Key biological processes driving metastatic spread of pancreatic cancer as identified by multi-omics studies.

Pancreatic ductal adenocarcinoma (PDAC) is an extremely aggressive malignancy, characterized by a high metastatic burden, already at the time of diagnosis. The metastatic potential of PDAC is one of the main reasons for the poor outcome next to lack of significant improvement in effective treatments in the last decade. Key mutated driver genes, such as activating KRAS mutations, are concordantly expressed in primary and metastatic tumors. However, the biology behind the metastatic potential of PDAC is not fully understood. Recently, large-scale omic approaches have revealed new mechanisms by which PDAC cells gain their metastatic potency. In particular, genomic studies have shown that multiple heterogeneous subclones reside in the primary tumor with different metastatic potential. The development of metastases may be correlated to a more mesenchymal transcriptomic subtype. However, for cancer cells to survive in a distant organ, metastatic sites need to be modulated into pre-metastatic niches. Proteomic studies identified the influence of exosomes on the Kuppfer cells in the liver, which could function to prepare this tissue for metastatic colonization. Phosphoproteomics adds an extra layer to the established omic techniques by unravelling key functional signaling. Future studies integrating results from these large-scale omic approaches will hopefully improve PDAC prognosis through identification of new therapeutic targets and patient selection tools. In this article, we will review the current knowledge on the biology of PDAC metastasis unravelled by large scale multi-omic approaches.


Introduction
The ability to spread and adapt to a hostile environment is one of the most dangerous characteristics cancer cells can acquire. This invasive feature encompasses multiple progression steps which start by acquiring the ability to migrate through and out of primary tissue, followed by invading and surviving in blood or lymphatic vessels, and finally establishing a new tumor microenvironment in the hostile receiver tissue 1 . Multiple mechanisms, e.g genetic instability and/or clonal expansion, explain the genetic evolution that tumors undergo to develop from a localized carcinoma into aggressive metastatic disease. Once these aggressive features are obtained, the survival of these patients reduces exponentially, and this is independent of the origin of the primary tumor. Metastases can develop very late after diagnosis implying a slow adaptation of a single cell in the distant microenvironment, but in some tumors, metastases develop synchronously during growth of the primary tumor 2 . This quick progression to metastatic disease is a key feature in pancreatic ductal adenocarcinoma (PDAC) patients, who often succumb from advanced local disease with widespread metastatic burden early after diagnosis 3 .
Patients suffering from PDAC have a 5-year survival of 7.7% 4 . This disease is the fourth leading cause of cancer-related deaths in the USA and is projected to be the second cause of cancer-related deaths by 2030 5 due to an increasing incidence, lack of effective treatments, and a high metastatic propensity. The metastatic risk is highlighted by the fact that 91% of PDAC patients are diagnosed with (regional) metastatic disease 4 . This quick progression warrants new studies to understand the key processes that drive metastatic behavior.
Several studies investigated the effect of localization and size of metastases and the timing of migration. Rapid autopsy programs revealed that the majority of metastases are located in the liver (76-94%), peritoneum (41-56%), abdominal lymph nodes (LN) (41%) and the lungs (45-48%) 3,6-8 ( Figure 1A). Moreover, these programs revealed the extent of metastatic disease in these patients, with on average 2.9 distant organs involved. Additionally, metastases occur mostly in a widespread fashion, with some patients harboring over 1000 metastases 7 . The size and the growth rate of metastases is inversely correlated to survival 3 .
PDAC is characterized by an interactive microenvironment with up to 90% of the primary tumor consisting of a stromal compartment 9 ( Figure 1B). Interestingly, liver metastases harbor pathological resemblance to the primary PDAC tumor with similar extracellular matrix (ECM) components 10 . Data suggest that PDAC cells can recruit local stromal cells and create an extracellular environment similar to the primary tumor upon colonization of the receiving organ 10 . Using a fluorescent lineage tracing mouse model, Aiello et al. 11 showed that there is a correlation between the size of the liver metastasis and the recruitment of stroma. This model was previously used to show that epithelialmesenchymal transition (EMT) proceeds metastasis and that PDAC cells retain their EMT state in circulation 12 .To allow growth of the metastasis, PDAC cells need to return to their epithelial morphology after their previous mesenchymal state ( Figure 1C). Upon multiplication of tumor cells, as early as in nano-metastases (2-10 cells), myofibroblasts were found to be in contact with tumor cells and ECM composition recapitulated that of the primary tumor 11 . This highlights the complex biology of PDAC and its metastatic features.
