD31, vWF, and AQP1. The control TMA was used to standardize the gating and classification from one stain run to the next. Any sample that had at least 2 cores and a consistent diagnosis (intestinal versus diffuse) were used. To classify the cores, a multistep analysis was used. From the first panel, the percentage of CD31 phantoms that were CD34⁺AQP⁻ was classified as outcome 1. From the second panel, the percentage of CD31 that was AQP⁺vWF⁻ was outcome 2, percentage of CD31 that was AQP⁺vWF⁺ was outcome 3, and finally the percentage of CD31 that was AQP⁻vWF⁺ was outcome 4. These four outcomes were normalized to 100% to reflect how many of the total vessels fell into each outcome. The percentage of vessels that fell into each of these outcomes was used for the vascular marker phenotyping to simplify the results and align the readouts with the bioinformatics analysis. Greater than 15% of outcome 4 = “mature.” Less than 16% of outcome 1 or >50% of outcome 1 + 2 = “immature.” Finally, all remaining cores were classified in “intermediate.”

The distribution of the eight possible phenotypes within the ACRG cohort was as follows (Signature 1/Signature 2/Signature 3): −/−/−, 32.5%; −/−/+, 6.02%; −/+/−, 0.80%, −/+/+, 13.65%; +/−/−, 19.28%; +/−/+, 3.21%; +/+/−, 3.61%; +/+/+, 20.88%. Because four dominant phenotypes were recognized, the four smaller subsets were distributed into the dominant classes as follows: −/−/+, −/+/−, and +/+/− grouped with −/+/+; +/−/+ grouped with +/−/−. See Supplementary Materials and Methods for additional details.

Reverse-phase protein array (RPPA) data from the TCGA cohort were mined from Supplementary Data in ref. 5. TCGA patients were classified into four stromal phenotypes as described, and mean protein expression values were reported for each group. Correlation analyses of analyte expression between stromal phenotypes and molecular classifications were performed using Spearman nonparametric analysis in JMP.

As in humans, mouse models also display a variety of responses to antiangiogenesis. We tested 17 patient-derived xenograft (PDX) gastric cancer models and found that they varied from very responsive to unresponsive to anti-VEGFR2 (DC101). A subset of these models is shown in Fig. 1A. The differential response was not obviously due to differences in tumor cell doubling time, histology or differentiation, but we did observe that stromal patterns correlated to some degree with response. The two examples shown in Fig. 1B represent extremes: the top panel illustrates a responsive tumor with numerous, large blood vessels with little hypoxia (as measured by Glut1 expression); the bottom panel depicts a nonresponsive tumor with small caliber vessels and pronounced hypoxia. We hypothesized that stromal phenotypes might be used to differentiate gastric cancers for susceptibility to stromal targeted therapies; i.e., blood vessel phenotypes for potential antiangiogenesis therapy and immune infiltrates for immunotherapy.

We started with a mouse model that drives a stromal response with an adenovirus expressing VEGF-A. This model is not only valuable for its ability to generate time-dependent vascular differentiation, but also allows the study of stromal remodeling associated with immune cell infiltration, ECM, and fibrotic tissue remodeling reminiscent of wound healing, which has been so often compared to cancer (13). The absence of tumor cells enables an enriched stromal environment to generate RNA signatures, which we then evaluated in human samples of gastric cancer.

In this model, viral infection and VEGF-A production is transient and largely localized to the initial site infiltration of immune cells and initiation of an immature vascular network. The large fragile blood vessels in this stage have been termed mother vessels (MV). At intermediate stages (days 5–20), the tissue is enriched by expansion of the vasculature and evolution of MVs into glomeruloid microvascular proliferations (GMP) and capillaries. Also at this stage, lymphatic vessels begin to form, and the immune infiltrate is cleared. Later stages (beyond day 20) are marked by a resolution/maturation/stabilization of vessels characterized by the heavy investment of pericytes, vascular malformations (VM), and remodeled extracellular matrix (12, 14, 15). This model was previously used to demonstrate that early-stage and intermediate-stage vessel subsets (MVs, GMPs, and capillaries) were sensitive to aflibercept (VEGF/PLGF Ligand trap) therapy, whereas more mature subsets of vessels (VMs, feeding arteries, and draining veins) were insensitive (Supplementary Fig. S1; refs. 15, 16).

