Phosphoproteomics Enables Molecular Subtyping and Nomination of Kinase Candidates for Individual Patients of Diffuse-Type Gastric Cancer

Summary

The diffuse-type gastric cancer (DGC) constitutes a subgroup of gastric cancer with poor prognosis and no effective molecular therapies. Here, we report a phosphoproteomic landscape of DGC derived from 83 tumors together with their nearby tissues. Based on phosphorylation, DGC could be classified into three molecular subtypes with distinct overall survival (OS) and chemosensitivity. We identified 16 kinases whose activities were associated with poor OS. These activated kinases covered several cancer hallmark pathways, with the MTOR signaling network being the most frequently activated. We proposed a patient-specific strategy based on the hierarchy of clinically actionable kinases for prioritization of kinases for further clinical evaluation. Our global data analysis indicates that in addition to finding activated kinase pathways in DGC, large-scale phosphoproteomics could be used to classify DGCs into subtypes that are associated with distinct clinical outcomes as well as nomination of kinase targets that may be inhibited for cancer treatments.

Graphical Abstract

Highlights

• •A phosphoproteomic landscape of diffuse-type gastric cancer (DGC) was depicted
• •DGC could be classified into three subtypes based on phosphorylation data
• •A bioinformatics workflow was used to identify 16 kinases as potential drug targets
• •A patient-specific strategy for nomination of kinases was proposed

Biological Sciences; Cancer Systems Biology; Proteomics; Systems Biology

Introduction

Gastric cancer (GC), one of the most common and fatal diseases in East Asian, is a heterogeneous disease with diverse histological and molecular characteristics (Wang et al., 2011, Cancer Genome Atlas Research Network, 2014). The widely used Lauren classification stratifies GC into diffuse-type, intestinal-type, and mixed types (Wu et al., 2019). Diffuse-type gastric cancer (DGC) has the worst prognosis and lacks treatment options, particularly for targeted therapy.

Over the past decades, genomic landscapes of GC has been mapped (Cancer Genome Atlas Research Network, 2014, Tan et al., 2011, Cristescu et al., 2015). The Cancer Genome Atlas (TCGA) project uncovered four molecular subtypes of gastric cancer. The Asian Cancer Research Group (ACRG) also described four molecular subtypes based on gene expression data of GC (Cancer Genome Atlas Research Network, 2014). Previously, we mapped the proteomic landscape of DGC of 84 paired tumors and their nearby tissues and showed that based on the altered protein expression alone, DGC could be subtyped into three major classes (PX1-3) that are associated with clinical outcomes (Ge et al., 2018). Our study allowed the nomination of more than 20 proteins that function in cancer growth, ROS and metabolism, cell-cell adhesion and adjunction, as well as immune-response pathways as potential drug targets.

The rapid development of kinase inhibitors raised the hope for targeted therapy and even truly individualized therapy (Ferguson and Gray, 2018). The concurrent development of phosphoproteomics that focused on identification and quantification of phosphorylated amino acid residues in proteins in biological specimens provided a needed support for the utilization of kinase inhibition as a therapy (Casado et al., 2017, Wu et al., 2019). The state-of-the-art phosphoproteomics now can measure tens and thousands of phosphorylation sites, from which kinase activities can be inferred and targeted kinase therapies may be developed (Casado et al., 2017, Wu et al., 2019). Large-scale mappings of phosphorylation landscapes were carried out in several cancers including breast cancer (Mertins et al., 2016), ovarian cancer (Mertins et al., 2016), prostate cancer (Drake et al., 2016), lung cancer (Rikova et al., 2007), blastoma brain cancer (Liu et al., 2018), gastric cancer (Mun et al., 2019), and hepatocellular carcinoma (Jiang et al., 2019). But the utilization of the phosphoproteomics data was poor, and there were few reports of using phosphoproteomics to subtype cancers to predict patient prognosis or nominate kinase targets for treatments in a systematic manner.