Large-scale omics have been performed in order to understand the biology of PDAC's aggressive behavior. Emerging insights from these studies can be used to elucidate the rapid progression to metastatic disease. In the following sections research involving genomic, transcriptomic and proteomic studies of PDAC and its metastases are reviewed.

Genetic events and clonal evolution of PDAC
Multiple genetic events are thought to occur before metastatic spread is initiated. Four commonly mutated genes characterize PDAC. The KRAS driver mutation is identified in more than 90% of the PDAC tumors 13 . Mutations at codon 12 (G12D or G12V) are most abundant and result in aberrant, persistent activation of the KRAS pathway. Inactivating mutations in TP53, CDKN2A and SMAD4 (DPC4) are also very common, occurring in 74%, 35% and 31% of all PDAC patients respectively 14 .
However, the presence of these somatic common mutations in the primary tumor does not clearly correlate to patients with a very long survival (more than ten years after resection) 15 . Yachida et al. 6,16 performed comparative analysis of primary tumors with matched metastatic tissue and showed that over 90% of tissues had concordant driver gene mutations between matched primary and metastatic material, indicating that these mutations are early events in PDAC tumorigenesis. These results were recently confirmed in another study, which showed low genetic heterogeneity between metastases and the primary tumor 17 . In contrast to very long survival times, the number of mutational driver genes in a tumor is correlated to the metastatic burden of the patients and disease free survival (DFS) 6 . Especially, mutations of TP53 and SMAD4 are associated with a higher metastatic burden and poor prognosis of PDAC patients ,6,7,10 , compared to KRAS or CDKN2A mutations, which are associated with oligometastases. The correlation of driver gene mutations with metastatic burden was independent of tumor stage or grade. Of note, in 37% of the patients analyzed all four driver genes were mutated 6 . This indicates that some PDAC tumors do not follow the original progression model of sequential mutations in the four known driver genes during the development from Pancreatic Intraepithelial Neoplasia (PanIN) precursor lesions to PDAC 19 . Most likely, a subset of tumors acquire mutations in a different order and do not need all four mutations for tumorigenesis. This finding was further highlighted by Notta et al. 20 who showed by whole genome sequencing that complex rearrangements in the genome of the tumors can take place by chromothripsis, where genomic rearrangements are clustered on a small number of chromosomes, and that 67% of analyzed tumors showed a non-conventional mutagenesis resulting in quick tumor progression and mutation of driver genes. This particular genetic rearrangement was also identified by Waddell et al. 14 and the mechanisms and its contributions to PDAC will need to be further explored.
To further elucidate the genetic aberrations driving this aggressive tumor, multiple other sequencing analyses of primary PDAC tumors have been performed. Jones et al. 21 sequenced protein-coding exon DNA from 24 tumors and showed an average of 63 mutations per tumor. Despite the identification of multiple mutations in this PDAC cohort, identical mutations in more than one patient were sparse. However, pathway analysis of this heterogeneous mutational landscape identified 12 aberrant pathways affected in PDAC tumorigenesis, which were altered in 67-100% of the tumors.
Gene expression analysis confirmed differential expression of these pathways in tumors compared to normal epithelial pancreatic duct cells. These pathways include known driver pathways, such as KRAS signaling, but also highlighted pathways that function in tumor-stroma crosstalk like Hedgehog, TGFβ, integrin and WNT-NOTCH signaling.
Exome sequencing and copy number analysis established additional frequently mutated genes in the core affected pathways, and axon guidance was identified as a new aberrant pathway 22,23 . The importance of this pathway in PDAC biology was underlined by epigenetic genome-wide methylation analysis of 167 tumors, where axon guidance was identified as one of the most significant epigenetic deregulated pathway with the promotor regions of SLIT/ROBO signaling being hypermethylated 24 .
The identification of axon guidance as a key pathway in PDAC is supportive of the fact that perineural invasion is a poor prognostic factor in PDAC 25 . Moreover, peri-neural growth and interaction of tumor cells with the nervous microenvironment has been shown to stimulate the migration of PDAC cells 26,27 . When surgical margins are clear, cell migration in the peri-pancreatic nerve system can indeed be a source for later recurrence and metastases 25 .