We performed a gene array analysis using flank samples from days 0, 5, 20, and 60. Hundreds of differentially expressed genes were identified between each of the time points (Fig. 2A), indicating the dynamic nature of this model throughout the entire time course. Hypergeometric analysis of the gene array dataset indicates that there is an enrichment of genes associated with a variety of pathological and disease states, including cancer and cardiovascular disease (Supplementary Fig. S2A; ref. 17) and a dynamic regulation of genes associated with VEGF-A (Supplementary Fig. S2B). In addition to endothelial genes, there were differentially expressed genes involved in multiple cell types of stroma. Temporally speaking, the genes that were highly expressed early are involved in proliferation (Ki67), inflammation (Il-6, Pik3cg, granulocyte marker Tlr1, myeloid cell marker Cd68), and remodeling of matrix (Timp1 and Tnc; Fig. 2B). Pan-endothelial expressed genes such as pecam1 (cd31), VE-Cadherin (Cdh5), and Vegfr2 (Flk1) showed peak activation in the middle (days 5–20) time points. Pericyte markers Pdgfrβ, Rgs5, and Acta2 (Smooth Muscle Actin/SMA) had peak later than the endothelial cell genes consistent with pericyte coverage following endothelial tube formation and demarking the remodeling process that stabilizes vessels. TEK, a gene enriched in maturing blood vessels, is more highly expressed at later time points, while CD36, an immature blood vessel marker, has higher expression at day 5. Genes involved in lymphangiogenesis, such as VEGFC, have high expression between days 5 and 20.

We used multiple sources of gene lists (see Materials and Methods) to select down to a smaller set of genes restricted to those involved in tumor stroma: pathologic angiogenesis, wound-healing response, and tumor microenvironment biology. This resulted in a pool of 1,549 genes that were used as seeds for additional bioinformatics analysis using ClaNC and PAMR algorithms to generate a 108 gene signature of the most differentially expressed genes (Fig. 3A and Supplementary Table S1), which segregated themselves into three stromal signatures that reflected biological processes such as signature 1 (immune cell activity), 2 (angiogenesis), and 3 (tissue remodeling) based on gene ontology (GO) classifications.

We applied this 108 gene set to a publicly available gene array data set from Avastin (anti-VEGF-A)-treated H1975 NSCLC xenografts (Fig. 3B; ref. 18). Vehicle-treated samples were highly enriched in genes from signatures 1 and 2, while signature 3 genes were underrepresented. Conversely, Avastin-treated samples had a significant decrease of the signature 1 and 2 genes with a concomitant enrichment of signature 3 genes compared with vehicle samples. This analysis reveals the selective loss of less mature stroma (represented by signatures 1 and 2) upon treatment with anti-VEGF, consistent with prior preclinical and clinical studies (18–21). Avastin only blocks human VEGF-A in this murine model, so we also explored an anti-mouse-VEGFA antibody (G6) in our Ad-VEGFA flank model. We observed similar results in that signatures 1 and 2 were robustly diminished, while signature 3 remained unchanged (Fig. 3C and Supplementary Fig. S3). An antibody to VEGFR2 (DC101) in the same study was less robust at reducing signatures 1 and 2 but did significantly reduce (as opposed to increase) signature 3. These data provide evidence that there are some biologic differences in targeting VEGF versus targeting VEGFR2, but whether these are significant enough to translate to clinical differences remains unknown. Finally, we asked whether high signature 2 could be predictive for DC101 efficacy, and in 7 gastric PDX models, the highest signature 2 pulled out the only strong responder in that cohort (Fig. 3D).

