A Unified Transcriptional, Pharmacogenomic, and Gene Dependency Approach to Decipher the Biology, Diagnostic Markers, and Therapeutic Targets Associated with Prostate Cancer Metastasis

Simple Summary This manuscript demonstrates how integrated bioinformatic and statistical reanalysis of publicly available genomic datasets can be utilized to identify molecular pathways and biomarkers that may be clinically relevant to metastatic prostate cancer (mPrCa) progression. The most notable observation is that the transition from primary prostate cancer to mPrCa is characterized by upregulation of processes associated with DNA replication, metastasis, and events regulated by the serine/threonine kinase PLK1. Moreover, our analysis also identified over-expressed genes that may be exploited for potential targeted therapeutics and minimally invasive diagnostics and monitoring of mPrCa. The primary data analyzed were two transcriptional datasets for tissues derived from normal prostate, primary prostate cancer, and mPrCa. Also incorporated in the analysis were the transcriptional, gene dependency, and drug response data for hundreds of cell lines, including those derived from prostate cancer tissues. Abstract Our understanding of metastatic prostate cancer (mPrCa) has dramatically advanced during the genomics era. Nonetheless, many aspects of the disease may still be uncovered through reanalysis of public datasets. We integrated the expression datasets for 209 PrCa tissues (metastasis, primary, normal) with expression, gene dependency (GD) (from CRISPR/cas9 screen), and drug viability data for hundreds of cancer lines (including PrCa). Comparative statistical and pathways analyses and functional annotations (available inhibitors, protein localization) revealed relevant pathways and potential (and previously reported) protein markers for minimally invasive mPrCa diagnostics. The transition from localized to mPrCa involved the upregulation of DNA replication, mitosis, and PLK1-mediated events. Genes highly upregulated in mPrCa and with very high average GD (~1) are potential therapeutic targets. We showed that fostamatinib (which can target PLK1 and other over-expressed serine/threonine kinases such as AURKA, MELK, NEK2, and TTK) is more active against cancer lines with more pronounced signatures of invasion (e.g., extracellular matrix organization/degradation). Furthermore, we identified surface-bound (e.g., ADAM15, CD276, ABCC5, CD36, NRP1, SCARB1) and likely secreted proteins (e.g., APLN, ANGPT2, CTHRC1, ADAM12) that are potential mPrCa diagnostic markers. Overall, we demonstrated that comprehensive analyses of public genomics data could reveal potentially clinically relevant information regarding mPrCa.


Introduction
Prostate cancer (PrCa) is the third most common cancer in the world, with a global incidence of 1,276,106 (7.1%) and mortality of 358,989 (3.8%), according to 2018 reports [1]. Among men, PrCa is the most commonly diagnosed and deadliest in 105 and 46 countries, respectively. Mortality rates are notably higher in Sub-Saharan Africa, the Caribbean, and African Americans [2]. cover information that may eventually be relevant to diagnostics and treatment of mPrCa. The method we employed involved the integration of publicly available transcriptional, pharmacological, and genetic dependency datasets for prostate tissue (metastasis, primary tumors, normal), as well as cancer cell lines (including PrCa). The genetic dependency dataset was generated from genome-wide CRISPR (clustered regularly interspaced short palindromic repeats) knockout studies. Results described in this manuscript include the identification of potential diagnostic markers (e.g., secreted proteins for ELISA-based assays, surface protein markers that can serve as targets of radiolabeled antibodies) and therapeutic targets for metastatic PrCa, as well as prediction of molecular and signaling pathways that may drive the PrCa metastatic process.

Public Genomic and Pharmacological Datasets
Several publicly available genomic and pharmacological datasets (further described in Table S1) were analyzed for this manuscript.