The genomes of PDAC tumors contain certain types of chromosomal rearrangements. In a comparative analysis of primary tumor and metastatic tissue, genomic instability was shown to be very heterogeneous between patients. Intra-chromosomal rearrangements were more common than deviations between chromosomes. Interestingly, fold-back inversions were commonly present in tumors 28 . Other genomic variation screens identified similar trends favoring intra-chromosomal rearrangements 29.14 . Four different genomic subtypes were identified, of which the locally rearranged type showed foci in possible target oncogenes, however with low penetrance in the whole patient group, and the unstable subtype was identified by wide-spread genomic instability, most likely due to DNA repair dysfunction 30 . Another study based on the mutational signatures also identified four subtypes, of which the DNA repair dysfunction signature was correlated with increased tumor immune response 31 (Figure 2A). The clinical value of these genomic subtypes still needs to be further explored. However, the genetic rearrangement seems to be relatively stable between the primary and metastatic tissue of patients, for example, fold-back inversions are identified in matched samples, indicating an early event in tumorigenesis 28,31 .
The moment of dissemination of PDAC cells during tumorigenesis remains a point of discussion. The genetic and clonal evolutions, which are needed for the PDAC cells to gain metastatic capabilities, can be explained by multiple mechanisms. Campbell et al. 28 showed that most genetic structural aberrations of the primary tumor were present in metastases in their cohort of 13 PDAC patients, however, some patients showed genetic variations between different metastases. This proves that multiple subclones in the primary PDAC can establish different distant metastases. In particular, some mutations and rearrangements can enhance the metastatic capabilities of a clone to colonize a specific organ site. For example, two patients harboring lung metastases showed a more extensive evolution from the primary tumor than abdominal metastases, and contained MYC and CCNE1 mutations, possibly enhancing the lung-seeding capacity of these cells. Interestingly, in the study of Witkiewicz et al. 32 , MYC amplification was correlated to poor survival, a correlation that will need to be further investigated in the perspective of metastasis.
By Sanger sequencing, Yachida et al. 16 recognized that mutations in metastases are most likely clonal, and that the genetic heterogeneity in different metastases results from subclones from the primary tumor. By computational modeling they estimated the average time of development from carcinoma in situ to gain of metastatic competence to be 6.8 years 16 . Another mathematical model based on clinical progression and autopsies, predicted metastasis most likely to be present at time of diagnosis. This risk is correlated to the size of the primary tumor, underlining the need for early detection and improved therapeutic options 3 . New models evaluating the genetically diverse subclones and their metastatic capability are needed to shed light on the exact timing of dissemination during carcinogenesis.
The clonal progression of these tumors has recently been further evaluated by a mouse model with confetti lineage labeling of tumorigenic cells 33 . By following the fluorescent cells, subclonal heterogeneity of primary mouse PDAC tumors can be tracked in the metastatic sites. Interestingly, the majority of the abdominal metastases were polychromatic, indicating that they resulted from multiple subclones of the primary tumor. This polyclonality was mostly a result of two founder cells.
Functional experiments validated that most likely, these metastases were formed from cell clusters rather than from seeding of multiple single cells. Remarkably, larger lung and liver metastases consisted mostly of monoclonal cells. This monoclonality was directly correlated to the size of the metastases, indicating a genetic advantage of a subclone after establishment of the distant tumor.
This clonal progression from polyclonality to monoclonality in liver and lung metastases could be an explanation for the genetic heterogeneity found in some previous genomic studies 16 , since upon monoclonality there is reduced resemblance to the primary tumor with multiple subclones.

Transcriptomics and gene expression of metastases
Although deep sequencing of the genome has identified multiple aberrantly regulated pathways contributing to pancreatic cancer, the mutational status does not explain the full spectrum of the aggressive PDAC phenotype. Since multiple cellular processes can influence gene expression, for example epigenetics, regulation by transcription factors, and cell-extrinsic factors, screening of differential expressed transcripts can deepen the understanding of tumor biology.