To develop IHC assays, we focused on four genes whose peak gene expression ranged from early (CD34), middle (AQP1), and late (vWF) time points. We optimized a multiplexed fluorescent two-panel angiophenotyping assay (10). CD31 stained all vessel structures from each time point, consistent with its role as a pan-endothelial marker (22). CD34 primarily stained early and intermediate vessels (seen at days 5 and 20), with very dim staining of later-stage vessels (represented by day 60). AQP1 does not stain early-stage vessels, but strongly stains intermediate vessels. vWF expression is primarily found in only intermediate- and late-stage vessels (Fig. 4A and B). Comparison of the multiplexed imaging analysis with previously described blood vessel subtypes (12) showed that these four markers can be used to map to histologically defined subsets (Fig. 4B). The earliest vessel types that are formed in angiogenesis are MVs and GMPs, which only highly express CD34. As vessels mature into VMs, they begin expressing AQP1. With additional maturation (and pericyte recruitment) the VMs and capillaries also express vWF and start losing CD34 expression. The most mature form of vessels, the AVMs (including feeding arteries and draining veins) exclusively express vWF and do not have CD34 or AQP1 expression (Fig. 4B, note arrows). Therefore, these endothelial markers can molecularly profile subsets of vessels that have up to this point only been characterized by their morphology.

Eighty-five human primary gastric tumors were profiled in TMA format for vessel phenotypes using the previously described IHC assay (Fig. 5). Using this approach, we were able to classify each of these tumors into their dominant stromal phenotype using vascular markers. Interestingly, we observed differences in the frequency of stromal/vascular phenotypes between the two major histologies in gastric cancer, Lauren's “diffuse” and “intestinal” (Fig. 5A) (21). About 50% of intestinal-type gastric tumors contained predominantly immature stroma, as marked by high proportions of poorly lumenized CD34-postive vessels with little to no vWF staining (Fig. 5B, left). Conversely, an overwhelming majority of diffuse-type tumors (∼70%) contained predominantly intermediate stroma, indicated by the high proportions of their vessels staining positive for AQP-1 and with few CD34-single-positive (immature) or vWF-single-positive (mature) vessels (Fig. 5B, right).

Because we had also observed a dynamic pattern of inflammatory cell infiltration from the bioinformatics analysis of the Ad-VEGF-A¹⁶⁴ model and immune cells are likely to be important for emerging stromal therapies, we expanded our IHC methodology to include an immune cell multiplexed panel also. Using GE's Multiomyx platform (11) to analyze 20 vascular and immune cell biomarkers on the same gastric TMA samples, we were able to describe both the vascular and immune cell status of the same gastric cancer patient samples (Fig. 5C). The immune cell panel included T-lymphocytes (CD3, CD8) and M2 macrophages (CD163). Using this approach, we were able to classify four primary stromal phenotypes in gastric cancer, which we describe as vascular immature/noninflammatory (VINI), inflammatory (I), vascular mature/inflammatory (VMI), and vascular mature (VM). VINI tumors were characterized as having rudimentary vessels found in immature stroma, but lacking the significant lymphocyte infiltration that also characterized the early infiltrate that we observed in the day 5 Ad-VEGF model. In contrast, tumors with immature angiogenic markers and with high levels of lymphocytes were termed “Inflammatory.” These vascular beds could be characterized as having nonproductive early angiogenesis akin to what is seen following Notch inhibition or extreme hypoxia (23). Intermediate and Late vascular markers signified the VM tumors, which could also be observed to have either immune infiltrate or not, so we called those without immune infiltrate “VM” and those with a strong immune component “VMI.” We hypothesized that these four stromal phenotypes represent fundamentally different biologic states that might have differential survival outcomes and/or implications on therapeutic efficacy.

To further address our stromal phenotype hypothesis, we analyzed a publicly available gene array dataset, TCGA (5), from 249 patients with gastric cancer with our stromal signature (Supplementary Fig. S4). We observe that a majority of the diffuse gastric tumors clustered together and show high expression of signature 2 and 3 genes, consistent with the IHC findings reported in Fig. 5B.