Pathways Analyses
The prediction of associated molecular pathways was accomplished using: (a) Gene Set Enrichment Analysis (GSEA) software available through the Broad Institute website (www.broadinstitute.org/gsea/, accessed on 15 May 2021) [28]. GSEA starts with the recognition that genes associate in particular groups (or gene sets), representing pathways and functionalities, such as those defined in Biocarta (http://software.broadinstitute. org/gsea/msigdb/genesets.jsp?collection=CP:BIOCARTA/, accessed on 15 May 2021), Reactome (https://reactome.org/, accessed on 15 May 2021) [29], KEGG (https://www. genome.jp/kegg/, accessed on 15 May 2021), and Hallmark [30] and (b) Reactome overrepresentation analysis. A more straightforward analysis of identifying the pathways associated with a given gene was conducted via the Reactome website. The gene identifiers for a select subset of genes were entered into the Reactome analysis entry box in this analysis. The built-in program then generates a list of over-represented pathways, along with the following values for each Reactome pathway (R): (a) the number of identifiers (or genes) submitted (or found) (F) in the analysis; (b) the total (T) number of genes curated to belong to pathway R; (c) the associated probability score (P), calculated using Binomial Test; and (d) false discovery rate (FDR) which estimates the false positives through the Benjamini-Hochberg procedure [31].
The genes CENPF, TPX2, EXO1, HJURP, and TOP2A, play pivotal roles in chromosome segregation, mitosis, and DNA replication. PLK1 is a serine/threonine kinase gene involved in mitotic regulation [32]. The top 500 genes with the highest signal-to-noise ratio (SNR; metastasis vs. PT) were subjected to Reactome over-representation analysis. Results indicated that the pathways with the lowest P score (i.e., most likely to be enriched) pertain to mitosis, cell cycle, cell cycle regulation, DNA replication, chromosomal segregation, and PLK-mediated events ( Figure 2A). Similar results were obtained by applying GSEA analysis, which, unlike the Reactome over-representation analysis, the entire dataset (i.e., all the genes) was used as input. The program was run to evaluate the enrichment of molecular signatures/pathways comprising the Biocarta, Reactome, KEGG, and Hallmark databases as described in the MSigDB website (https://www.gsea-msigdb.org, accessed on 15 May 2021). The Reactome pathways which registered the highest Enrichment Scores  Figure 2B (plus more comprehensive lists in Table S3) are the enrichment plots of some of the above pathways/gene sets. The enrichment of the Reactome pathway "Unwinding of DNA" can be explained by high (and highly ranked) SNR values for several of its component genes such as GINS1, MCM4, and MCM6. These genes are considered the "core enrichment genes" for this gene set. The core enrichment genes for the Hallmark pathway "E2F Targets" include TOP2A, MELK, MKI67, CDK1, DLG, and AURKA.

PLK1-Driven Mitotic Events Are Likely Upregulated in Prostate Cancer Metastasis
As illustrated in Figure 3, the components of the Reactome pathway PLK1-related events are predominantly upregulated in PrCa metastasis relative to PT (but not in PT relative to normal prostate tissues). In this signaling pathway, a PLK1 phosphorylated at threonine 210 (or P-T210) will activate the phosphatase CDC25C, which will then translocate to the nucleus. In the nucleus, the activated CDC25C will activate the cyclin B1/CDK1 complex, which will promote early mitotic events [44]. In contrast, PLK1's phosphorylation of the CDK1-inhibitor WEE1 may serve as the latter's signal for degradation [45]. In addition, PLK1 (P-T210) can also phosphorylate the transcription factor FOXM1, which will then upregulate the expression of various genes involved in G2 to M transition (CCNB1, CCNB2, CENPF, CDC25A, and PLK1). As shown in Figure 3, the elevated expression of the genes mentioned above is consistent with a more active PLK1-driven signaling pathway that eventually leads to an increased mitotic rate, which is likely what happens in metastasizing prostate cancer cells [32]. One exception is WEE1, whose expression is lower in metastatic compared to PT tissues. It makes sense since WEE1 has a mitotic inhibitory function, as explained above.