Efforts identifying prognostic important subtypes in PDAC resulted in multiple transcriptomic classifiers ( Figure 2A); however, different approaches to the heterocellular consistency of PDAC have been used. Collisson et al. 34 were the first to describe three subtypes in PDAC, each with a different biology and prognosis. Their dataset consisted of 27 microdissected tumors to enrich for the epithelial compartment of PDAC. Of the three subtypes identified (quasi-mesenchymal (QM), exocrine and classical), the QM subtype was associated with the poorest prognosis, while patients identified retrospectively with the classical subtype showed relative good overall survival. The relatively higher expression of mesenchymal genes in the QM subtype very likely contributes to a higher occurrence of EMT and thus the ability to metastasize, leading to poor prognosis. However, the need for elaborate microdissection and extensive genetic analyses will hamper the clinical applicability of their classifier. A more feasible approach recently was suggested in a study which identified immunohistochemical classification markers (KRT81 and HNF1A). Interestingly, in this study the exocrine subtype was more resistant to paclitaxel treatment and tyrosine kinase inhibitors due to cytochrome P450 3A5 expression 35 . These results indicate that subtyping can have clinical applications. Another way to overcome the problem of microdissection, is bulk tumor analysis of high percentage epithelial tumors. Bailey et al. 23 identified four stable subtypes on an initial patient dataset (n=96) consisting of tumors with minimally 40% epithelial cellularity. These subtypes were validated on a larger cohort with a more natural distribution of the stromal compartment. Their poor survival "squamous" subtype resembled the QM subtype previously described. Other subtypes identified were "pancreatic progenitor", "immunogenic" and "aberrantly differentiated endocrine exocrine", which partly overlapped with the previous classical and exocrine subtypes 34 .
To account for the interaction between stromal and epithelial compartments and reduce the selection bias for only high epithelial tumors in analyses, Moffitt et al. 36 explored bioinformatics tools to dissect gene expression profiles of both. They identified two stromal subtypes, "activated" and "normal" stroma. These stromal subtypes were correlated to prognosis of PDAC patients, however their implication in metastases was not described and has not yet been further investigated.
Interestingly, the stroma subtypes were relatively underrepresented in the metastases in their dataset, indicating less stroma signatures in metastases. This study also identified two prognostic PDAC tumor subgroups, basal-like and classical, which resemble in part two of the Collison subtypes ( Figure 2A). The basal-like subtype was specifically enriched in metastatic tissue, implicating that transformation to a basal-like state is necessary for dissemination, or that some PDAC subtypes are more prone to metastasize. Interestingly, upon differential analysis of their classifier transcripts from matched primary and metastatic tissues, low heterogeneity was identified providing evidence that subtype specific gene expression is preserved in metastatic tissue.
Even though the computational dissection of bulk tumor identified plausible subtypes, consensus of all the different subtyping efforts is needed to further evaluate clinical relevance and utility. A largescale laser microdissected dataset to experimentally prove gene expression profiles from different compartments of bulk tumor will help to improve subtyping these tumors and define consensus subtypes. The QM subtype was concordantly identified so far in the large-scale studies and is consistently correlated to poor outcome and increased metastatic potential. This finding is in line with transcriptomic analysis of other tumor types where the mesenchymal classification is the most aggressive subtype 37-39 .
Gene expression profiling has been used to identify important genes and pathways in metastases. In a comparative analysis, genes with functions in cell proliferation, cell cycle regulation, were overexpressed in tumors with lymph node metastases. Moreover, apoptosis and cell motility were down regulated 40 . Stratford et al. 41 compared tissue from primary tumors with and without metastases and identified 6 genes related to metastases (FosB, KLF6, NFKBIZ, ATP4A, GSG1, SIGLEC11). These genes were prognostic for survival. Interestingly, another study with matched metastatic tissue from multiple locations did not show evident differential gene expression 42 . This could be explained by the hypothesis that the primary tumor already acquired the metastatic abilities and that the majority of gene expression does not show substantial heterogeneity from the primary tumor. Finally, a recent study explored the metabolic differential expression of PDAC and its metastatic tumors. Compared to normal pancreatic tissue there was enrichment of aerobic glycolytic genes. The majority of metabolic genes were comparable between tumor sites. Specifically, glucose transporter (GLUT1) was overexpressed in all tumor sites. SLC2A2 was overexpressed in liver metastases and IDH3 was overexpressed in lung and lymph node metastases, indicating some differential energy acquiring process in different metastatic locations 43 . Since the RNA technology is advancing with increasing depth of analysis, future studies will define more differential gene expression profiles of subtypes and prognostic profiles for metastatic disease. This hopefully will lead to better patient staging and clinical management for subgroups of patients. Integrative analysis combining mRNA and miRNA profiles can highlight their regulatory network. This was shown in a cohort of nine patients, where 3 miRNAs (miR-21, miR-23a, miR-27a) were identified as regulators for multiple known tumor suppressors 51 . Future large-scale efforts combining RNAseq and small RNAseq of primary and matched metastatic tumors will further deepen our knowledge of this regulatory network and its importance in metastases of PDAC.