The TCGA consortium devised a subclassification of gastric cancers into four distinct groups (C1–4) based upon a 40 gene tumor-enriched RNA signature. A cross-comparison of our stromal signature with these did not reveal any overlapping genes. Using the “C1–4” designation, we clustered these patient samples with our stromal signature (Supplementary Fig. S5A). Interestingly, as observed in the Lauren diffuse histology, the C1 tumors appear to have the highest expression of stromal signatures 2 and 3. This group of patients was defined as genomically stable with mRNA expression profiles characteristic of tumor cells undergoing epithelial-to-mesenchymal transition (EMT).

Application of our signature to another independent gastric cancer dataset from the ACRG (6) revealed a high expression of the signature 2 and 3 genes in the group of patients characterized as microsatellite stable MSS/EMT (Fig. 6A). 100% of TCGA C1 subtype and 95% of ACRG MSS/EMT tumors were predominantly enriched for signature 2 (Fig. 6B). A third independent gastric cancer cohort (24) also showed a very high correlation of signature 2 genes to their “Mesenchymal” molecular phenotype (Supplementary Fig. S5B and S5C). However, tumors from non-C1 TCGA and non-MSS/EMT ACRG molecular classification groups associated strongly with stromal signatures 1 and 3, albeit with heterogeneous distributions. This was also apparent from the heatmaps (Fig. 6A and Supplementary Fig. S5A), where some patients show high or low expression of genes from all clusters within the stromal signature or preferentially show gene expression in two of the three clusters. Therefore, using this forced clustering approach some patients will be placed into a particular cluster despite having an overall low gene activation score for that cluster.

The heterogeneity of signatures in each patient suggested that we needed to move away from forced clustering into a gene-by-gene binary segmentation of gene activation scores approach to describe a patient's stromal biology. We set a value of zero as the median expression score, such that scores above zero would be scored “+” and scores below zero would be scored “−”for signatures 1, 2, and 3, respectively. When we did this, four dominant combinatorial profiles were observed in both the TCGA and ACRG datasets, −/−/−, +/−/−, −/+/+, and +/+/+, which compared favorably with the four phenotypes observed by IHC (Fig. 6C). These four profiles accounted for 88% of tumors in the TCGA set and 86% of tumors in the ACRG set. The tumors within the other minor profiles were combined with the four dominant profiles for future analyses (see Materials and Methods). A comparison of the signature 1 genes confirms that these genes are highly correlated with inflammatory genes expressed in the TCGA and ACRG cohorts (Supplementary Figs. S6 and S7). Conversely, both signatures 2 and 3 correlated with “Endothelial” cell gene expression. Therefore, tumors with signature 1 scores above zero were expected to have an inflammatory phenotype (Fig. 6C). Similarly, tumors with signature 2 scores above zero were termed VM and are considered fundamentally associated with angiogenesis. Positive scores in both signatures 1 and 2 were grouped as VMI and negative scores in both were grouped as VINI. In the end, signature 3 did not reveal robust differentiation in the gastric cancer samples.

Next, tumors from the ACRG and TCGA cohorts were classified by these four stromal phenotypes, and the subsequent distributions within each molecular subtype were calculated (Fig. 6D). Consistent with the earlier observation that TCGA C1 and ACRG MSS/EMT subtypes had enrichment for stromal signature 2 genes, almost all tumors from these groups had a VM gene expression phenotype. However, a large percentage of these tumors (52% in TCGA C1 and 65% in ACRG MSS/EMT) also had expression of inflammatory genes, making them VMI. The TCGA C2 subtype was composed largely of tumors with inflammatory stromal phenotype (39%) with another significant portion of tumors having a VMI phenotype (30%). The ACRG microsatellite instability (MSI) group had a similar proportion of inflammatory stromal phenotypes (38%) but fewer VMI tumors (13%). Both TCGA C2 and ACRG MSI classes had very few purely VM stromal gene expression profiles (6% and 8%, respectively). Therefore, the TCGA C1 and ACRG MSS/EMT tumors can be classified as dominantly VM, while the TCGA C2 and ACRG MSI are dominantly inflammatory. The other two classes from the TCGA and ACRG cohorts had varied distributions among the four stromal phenotypes. Supplementary Table S2 describes relationships within our stromal subtypes for important genetic features, such as MSI, EBV, Her2, and others.