Here, we report the phosphorylation landscape of DGC as a part of a concerted effort of the Chinese Human Proteome Project (CNHPP). We carried out a quantitative measurement of phosphoproteomics of 83 DGC tumors and their matching nearby tissues. Applying an unbiased bioinformatics analysis workflow that we developed recently (Tong et al., 2018), we showed that using phosphoproteomics data alone, DGC could be classified into three subtypes that associated with distinct clinical outcomes and we nominated druggable kinase candidates for each individual patient, allowing for drug prioritization on an individualized basis for DGC patients.

Results

General Features of the DGC Phosphoproteome

We enriched phosphopeptides with TiO2-coupled beads from tryptic digests of tumor samples and their nearby tissues of 83 DGC patients obtained from the tumor tissue bank of Beijing Cancer Hospital (Ge et al., 2018) (Table S1) and measured enriched phosphopeptides with LC-MS/MS. Phosphorylation was identified through database using a data analysis platform Firmiana (Feng et al., 2017) (Figure 1A). After quality control including the false discovery rate (FDR) of peptides and Mascot delta score (Wu et al., 2011) for phosphorylation sites (Figure 1B), 28,016 phosphorylation sites were used for subsequent analysis (Figure S1, Table S2), which include 22,744 (81.18%) phosphoserine sites, 4,889 (17.45%) phosphothreonine sites, and 383 (1.37%) phosphotyrosine sites (Figure 1C).

Figure 1.

A Summary of Phosphoproteome Analysis of Diffuse-Type Gastric Cancer

(A) General workflow of the phosphopeptide enrichment and quantitative mass spectrometry protocol.

(B)Phosphoproteomic datasets filtered at different levels for various statistical analyses.

(C) Distribution of modification types.

(D) Principal component analysis (PCA) to visualize tumor and nearby tissue samples.

(E) Top ranked pathways that are significantly altered in tumors compared with nearby tissues (FDR<0.1, the minimum proteins in one pathway is 5). Yellow words represent the pathways enriched by the proteins with upregulated phosphorylation sites in tumors compared with nearby tissues; blue words represent the pathways enriched by the proteins with downregulated phosphorylation sites in tumors compared with nearby tissues.

Also see Figure S1.

Among the 28,016 phosphorylation sites in the DGC phosphoproteomics, 21,282 sites were found in both tumors and nearby tissues; 4,447 sites and 2,287 sites were detected only in tumors and nearby tissues, respectively (Figure S1B). Principle component analysis (PCA) revealed that tumors could be separated from the nearby tissues based on phosphorylation profiles (Figure 1D). An SAM (significance analysis of microarray) analysis (Tusher et al., 2001) identified 445 upregulated and 819 downregulated phosphorylation sites in tumors compared with nearby tissues (FDR <0.01 by SAM and differential expression percentage >0.5 or <−0.5, Figure S2A, Table S3). Proteins with upregulated phosphorylation in tumors were enriched in cell-cycle-related pathways (DNA replication and cell division), cell-cell adhesion, DNA repair, and mRNA splicing pathways (Figure 1E, Table S4), whereas proteins with upregulated phosphorylation in the nearby tissues were enriched in cell-cell adhesion, gastric acid secretion, and regulation of Rho protein signal transduction.

Subtypes of DGC and Their Associations with OS and Chemosensitivity

Based on the intensity of 28,016 phosphorylation sites in the tumors (Figure 1B, Table S2), we employed consensus clustering (Wilkerson and Hayes, 2010) to identify DGC subtypes. Three clusters (Ph1-3) were apparent (Figures S2B, S2E, and S3A). We further identified 302 differentially phosphorylated sites in tumors among the subtypes (Anova, FDR<0.001, Table S4). As shown in Figure 2A, hierarchical clustering revealed four groups of phosphorylated sites (Tables S5 and S6). To investigate subtype-specific pathway alterations, we further identified significantly altered phosphorylation sites between tumors and nearby tissues within the same subtype (FDR <0.01 by SAM and differential expression percentage >0.5 or <–0.5, Table S7). Ph1 to Ph3 contained 28, 133, and 5 uniquely upregulated phosphorylation sites in tumors, respectively, and many phosphorylation sites were downregulated in tumor (Figure 2B). These dysregulated phosphoproteins suggested that the Ph1 subtype was more or less normal in the basic function of the stomach while upregulated rRNA processing and RNA polymerase II promoter activity (Figures 2C and 2D, Table S7). The Ph2 subtype mainly upregulated DNA metabolic process and DNA repair while losing the basic function of the stomach including gastric acid secretion (Figures 2C and 2D, Table S7). The Ph3 subtype upregulated chromosome segregation and mainly lost cell-cell interaction and communications (Figures 2C and 2D, Table S7)

Figure 2.