Proteomics applied to PDAC to identify new biomarkers and protein subtypes
As outline above, PDAC is caused by alterations in DNA that yield altered gene products that make cells grow in an uncontrolled way and spread throughout the body. Comprehensive analysis of the alterations in each tumor's complete set of functionally relevant proteins, the proteome, can add a complementary layer of information that is expected to increase our understanding of how molecular changes interact to drive the disease.
In recent years, the field of proteomics has evolved from limited protein inventories to in-depth (close to) proteome-wide discovery due to massive improvements in mass spectrometry (MS) technology and (bio)informatics tools. Together with robust relative protein quantitation based on label-free or stable isotope labeling, proteomics studies are increasingly being used in cancer research. Moreover, the integration of proteomic and genomic data by which MS/MS data is searched against customized databases of individual matched DNA/RNA sequence data, referred to as proteogenomics, has enabled a more comprehensive view of the molecular determinants that drive cancer than genomic analysis alone and may help to identify the most important targets for cancer detection and intervention. This was recently shown for colon, breast and ovarian cancer 52,53,54 . Importantly, from these studies it also became apparent that proteome profiling data can outperform transcriptome profiling data for co-expression based gene function 55 , underlining the importance of proteomics in gene function and human disease studies.
Mass spectrometry based identification of proteins in complex biological samples such as tissues and biofluids has been performed to develop multiple cancer diagnostic applications. To identify protein biomarkers for non-invasive applications, proximal fluids that contain relatively high levels of tumor secreted proteins are an appealing biomarker source since they do not contain high levels of albumin that will mask low abundant biomarkers in blood-based screens. Another option to facilitate biomarker identification is to remove highly abundant proteins from blood. In PDAC, multiple studies have been performed in the recent years applying these techniques. A summary of the studies is described in Table 1   . We will describe some of the different approaches in more detail below. These proteins showed very good sensitivity and specificity as a panel in a validation cohort of over 300 patients. These studies prove that multiple biofluids can be a source for protein biomarker identification. Future validation is needed to establish these proteins for clinical use.
Another way to implement protein biomarkers is to classify subtypes previously established by transcriptomics. Kuhlmann et al. 59 made use of primary cells lines representing three Collison transcriptomic subtypes to detect subtype-specific protein biomarkers. Interestingly, in their cohort only the exocrine subtype had differential protein expression detectable on their cell surface or as secreted proteins. Further proteomic studies are needed to validate the transcriptomic subtypes in PDAC and their value or to establish whether proteome data may yield a different classification system. For the latter purpose, large-scale proteome profiling of clinical samples is needed as recently performed for other tumor types 82 .

Identification of prognostic and metastatic protein markers in PDAC
To understand the variability in survival of PDAC patients, several studies have explored differential proteome landscapes of long versus short survival or metastatic versus non-metastatic disease.
Matched formalin-fixed paraffin-embedded (FFPE) tissues from very long (more than ten year survival post-surgery) and short surviving patients were screened to understand the different underlying biology. In the short survival group, proteins associated with the cytoskeleton were increased as well as RNA processing / protein biosynthesis. This can point to higher motility and metastatic capability. Interestingly, one of the upregulated proteins identified was galectin-1 (LGALS1) which is mainly expressed in cancer-associated fibroblasts (CAF) in PDAC 83 . Knockdown of this gene reduced their ability to migrate and therefore to stimulate PDAC cells 84 .