Genes for Secreted Proteins Are Also Upregulated in Metastatic Prostate Cancer
The PSA test's non-invasive and easily accessible nature is what made it a very popular early detection test for PrCa. However, its reliability is questionable because of the low specificity (high false-positive rate) resulting from the test [9][10][11]. This prompted the search for alternative ELISA-based tests to detect more reliable serum-or urine-based markers [62,63]. Recently proposed is detecting two glycoproteins (thrombospondin 1 or THBS1, and cathepsin D or CTSD) in the blood. We aimed to identify potential PrCa metastatic-specific, secreted protein markers by simply asking which of the most highly upregulated mRNAs (metastasis relative to PT) are also most likely to be translated to secretable proteins. We assume that the expressed protein is likely secreted if it passes either of the following filters: (a) the proteins are tagged as "secreted (curated)", "secreted (highly likely)", or "secreted (likely)", based on information derived from MetazSecKB [25], or (b) the extracellular location is "predicted to be secreted", according to information derived from the Human Protein Atlas [23]. As shown in Table 2, the metastatic-specific genes that may code for such proteins include C12orf49, ESM1, APLN, FNDC1, EDA, ANGPT2, PDGFB, and STC2. The expression profiles of APLN (apelin) and ANGPT2 (angiopoietin 2) are illustrated in Figure 1D. Both genes have been reported as potential serum-based markers for metastatic colorectal cancer [64,65]. ANGPT2 also proved to be a serum marker of poor prognosis in lung cancer [66]. An elevated level of APLN in serum correlated with esophageal squamous cell carcinoma metastasis [67]. Other secreted protein markers, listed in Table 2, that have been demonstrated as indicators of cancer are CTHRC1 (metastatic colon cancer) [68], ESM1 (metastatic colon cancer) [69], ADAM12 (advanced stage prostate cancer) [70], PDGFB (oral squamous cell carcinoma) [71], and STC2 (laryngeal squamous cell carcinoma) [72,73].

Potential Metastatic Therapeutic Targets (Such as PLK1, INCENP) Also Exhibit High Genetic Dependency
In Project Achilles, >800 cancer cell lines (n = 808, according to the 20Q4 data release) were subjected to a genome-wide CRISPR/Cas9 knockout screen [20,21]. The resulting sgRNA sequencing and cell viability data were then used to calculate the probability (P) that the knockout of a given gene (G) will affect the viability of a particular cell line (C). The P GC score (also referred to as "gene dependency" or GD) for 18,119 genes ranged from 0 (i.e., gene knockout did not influence cell viability) to 1 (i.e., gene expression is very vital to cell viability). The average GD scores (separately for 368 PT and 253 metastatic cell lines) for each of the top 300 PrCa metastasis-upregulated can be found in Table S2. As exhibited in Figure 4, the average GD scores for the potential PrCa metastasis diagnostic markers (e.g., the surfaceome genes LRFB1, NUP210, ABCC5, and NRP1, as well as the secretome genes VASH, ANGPT2, LPL, and EDA) are closer to 0 than they are to 1. It is also noteworthy that the average expression levels of the genes mentioned above are higher among metastatic compared to PT-derived PrCa lines. In contrast, the average GD scores for INCENP, PLK1, and CDK1 are close to 1. The expression levels of these genes are similarly higher in metastatic relative to PT PrCa cell lines. However, for the gene AR, a high gene dependency value (close to 1) is only evident in the PrCA line VCap, reflective of the gene's unique role in PrCa progression.

A Tyrosine Kinase Inhibitor (Which Targets PLK1, AURKA, MELK) Exhibits Higher Efficacy against Cancer Cell Lines of Metastatic Origin
In the PRISM drug repurposing project, pools of 468 molecularly barcoded cancer cell lines were treated with 4686 drugs (majority of which have been approved for diseases other than cancer) (information taken from the 19q4 version of the public dataset) [22]. The abundance of these barcodes (relative to cells treated with DMSO) served as a measure of change in the viability of the cell lines post drug treatment. Given that there is a great deal of overlap (i.e., cell lines) between PRISM and CCLE molecular profiling datasets, it is theoretically possible to identify potential predictive or resistance markers for many of the drugs included in the PRISM project. As mentioned above, we are particularly interested in the drug fostamatinib, which targets a family of kinases including PLK1. Both genome-wide transcriptional and fostamatinib viability data are available for 464 cell lines. We arbitrarily divided the cell lines into two subgroups: (a) Group A includes cell lines that were "responsive to fostamatinib" (i.e., log fold change viability ≤ −0.5; n = 193), and (b) Group B covers those which were "non-responsive to fostamatinib" (i.e., log fold change between −0.5 and 0.5; n = 271). We then identified the highly differentiated genes between the two groups. As shown in Figure 5A (and Table S4), the upregulated genes in Group A include COL24A1, COL7A1, and many other genes related to invasion processes. Indeed, when the top 150 of such genes were subjected to Reactome analysis, we observed that the most highly dysregulated pathways (in Group A relative to Group B) are related to invasion as well as degradation of ECM, molecular pathways that are definitive signatures of metastasis ( Figure 5B, Table S5). These pathways include "assembly of collagen fibrils and other multimeric structures", "crosslinking of collagen fibrils", "collagen formation", "collagen chain trimerization", "interleukin-4, and interleukin-13 signaling", "anchoring fibril formation", "elastic fiber formation", "ECM proteoglycans", "collagen biosynthesis and modifying enzymes", "collagen degradation", "extracellular matrix organization", "degradation of the extracellular matrix", "platelet degranulation", "molecules associated with elastic fibers", "MET activation of PTK2 signaling", and "the RND3 GTPase cycle".
In essence, what these results suggest is that cell lines exhibiting signatures related to invasion and metastasis seem to be more responsive to inhibition of kinases such as PLK1, CDK1, MELK, and NEK.