Activation scores for multiple gene sets previously described in the literature and associated with unique biology, such as stroma, inflammation, regulatory T cell (Treg) biology, were calculated for each patient's tumor from the ACRG and TCGA cohorts and were subjected to pairwise correlation analysis (Fig. 6E and Supplementary Figs. S6 and S7). The gene activation scores for various gene groups were consistent with their stromal designations. For example, inflammatory tumors were enriched in immune cell genes, including macrophages and T cells, whereas VM tumors were highly enriched in angiogenesis genes. Additionally, tumors with a combined VMI phenotype tended to have significantly higher expression of genes associated with downmodulation of the immune response (Fig. 6E and F). These include genes such as HAVCR2 (TIM3 ref. 25), IL10 (26, 27), and TGFβ1 (26, 28) that are associated with Tregs and suppression of Th1 immune responses and VCAM1 (29), CD163 (30), and ITGAM (CD11b, ref. 31) that are associated with M2 macrophage infiltration. Conversely, tumor with an inflammatory phenotype have significantly higher expression of immunostimulatory genes (TNF, IFNG, and ICAM1; refs. 26, 29, 31) and markers of M1 macrophage activation (CD274, IDO1, and LAG3; refs. 26, 32). This suggests that the immune cell components within the Inflammatory and VMI phenotypic groups represent distinct biologic activation states and are likely to influence their respective tumor growth in disparate ways.

We hypothesized that gastric tumors with similar stromal phenotypes would also share similarities in protein expression and cellular signaling, and ultimately relate to overall patient survival. To explore these ideas, we first mined the RPPA data from the TCGA cohort (5) to examine protein expression profiles within the different stromal phenotypic classes. Of the 191 analytes examined (5), 43 were highly differentially expressed across the four stromal phenotypic classes (Fig. 7A). Plots of the mean relative expression value for these analytes show two distinct patterns of expression, with the VM and VMI classes having very similar trend lines. The inflammatory and VINI also share a common trend line with each other, but in an opposite direction to the VM and VMI expression profiles. A closer look at some of these analytes reveals striking differences in expression of proteins involved in angiogenesis (VEGFR2, ACVRL1, Rictor; refs. 33, 34) and inflammatory processes (Syk, CD49b, PAI1; refs. 35–37; Supplementary Fig. S8A). Analysis of the entire 191 RPPA set shows that the overall protein expression profiles within the stromal phenotypic groups correlate to the same molecular classes, as was demonstrated with the gene expression profiles (Supplementary Fig. S8B). This suggests that gastric tumors defined by different gene expression stromal phenotypes also have differential protein expression patterns and, by extension, a different tumor biology.

Previously, survival analysis of the ACRG cohort was performed with a patient follow-up time up to almost 9 years, with a minimum of 53 months of census data (6). A new Kaplan–Meier analysis was performed with updated data that include a minimum of 84 months of census data for each patient. The same trends that were previously reported still held true for each of the four ACRG subclasses in this updated data (Supplementary Fig. S9). As observed with the RPPA data, two main stromal phenotypic classes appear to exist: VM and vascular immature. Kaplan–Meier analysis of the ACRG cohort subdivided into angiogenic and nonangiogenic groups reveals a robust difference in survival (Fig. 7B). Therefore, the stromal status of gastric tumors has prognostic implications, such that patients with VM tumors fare more poorly than those with Vascular Immature tumors. Although the differences between each subgroup are smaller when using all four stromal phenotypes (Fig. 7C), there is still statistical significance and some interesting implications. Similar to the ACRG MSS/EMT and MSI classes, the VM group has the poorest overall survival while the I group has the best survival. The VINI and VMI groups have intermediate survivals, with the VMI trending closer to the VM group.