In another approach to understand the basic principles of metastatic capability, Naidoo et al. 85 analyzed seven primary PDAC samples and their LN metastases. This yielded 856 commonly expressed proteins, of which the majority clustered in the biological functions of cell proliferation and growth, cell death and cellular movement. Only a small subset of proteins was differentially regulated, implying again that malignant epithelial cells in LN metastases are not very different from primary tumor cells. One of the proteins differentially expressed was S100P, which was validated by IHC. This protein was previously identified as an important player in trans-endothelial migration of PDAC cells and upon knockdown, less migration into the vasculature and less metastases were seen in a fluorescent zebrafish model 86 . These results show that this protein could be an interesting target to inhibit migration of tumor cells, and moreover, that differential protein expression can lead to new targets and understanding of cancer biology.
Another factor thought to contribute to metastatic disease, is the population of cancer stem cells (CSC). These cells have, or have regained, the ability to self-renew and are recognized as important modulators of metastatic capabilities and chemoresistance 87 . In an effort to elucidate their biology, Brandi and collaborators 58,88 profiled the proteome and the secreted proteins of a CSC subpopulation in the PDAC cell line PANC1. These analyses showed that CSCs upregulated multiple metabolomic pathways. Moreover, CSC were relatively sensitive to metabolic inhibition by existing drugs 89 . Similar metabolomics pathways were dysregulated in the secretome of CSCs 90 . Interestingly, the secreted proteins could be identified by ELISA assays in blood sample of PDAC patients, indicating their possible use as biomarkers (Table 1).

Exosome protein content and establishment of the metastatic niche
In recent years, we have come to understand that migration of cancer cells into the vasculature or lymph vessels by itself is not enough to establish metastasis. The hostile environment of the distant site requires certain changes that enable tumor cells to attach and thrive. Exosomes can be a player at the metastatic site to assist adhesion and growth of a tumor. Exosomes are small extracellular vesicles of endosomal origin that can carry nucleic acids and proteins, and have been shown to be able to influence the migratory capacity of tumor cells 91  Kim et al. 102 analyzed three primary cell lines of one patients' metastatic sites. In the normal proteome of these cell lines, 58% of the protein expression was similar. However, the phosphoproteome was highly variable between the different metastatic cell lines. This can indicate a dynamic state of phosphorylation, or differential aberrant activations after clonal evolution. It was found that AXL was phosphorylated in liver and lung metastatic cells, but not in peritoneal metastasis. This phosphorylation status was correlated to sensitivity of AXL inhibitors, proving this approach for treatment selection. On a side note, the finding of phosphotyrosine heterogeneity complicates a single drug regimen selection for these patients and underlines the difficulty of targeting this disease.
The phosphoproteomic approach can be used for identification of signaling that can explain some of the aggressive PDAC traits. By identifying an aberrant phosphosite of the kinase SGK223 via tyrosine screening, Tactatan et al. 103 explored the function of this kinase, which turned out important for STAT3 transcription and invasion and migration. Importantly, as discussed before, PDAC is a multicellular disease with interaction of stromal cells and tumor cells, which influences signaling.
Tape et al. 104 have expanded our knowledge of this interaction by analyzing different signaling events upon co-cultures. They identified that only 7% of the tumor signaling is regulated by KRAS activation.
Interestingly, the phosphoproteome was influenced on a similar level by CAF interaction as by KRAS, highlighting the influence of the microenvironment on PDAC. One of these stimulatory events is activation of the IGF1R/AXL-AKT axis. This crosstalk signaling can stimulate the tumor cells on a different level than tumor-tumor interaction. Future research of multi-cellular systems or whole tumors will elucidate the importance of activated pathways by the microenvironment and will guide towards new targeted therapies.

Future perspectives and concluding remarks
During the last decade, large-scale omic approaches have greatly expanded our knowledge of PDAC genetics and the resulting tumor biology. Subtyping at the transcriptome level has identified poor prognosis subtypes characterized by the expression of mesenchymal genes and other programs that contribute to poor outcome 23,34,36 . Whole genome analyses of these tumors have proven that carcinogenesis is not always a sequential progression of mutations 20 . Even though numerous mutations have been revealed by whole-exome sequencing, no commonly mutated genes other than the known four driver genes (KRAS, P53, CDKN2A, SMAD4) have been identified, but the mutational status of PDAC does seem to cluster around certain pathways 21,22 . Future analyses identifying the concordance of these pathways in metastases and new studies targeting these pathways are needed to validate the importance of these findings. Hopefully, this will result in clinical translation and applications. This can be in the form of new therapeutic targets and/or stratification tools that could be used to improve the use of currently available treatment modalities.