Phosphoproteome Subtyping of DGC with Different Overall Survival and Chemosensitivity

(A) 301 sites differentially phosphorylated among the subtypes (Anova, FDR<0.001).

(B) The number of differentially expressed phosphorylation sites in tumors compared with nearby tissues for each type.

(C and D) Heatmaps of selected phosphoproteins representing major altered signaling pathways in each type.

(E) The association of molecular subtypes with overall survival of patients; Kaplan-Meier analysis, p value from logrank test.

(F) The association of adjuvant chemotherapy with overall survival in each subtype. Only the patients who have the chemotherapy information were shown.

(G–I) Other clinical parameters across 83 patients in Ph1–3. Intratumoral TILs, intratumoral tumor-infiltrating lymphocytes.

Also see Figures S2 and S3.

We further investigated the OS of these patients. The Ph1 subtype had the best OS, whereas both Ph2 and Ph3 had worse survival (logrank p=0.012, Figure 2E). Moreover, we found that the Ph2 patients tend to be more sensitive to chemotherapy (logrank p = 0.044, Figure 2F), but the Ph1 and Ph3 groups exhibited no statistically significant prognosis improvement by chemotherapy (logrank p> 0.1, Figure 2F). In addition, a multivariable Cox analysis showed that the phosphoproteomics subtyping remained significantly associated with patients' OS after adjusting for age, gender, adjuvant chemotherapy, tumor site, stage, and TP53 (using Ph1 as the reference, HR = 8.67, p = 0.038 for Ph2; HR = 9.87, p = 0.029 for Ph3, Table 1). Other clinical characteristics were also associated with the Ph1–3 subtypes. Firstly, age (≥50 years vs <50 years), stage (I/II/III/IV), and MSI status (microsatellite unstable vs microsatellite stable) in each subtype were significantly different (Table S8, Figure S3B, Fisher-test, p<0.05). The Ph1 subtype contained younger patients (<50 years), as well as early stage (stage II) patients, whereas Ph2 and Ph3 contained older and more advanced stage (stage III–IV) patients. Cancer driver gene mutation rate that we measured previously (Ge et al., 2018) seemed to be higher in Ph1 than in Ph2 and Ph3 (Figure 2G, p = 0.044, Wilcox test), for example, the Ph1 subtype contained three out of five patients with MSI. Secondly, Ph1 was enriched with higher intratumoral TILs (tumor-infiltrating lymphocytes) and mesenchymal cells than Ph2 and Ph3 (p<0.05, Wilcox test Figures 2H and 2I). It has been reported that increasing intratumoral TILs implies a better prognosis in GC (Kang et al., 2017, Yu et al., 2016, Grogg et al., 2003). These correlations were consistent with and may explain the OS difference between the three molecular subtypes. Previously, we subtyped DGC of the same samples into three subtypes (PX1-3) based on protein profiling alone (Ge et al., 2018). As showed in Figure S3B, most of the patients with worse survival in the Ph2 and Ph3 (85.25%) were also classified to PX2-3, demonstrating that the two methods reached good agreement. Moreover, Ph1 appeared to include more patients with better survival than PX1 did, suggesting that phosphoproteomics-based subtyping may be superior to that from protein profiling for correlation with OS.

Table 1.