Given the available treatment options for gastric cancer and the increasing interest in immunotherapy, differentiation of the inflammatory and VMI phenotypes in terms of Th1/2, dendritic cell activation and Tregs may help determine where combinations of antiangiogenesis and checkpoint inhibitor blockade could be best tested with experimental agents. Because these tumor samples were from treatment-naïve patients with mixed grades and histology, there is the greatest opportunity for these subgroups to most closely reflect a first-line setting, though it remains to be tested. In addition, there was a higher percentage of early-stage samples in the TCGA data set so in the end the distribution of subsets from any analysis including this one will vary depending on the samples collected. Even when focused on just the stage 3 and 4 samples, our stromal subtypes retained their prognostic value. At this time, the only targeted therapy in first-line gastric cancer is for IHC HER2 3⁺ and/or HER2-amplified tumors. Therefore, we assessed how these potential stromal subgroups would divide when the HER2 high expressers were defined separately (Fig. 7D). Data with Her2 included can be found in Supplementary Table S2.

Expression profiling has been used to explore molecular classification of cancer across multiple indications. With the availability of deep-sequencing data in recent years, there is an opportunity to take different views of subclassification and when appropriate, integrate data from the DNA, RNA, protein profiles with clinical outcome data. To date, the success in identification of patient subsets for therapeutic matching has primarily been based on single gene mutations or high expression in tumor cells. When this has been successful, such as for anti-Her2–targeted therapies in breast and gastric cancer, it can largely be attributed to the targeting of dominant driver mutation. But for molecules with targets in the tumor stroma, whether antiangiogenesis or immunomodulatory, it would seem equally appropriate that our focus turn toward phenotyping the stroma. Of course tumor genetics undoubtedly influences the gene expression that contributes to a stromal phenotype. Gastric cancer has few target-directed options and only trastuzumab is currently approved for HER2 high expressing cancers. Recently ramucirumab was approved in the second-line setting as a monotherapy or in combination with paclitaxel (38, 39), and there are several early-phase clinical trials ongoing with checkpoint inhibitors. This study was undertaken to investigate the heterogeneity of gastric cancer stroma to support the future study of available stromal-targeted therapies. Ultimately, whether stromal signatures, tumor signatures, genetics or combinations of the above most potently predict therapy remains to be explored.

Technically, it is very difficult to develop a stromal signature from tumor tissue, as there can be such variability in stromal content from sample to sample. For the TCGA profiling, an effort was made to select sections enriched for tumor cells (5). Mouse tumor models pose an even greater challenge as we can readily see the relative dearth of stroma in cell line xenografts as well as PDX tumor models compared with human cancer (unpublished). Therefore, the stromal model used here had clear advantages in being temporally and phenotypically reproducible. While this model uses a Nude mouse host, a robust myeloid immune infiltrate found there was able to be extended to lymphoid phenotypes by studying human samples. While VEGF-A is the dominant initiator of angiogenesis in this model, its viral expression is short lived, and additional angiogenic factors arise to contribute to a complex vascular bed. There is also some influence of the Adenovirus itself in the immune cell influx.

Using molecular classification of cancer to supplement classical pathology can reveal differences in prognosis or be predictive of treatment outcomes for patients that are otherwise indistinguishable by histology. ACRG survival data paired with TCGA protein array data showed that the most striking difference in our phenotypes and their prognosis is seen between the patients with productive angiogenesis from those without it (Fig. 7). However, when more closely examined there were differences in prognosis also when considering the immune phenotypes in the VMI and inflammatory subgroups. The profile of the immune markers that dominate those groups suggests that the inflammatory subgroup trends toward a more productive immune response with markers of cytotoxic T cells and low incidence of immunosuppressive gene expression, similar to what has been reported (40). Notably, the MSI cohorts described by the TCGA most closely associated with the nonangiogenic phenotype, but were split relatively equally between the I and VINI groups (Supplementary Table S2). MSI high tends to be enriched for activated T cells and checkpoint regulators such as PD1 and PD-L1 (41). Interestingly, thus far clinical data suggest that only about half of the MSI high patients respond to checkpoint inhibition.