Discussion
In recent years, the integration and reanalysis of various publicly available genomic, epigenomic, proteomic, and metabolomic datasets have been a significant source of discoveries in different areas of cancer research: from basic cancer biology to translational studies (e.g., molecular diagnostic markers, therapeutic targets). This is not surprising given the enormity of information buried in those datasets. In this report, we demonstrate the possibility of predicting the molecular pathways, non-invasive diagnostic markers, and molecular drug targets associated with the metastatic progression of prostate cancer by merely integrating various genome-wide transcriptional, gene-dependency, and pharmacological datasets.
Based on our results, it is clear that the progression from localized PrCa PT to metastasis is defined by elevated expression of many genes involved in cell division, cell cycle regulation, and DNA replication and repair process. These genes include TPX2 (TPX2 microtubule nucleation factor), PLK1 (polo-like kinase 1), ANLN (anillin actin-binding protein), EXO1 (exonuclease 1), PRC1 (protein regulator of cytokinesis 1), KIF20A (kinesin family member 20A), POC1A (POC1 centriolar protein A), CENPF (centromere protein F), HJURP (Holliday junction recognition protein), MCM2 and MCM4 (minichromosome maintenance complex components 2 and 4), and TOP2A (DNA topoisomerase II alpha) (See Table S2). A more comprehensive pathways analysis (GSEA) or identification of common pathways/functionality signatures of genes highly upregulated in PrCa metastasis (through Reactome) would also reveal that pathways such as "unwinding of DNA", "DNA replication", "PLK-mediated events", and "cell cycle checkpoints" are relatively enriched in PrCa metastasis samples. These observations are consistent with a recent report by Hsu and colleagues [98], wherein the transcription levels of MCM genes 2,3,4, and 6, which code for components of a complex involved in genome replication initiation, are elevated in Neuroendocrine PrCa (NEPC). Moreover, the inhibition of the MCM2-7 complex (by the drug ciprofloxacin) reduced cell proliferation and migration in vitro. Kauffmann and colleagues [99] reported similar observations regarding metastatic melanoma. The authors posited that a more active DNA replication and repair machinery enable metastatic melanoma to circumvent DNA damages caused by chemo-and radiation therapy.
The elevated expression of PLK1 in mPrCa may be tied to its prominent regulatory role in mitosis. A phosphorylated PLK1 can phosphorylate (and activate) the phosphatase CDC25C, the CDK1-inhibitor WEE1, and the transcription factor FOXM1. The downstream targets of these activated proteins are other proteins that regulate the G2 to M transition of cancer cells [32,44]. In addition, a recent study has shown that phosphorylation by PLK1 is also necessary to suppress the proapoptotic activity of the transcription factor FOXO1 (forkhead box protein O1) in PrCa cells [100]. Hence, targeting PLK1 by a drug can potentially inhibit or slow down PrCa's (or any other cancer type's) metastatic potential. This was recently demonstrated by Montaudon and colleagues in which the size of an ERpositive BrCa patient's PT and bone metastasis-derived PDX (patient-derived xenograft) rapidly shrunk after treatment with volasertib, an inhibitor of the PLK1 (which along with AURKA and CDK1 were upregulated in the PDX) [77]. Shin and colleagues have also demonstrated that PLK1 is a target for the flavonoid genistein and that the drug was found to be selective against TP53-mutated cell lines [101]. In another recent study, fostamatinib (which inhibits PLK1 as well as other serine-threonine kinases) was shown to be effective against the prostate cancer cell line (PC3) [102]. The anti-cancer activity of fostamatinib was also evident against head and neck squamous cell carcinoma [103], hepatocellular carcinoma [104], breast cancer [105], and diffuse large B cell lymphoma [106] cell lines. Moreover, a fostamatinib derivative, NSC765691, also exhibited antiproliferative activity against the panel of NCI-60 cell lines [107]. The drug was also shown to have significant clinical activity when treating non-Hodgkin lymphoma and chronic lymphocytic leukemia patients [108]. Among the ongoing clinical trials (source: clinicaltrials.gov) which involve fostamatinib are NCT05030675 (Phase I; against lower-risk myelodysplastic syndromes or chronic myelomonocytic leukemia who have failed hypomethylating agents) and NCT03246074 (Phase I; combined with paclitaxel, against recurrent ovarian, fallopian tube, or primary peritoneal cancer).