Following our understanding of the genomic rearrangements and gene expression that drive these tumors, analyses at the protein level are a logical next step to help understand the disease and improve survival of these patients. Proteome analysis adds another level to known genomic and transcriptomic data since it identifies the functional players in cell biology. For example, phosphoproteomics can give detailed insight in key signaling pathways that drive growth of tumors that might appear similar at the genetic level ( Figure 2B). Large-scale profiling studies of wellcharacterized clinical cohorts are needed to reveal the proteome landscape of PDAC and exploit this functionally relevant information in a more comprehensive way than was possible to date.
At this moment, an integrated analysis of all levels of molecular profiling data is called for to improve our understanding of how they interact to contribute to the disease, and clinically compatible assays need to be developed in order to capitalize on these findings. For example, although transcriptomic subtypes show survival differences, the transition to clinical practice is not easy.. This is partly due to the number of genes in a classifier. Adding a proteomic analysis to matched samples of subtypes could identify the most differential proteins, which can be evaluated by standardized immunohistochemical techniques in pathology laboratories. Additionally, even though many analyses were performed with one particular technique, all the levels of these omics approaches are interwoven and influence each other. The connectivity of these data should be used to create multilevel biological networks that explain the acquired aggressive capacities more faithfully ( Figure 2B).
So far, treatment of PDAC patients with advanced disease still relies on cytotoxic agents 105 . Some survival improvement has been established during the last decade. Especially, neoadjuvant therapy can possibly make a difference in the treatment of this disease. Preliminary retrospective studies show improved survival upon multi-regimen treatment preoperatively 106 . This improvement is in line with a computational analysis that calculated the effect of inhibition of proliferation to have a bigger impact on survival than just tumor bulk reduction by surgery 3 . One of the presumed mechanisms underlying the benefit of the neoadjuvant approach is the avoidance of development of genetic heterogeneity of subclones and its associated resistance, with progression of the cancer and dissemination of different metastases. Large randomized clinical trials are needed to show the survival benefit of neoadjuvant treatment, preferably in collaboration with biomarker discovery studies to understand resistance and improve patient selection.
Another future treatment perspective will be immunotherapy. As discussed before, exosomes can influence the metastatic niche and influence local immune response against PDAC cells 93 . PDAC has immune-evasive capacities but certain genetic subtypes do initiate a more immunogenic response, which is translated from the primary site to the metastatic tumor 31 . Interestingly, Steele et al 107 identified CXCR2 as an immune modulator which after inhibition induced an enhanced T-cell response and reduced metastatic burden. The promising data of patients with mismatch-repair deficient colorectal cancer responding to immunotherapy might also be relevant for a small number of patients with pancreatic cancer 108 . These studies indicate that next to tumor targeting, the microenvironment and immune response of primary, as well as metastatic PDAC, will need to be further explored to change the prognosis of these patients.
Future studies, including proteomics studies, should also identify novel biomarkers in order to select the group of patients who may gain the most benefit of cancer immunotherapy, as well as implement the design of novel clinical trials designs that allow tumor sample collection in order to understand the mechanism of action and resistance of PDAC (and its metastasis) to (immune)therapy. Biomarker-based selective clinical trials for targeted therapy are indeed incorporated into many ongoing trials, raising hope that future studies and treatments can be given more efficiently. A new clinical study by the Pancreatic Cancer Action Network will make use of molecular profiling of PDAC patients to select patients for specific tracks in their clinical trial 109 .
Moreover, in the near future readout of aberrantly activated kinases identified by phosphoproteomics will hopefully aid in the stratification of patients for targeted therapy.
Finally, in a possible and desirable future, with the availability of genome/proteome-wide screening platforms at reasonable costs, a thorough omic analysis of both the tumor and the metastastic specimens in conjunction with user-friendly computational tools will help clinicians to identify the most appropriate drug regimen to be administered to the patient. Hopefully, this approach will become a strategic companion for patient stratification and optimization of currently available cytotoxic treatments as well as novel anticancer drugs in clinical development.

Conflict of interest
The authors declare that there are no conflicts of interest.