Univariate and Multivariate Analysis of Overall Survival in 81 Patients

Characteristics (n)	Univariate Analysis		Multivariate Analysis
	HR (95% CI)	p Value	HR (95% CI)	p Value
Agea	1.033 (1.002–1.065)	0.036	1.039 (0.99–1.08)	0.054
Gender
Male (50)	1		0.68 (0.30–1.55)	0.36
Female (31)	0.662 (0.323–1.36)	0.26
Adjuvant chemotherapy			0.60 (0.23–1.58)	0.3
Without (17)	1
Withb (64)	0.439 (0.201–0.961)	0.04
Tumor site
Cardia, GEJ (19)	1
Body (32)	1.484 (0.569–3.870)	0.42	1.62 (0.53–4.98)	0.4
Antrum (26)	1.342 (0.487–3.696)	0.57	1.27 (0.38–4.27)	0.7
Clinical stagea
(Ib to IV)	1.82 (1.01–3.29)	0.048	2.09 (1.12–3.88)	0.02
TP53 mutation
Wild-type (45)	1
Mutant (36)	1.23 (0.60–2.51)	0.58	1.10 (0.53–2.31)	0.79
Phosphoproteome cluster
Ph1 (22)	1
Ph2 (32)	10.53 (1.39–79.64)	0.023	8.67 (1.12–67.01)	0.038
Ph3 (27)	11.64 (1.52–89.14)	0.018	9.86 (1.26–76.95)	0.029

GEJ, gastroesophageal junction; HR, hazard ratio; CI, confidence interval.

^aContinuous variable.

^bPatients proceed at least one cycle of adjuvant chemotherapy. Significant data are emphasized in bold.

Nomination of Kinases as Potential Therapeutic Targets

As the activity of a kinase can be inferred by the intensity of its substrates (Casado et al., 2013), we built a kinase-substrate dataset that contained 3,321 substrates and 250 kinases to find activated kinases from the measured substrates (Table S9, Figure 3A). Our DGC phosphoproteomes contained 1,896 protein substrates in the kinase-substrate database (see Methods, Table S9). For each kinase, the average number of substrate sites is 13 and the median number is 4 (Figure S4, Table S9). We then calculated the normalized value of p-site-FOT^TiO2/protein-FOT^profiling to correct for altered protein expression (Wu et al., 2011) and obtained a normalized ratio of p-site between tumor and nearby tissue for each patient. We used the average fold difference for all detected substrate sites for the same kinase as a measurement of the kinase activation/inhibition (Table S10). The highly activated kinases (with top ranked values of average fold difference and top frequencies detected in the patients) included PRKACA, CSNK2A1, CDK1/2, MAPK1/3, GSK3B, PRKCA, AKT1, CDK4, and CDK6 (Figure S5A). We also adapted three other methods including Z-test, kinase substrate enrichment analysis (KSEA), and the multiple linear regression (MLR) model to identify more activated kinases (Hernandez-Armenta et al., 2017) (Figure 3A, see Methods). We computed “kinase activity” using each of these four methods for every patient and generated a kinase-patient matrix, respectively (Figure 3A).

Figure 3.

Nominating Potential Druggable Kinases for DGC

(A) The workflow of the identification of druggable kinases.

(B) Survival curves of the four kinases and boxplots of the kinase substrates in the phosphoproteome and proteome data. Low (Kinase = 0)/high (Kinase = 1): patients with kinase activity lower/above than cut-off. For kinases were both identified by more than one kinase activity prediction method, only the smallest p value of the kinase was shown.

Also see Figures S4–S10.

In order to evaluate the accuracy of the four prediction methods, we calculated correlation of the predicted kinases activities with phosphorylation intensities of the kinase activation loops. Phosphorylation of the activation loop is often critical in regulating kinase activity in many cases (Nolen et al., 2004). We found that the kinase activities predicted by the four methods were positively correlated with the intensities of the 43 activation loop phosphorylation sites that we could measure (Figure 3A, panel 2; Figure S6). The Mean value, KSEA, and regression methods performed significantly better than the Z-test method (Wilcox test p<0.05; Figure 3A, panel 2) and were retained for subsequent analysis.