In this current report, we also evaluated the transcriptional signatures that may be indicative of fostamatinib's antiproliferative activity in cancer cell lines. Results indicated that fostamatinib-responsive cell lines exhibited relatively higher expression of genes belonging to the family of fibrillar and fibrillar-like collagens (COL24A1, COL6A2, COL1A1, COL1A2, COL5A1, and COL6A3). Collagens are the most abundant proteins in the ECM and provide the bulk of mechanical strength that drives cell migration [109][110][111][112]. Other genes whose expression is higher among fostamatinib-responsive cell lines are the fibrillin gene FBN1, the bone morphogenetic protein 1 (BMP1), lysyl oxidase-like 2 (LOXL2), the integrin genes ITGB1 and ITGA5, the adamlysin gene ADAM12, and the growth factor genes PDGFC and TGFB1. Fibrillins are microfibrillar proteins that are also components of ECM. Integrins are heterodimer cell surface receptors utilized in downstream signaling from the ECM. The metalloprotease BMP1 cleaves the collagen precursor's carboxy terminus, a necessary step in matrix assembly. Lysyl oxidases are enzymes needed for crosslinking collagen and elastin molecules in the ECM. Adamlysins are endopeptidases whose ability to degrade the matrix during ECM remodeling also allows cell migration during metastasis. Predictably, the results of our Reactome analysis indicated that the fostamatinib-responsive cell lines are characterized by enhanced signatures of pathways such as "assembly of collagen fibrils and other multimeric structures", "extracellular matrix organization", "anchoring fibril formation", "crosslinking of collagen fibrils", and "collagen degradation". Overall, these observations point to the possibility that inhibitors to PLK1 (and related kinases) may help suppress prostate cancer metastasis.
We were able to identify genes coding for cell surface-bound proteins, which can potentially be explored as targets for radiolabeled monoclonal antibodies for positron emission tomography (PET)-based detection of metastatic prostate cancer. These markers include ADAM15 [48], CD276 [49], NRP1 [52,53], SCARB1 [54], and PLXNA3 [56], all of which have been reported to be overexpressed in metastatic PrCa. Elevated expression of genes such as ABCC5 [50], LRFN1 [59], ELOVL6 [58], and HTR2B [61] have been associated with metastasis in other cancer types. Recently, PET-based detection and monitoring of metastasis cancer has utilized the following antibodies: 111 In-labeled anti-CDH17 (gastric cancer) [114], 177 Lu-labeled anti-CD55 (lung cancer) [115], and radio-labeled anti-ERBB2 (various labeling, including 89 Zr, 64 Cu, 111 In) (breast cancer) [116]. The gene FOLH1 (folate hydrolase 1) is of particular interest since it codes for the transmembrane metalloenzyme PSMA (prostate-specific membrane antigen). PSMA is the target for an FDA-approved 68 Ga-based peptidomimetic radiotracer for PET imaging of PrCa [117]. Although FOLH1 is not included in Table 1 or Table S2, the gene's transcriptional upregulation is significant for both PrCa primary tumors (fold change and SNR relative to normal prostate are 1.42 and 0.20, respectively), and PrCa metastasis (fold change and SNR relative to primary tumors are 1.89 and 0.30, respectively).