On the other hand, the immunologic microenvironment in the subgroup associated with angiogenesis was weighted toward immunosuppression, perhaps not surprising given the role VEGF-A can play in promoting M2 myeloid phenotypes and inhibition of dendritic cell maturation (42). In this setting, we hypothesize that it might require more than just relieving checkpoint inhibition to enable a robust immune response and combination approach with antiangiogenics added to checkpoint inhibition may be worth exploring. Finally, it is worth mentioning that there was a subgroup with relatively little productive angiogenesis or immune cell infiltration, the VINI subgroup. This group is a major subset of the p53 mutant/microsatellite stable ACRG subgroup (see Fig. 6D), which was described by the ACRG consortium as having the most RTK amplification (6). Thus, more targeted therapies may emerge that are well suited to this cohort in the future.

M.T. Uhlik is Vice President, Translational Oncology, at Biothera Pharmaceutical, reports receiving commercial research grant from Biothera Pharmaceutical, and has ownership interest (including patents) in Biothera Pharmaceutical and Eli Lilly and Company. H.F. Dvorak reports receiving commercial research grant from Lilly. No potential conflicts of interest were disclosed by the other authors.

Conception and design: M.T. Uhlik, J. Liu, B.L. Falcon, S. Iyer, C. Lowes, S. Chintharlapalli, A. Fischl, D. Gerald, Q. Xue, S.-c. Jaminet, A. Nasir, H.F. Dvorak, L.E. Benjamin

Development of methodology: M.T. Uhlik, J. Liu, B.L. Falcon, S. Iyer, J. Stewart, M. O'Mahony, C. Sevinsky, D. Gerald, Q. Xue, Y. Al-Kofahi, F. Ginty, A. Nasir, H.F. Dvorak, L.E. Benjamin

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M.T. Uhlik, J. Liu, B.L. Falcon, S. Iyer, J. Stewart, H. Celikkaya, M. O'Mahony, C. Sevinsky, C. Lowes, L. Douglass, C. Jeffries, D. Bodenmiller, Q. Xue, J.-y. Lee, J.H. Carter, A. Nasir, J.A. Nagy, H.F. Dvorak

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.T. Uhlik, J. Liu, B.L. Falcon, S. Iyer, J. Stewart, H. Celikkaya, C. Sevinsky, L. Douglass, C. Jeffries, D. Bodenmiller, S. Chintharlapalli, A. Fischl, Q. Xue, A. Santamaria-Pang, Y. Al-Kofahi, Y. Sui, K. Desai, T. Doman, A. Aggarwal, J.H. Carter, B. Pytowski, S.-c. Jaminet, F. Ginty, A. Nasir, H.F. Dvorak, L.E. Benjamin

Writing, review, and/or revision of the manuscript: M.T. Uhlik, J. Liu, B.L. Falcon, S. Iyer, C. Sevinsky, C. Lowes, S. Chintharlapalli, A. Fischl, D. Gerald, A. Santamaria-Pang, Y. Al-Kofahi, A. Aggarwal, F. Ginty, A. Nasir, J.A. Nagy, H.F. Dvorak, L.E. Benjamin

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M.T. Uhlik, B.L. Falcon, S. Iyer, H. Celikkaya, L. Douglass, Q. Xue, A. Santamaria-Pang, J.H. Carter, B. Pytowski, F. Ginty

Study supervision: M.T. Uhlik, S. Iyer, H.F. Dvorak, L.E. Benjamin

Other (developed image and data analysis algorithms blood vessel segmentation and classification of immune cells): A. Santamaria-Pang

The authors thank G. Plowman and D. Ferry for critical reading and editorial suggestions, C. Reinhard for helpful discussions, and Sheng Guo at CrownBio for analysis of PDX expression profiles.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.