We then postulated that activated kinases might be therapeutic targets if their activities are associated with poor OS. We stratified patients into two groups according to high and low kinase activity. The cut-off value was individually determined by the lowest p value of OS according to the logrank test (Figure 3A, panel 3; Table S11). We identified 19 kinases whose activities were significantly associated with poor OS (10 from the mean value method, 7 from MLR, and 8 from KSEA) (Figure 3A, panel 4). At least two substrates were identified for each kinase with a fold change of >1.5 between tumor and the near-by tissues. Finally, 16 non-overlap kinases were found after controlling the logrank p value <0.05 at least in three continuous cut-offs, which was to reduce the false-positive rate of the survival analysis (Figures S6–S8). The higher activities of these kinases were significantly associated with poor OS (logrank p<0.05, HR>2, Figure 3B, Table 2). We thus nominated these 16 kinases as potential therapeutic kinase targets (PTKT). Using the same strategy, we could identify only two kinases, namely SRPK2 and MAPKAPK2, whose increase in protein abundance correlated with poor OS (Table S11, Figure S9). SRPK2 was the only kinase that was identified as PTKT based on activity or abundance. Therefore, using kinase activity could identify more PTKTs.

Table 2.

Kinases Associated with Overall Survival of DGC Patients

Kinases	Method	HR	p Valuea	#Patientsb	Summary
PI3K/AkT/mTOR pathway
PDPK1	KSEA	2.85	0.0031	24	AKT1 activation
MTOR	Regression	2.54	0.0091	32	Cell growth and metabolism
RPS6KB1	Regression	2.29	0.021	27	Cell growth and proliferation
RPS6KA3	Mean values	2.69	0.0051	29	Cell growth and proliferation
PKN1	Mean values	3.18	0.00136	33	Phosphorylated by PDPK1
STK11	KSEA	2.58	0.0084	34	A tumor suppressor
Cell cycle; apoptosis
CDK7	Mean values	2.66	0.012	15	Cell cycle
CSNK2A1	KSEA; regression	2.25; 2.45	0.028; 0.014	22; 21	Cell cycle; apoptosis
CSNK2A2	Mean values	2.39	0.016	22	Cell cycle; apoptosis
ATM	Mean values	3.3	0.0007	29	DNA damage
MAPK pathway
MAPK3	Mean values	2.67	0.0066	35	Proliferation, differentiation, and survival
MAP3K7	Regression	3.35	0.0012	15	Proliferation, differentiation, and survival
Cell adhesion and actin organization
PAK4	Regression	2.49	0.013	17	Actin organization and cell adhesion
Microtubule affinity
MARK2	Mean values; regression	3.22; 2.39	0.00096; 0.014	20; 27	Controls the stability of microtubules
RNA splicing
SRPK2	Regression	2.66	0.0090	16	Pre-mRNA splicing; cell cycle regulation and cell apoptosis
G protein-coupled receptor
GRK6	Mean values	2.75	0.011	47	G protein-coupled receptor kinase

^aLogrank p values.

^bNumber of patients with a kinase activated.

Major Dysregulated Kinase Pathways in DGC

The above 16 PTKTs could be classified into seven pathways: (1) MTOR signaling, (2) cell cycle and apoptosis, (3) MAPK signaling, (4) cell adhesion and actin organization, (5) microtubule affinity, (6) RNA splicing, and (7) G protein-coupled receptor (Figure 4A). The MTOR signaling pathway contained six PTKTs, including PDPK1, MTOR, RPS6KB1, RPS6KA3, PKN1, and STK11, representing the most frequently dysregulated pathway in DGC (Figure 4A). It was reported that overexpression of PDPK1 was associated with poor OS in gastric carcinoma (Gagliardi et al., 2018). The cell cycle pathway ranked the second highly dysregulated pathway including PLK1, CDK7, CDK4, and CDK6. Upregulation of PLK1 and CDK7 in gastric cancer correlated with poor OS (Otsu et al., 2016, Wang et al., 2016). Moreover, CDK4/6 inhibitors demonstrated significant activity against several solid tumors (Goel et al., 2017). Recently, it was reported that MTORC1-S6K1 and CK1 phosphorylate SRPK2 to induce its nuclear translocation and turn on a splicing program to activate lipid metabolism to fuel cancer growth (Lee et al., 2017). Our findings that the activities of MTOR as well as its downstream kinases RPS6KB1/RPS6KA3 and SRPK2 were associated with poor OS (Figure 3B) strongly support the notion that the MTOR-S6K1-SRPK2 signaling is at work in DGC tumorigenesis. SRPK2 was reported to be overexpressed in several cancer types (Lee and Abdel-Wahab, 2016) including leukemia, lung, colon, and head and neck squamous cell carcinoma; its overexpression in gastric cancer was not reported yet. Summaries of other kinases are provided in Table 2. Table S12 displayed the support validation evidences for kinases in published paper and the immunohistochemistry results of kinases in 11 stomach cancer tissues from human protein atlas (https://www.proteinatlas.org). Among the 16 kinases we identified, 10 kinases have been validated as potential targets in different cancers.