The popular but very controversial PSA test is an ELISA-based test for the presence of PSA protein (coded by the gene KLK3) in serum and is intended for early detection of PrCa. Tests to detect the presence of proteins THBS1 (thrombospondin 1) and CTSD (cathepsin D) are among those being proposed as alternatives to the PSA test [63]. A noninvasive detection or monitoring of metastasis by interrogating specific proteins in patient serum (or urine) may also be feasible and backed by many publications. Several PrCa metastasis-upregulated proteins predicted to be part of the secretome have been proved experimentally as potential markers for ELISA assays. These include the proteins APLN (apelin) [64,67], ANGPT2 (angiopoietin 2) [66], CTHRC1 (collagen triple helix repeat containing 1) [68], ESM1(endothelial cell-specific molecule 1) [69], ADAM12 (ADAM metallopeptidase domain 12) [70], PDGFB (platelet-derived growth factor subunit B) [71], and STC2 (stanniocalcin 2) [72,73]. It will not be surprising if more proteins listed in Table 2 may also prove good candidates for serum-or even urine-based tests for PrCa metastasis detection and monitoring. Nonetheless, it should be pointed out that more studies are needed to ascertain the clinical utilities of these secreted proteins as diagnostic markers for mPrCa.
Apart from PLK1 (and the related serine/threonine kinases), our analysis identified a relatively long list of proteins whose inhibition can potentially (or, in theory) repress PrCa metastatic potential. It is encouraging to know that inhibitors already exist for many of these proteins, some of them FDA-approved for diseases other than cancer. Recent reports have demonstrated that inhibition of some of these proteins can potentially hinder metastasis. For example, the inhibition of the protein INCENP (by the drug reversine) led to a reduction of migration potential of colon cancer cells [118] and cell motility and invasion potential of breast cancer cells [119]. Another potential target is SSTR1(somatostatin receptor 1). The drug pasireotide, which targets this protein, exhibited efficacy against mCRPC [120] and metastatic carcinoid disease [121]. Targeting the protein BIRC5 (by the drug berberine) reduced the metastatic ability of PrCa cells [122].
The genome-wide CRISPR-generated gene dependency (GD) data incorporated in our analyses provided vital information on how a given gene's inactivation affects cancer cells' survival. As indicated in Table S1, PLK1 and its related kinases (AURKA, CDK1, MELK, NEK2) have GD values very close to 1 across all cell lines (irrespective of the type of cancer or whether it is PT or metastatic origin), signifying that the protein products are essential to the survival of cancer cells, thus ideal therapeutic targets. Other genes that exhibited high GD values include INCENP, TPX2, PRC1, TOP2A, MCM2, and MCM4. The genes mentioned above are part of cells' DNA replication and cell division machinery. An example of a PrCa-upregulated gene with a very low (close to zero) GD value is MKI67 (a marker of proliferation Ki-67), which codes for a nuclear protein that has become a well-studied immunohistochemical marker of cancer proliferation [123]. The low GD value for MKI67 indicates that it is not an ideal therapeutic target despite being a proven marker of proliferation. Indeed, there is experimental evidence proving that altering MKI67 does not significantly affect proliferation [124]. Oher genes that surprisingly have low GD values are SSTR1, ABCC5, PLXNA3, EZH2, and LRFN1.

Conclusions
Although a bioinformatic exercise, this report stemmed from meticulous analyses of publicly available genomic and pharmacological data from >200 tissues and >1000 cell lines. Overall, we both validated previously reported observations and presented new and interesting observations regarding the biology, diagnostics, and molecular targeting of metastatic prostate cancer. These bioinformatic observations may also serve as a springboard for a wide array of experimental validations.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/cancers13205158/s1, Table S1: List of publicly available datasets re-analyzed in the study, Table S2: Top 300 most highly upregulated genes (Mets relative to PT), Table S3: Results of GSEA Analysis (prostate cancer metastasis vs. primary tumors), Table S4: Expression levels (across all cell lines) of top 150 genes exhibiting the highest signal-to-noise ratios ("fostamatinib responsive" vs. "fostamatinib non-responsive), Table S5: The resulting top 20 Reactome pathways (the ones with the lowest Entities p values) when the top 150 PrCa metastasis-upregulated genes were used as input in the Reactome analysis.

Data Availability Statement:
The datasets used in the study were downloaded from GEO and DepMap websites. Links to the datasets are listed in Table S1.

Conflicts of Interest:
The authors declare no conflict of interest.