Figure 4.

Major Pathways Mapped by the Kinases Activated in DGC

(A) Potential kinase targets in DGC.

(B) Coexpression modules of unfavorable phosphorylation sites prognostic markers.

(C) Survival curves of the phosphorylation sites within the three modules.

In addition, we identified 287 unfavorable prognostic phosphorylation sites, for which higher expression of a given phosphorylation site was correlated with a poor patient survival outcome (FDR<0.1, Table S13). The best ratio cut-offs were determined using the same strategy as the kinase identification. Interestingly, we found these unfavorable prognostic phosphorylation sites had three co-expression modules (Figure 4B). A functional gene ontology (GO) analysis and kinase enrichment (see Methods) were performed for the modules (Table S14). The Module1 was related to Rho protein signal transduction and cell-cell adhesion. Module 2 contained many phosphorylation sites associated with RNA splicing. These phosphorylation sites included S312-THOC5, S597-SRRM1, S549-SRRM1, S965-SCF1, and S796-SCAF11 (Figure 4B). The Module 3 enriched in cell growth and regulation of small GTPase-mediated signal transduction pathways. Three representative phosphorylation sites, S1507-AKAP13, S1261-MTOR, and T1353- MYO9B, were displayed in Figure 4B. Many kinases such as AKT1, RPS6KB1, RPS6KA1, RPS6KA3, PRKCA, PRKCD, RAF1, and PAK4 were enriched. These co-expression modules results confirmed that Rho protein signal transduction, cell-cell adhesion, MTOR signaling pathway, and RNA splicing were major altered pathways in DGC.

Nomination of PTKTs and Inhibitors for Individual Patients

To help nominate PTKTs for DGC patients, we calculated kinase activity for each individual patient. We then selected the top three PTKTs according to their activated kinase activities for the 83 patients (Figure 4B). It was evident that each patient had a unique pattern of activation for the top three PTKTs. The most frequently activated PTKT, CSNK2A1, was identified in 42/83 (50%) of the patients, whereas the second most frequently activated kinases MTOR, MAPK3, and GRK6 were each identified in 25/83 (30%) of the patients and the least frequently activated PTKT, PDPK1, was only identified in 4/83 (5%) of the patients.

Silmitasertib, the CSNK2A1 inhibitor, was in several clinical trials for treating cancers as a single agent or in combination with chemotherapy (Chon et al., 2015). Thus, we nominated Silmitasertib as an actionable reagent. The second best available actionable kinase inhibitors were the MTOR inhibitor, everolimus, and the MAPK3 inhibitor Ulixertinib. The oral MTOR inhibitor everolimus was evaluated in the phase III GRANITE-1 (First Gastric Antitumor Trial With Everolimus; Clinical Trial No. NCT00879333) (Ohtsu et al., 2013) and did not significantly improve OS for advanced gastric cancer. We speculate that inhibition of MTOR might be more effective if patients with hyperactive MTOR were selected in the trial. Ulixertinib is a potent and selective small molecule inhibitor of ERK1 and ERK2. It holds promise as a treatment for ERK-dependent cancers including colon cancer and melanoma with BRAF mutation (Germann et al., 2017). Although the frequency of BRAF mutation is rare in DGC, its downstream effector MAPK3 kinase activity is elevated, rendering them amenable to Ulixertinib treatment. Given the diverse PTKT activation profile of the 83 DGC patients, we propose that kinase inhibitions as a means of cancer treatment needs to be done on an individual basis for this cohort of patients.

Validation of CSNK2A1 Inhibition with Silmitasertib in GC Cell Lines

To test the hypothesis that the computed kinase activity from phosphoproteomics data could predict sensitivity to kinase inhibition, we repeated the above approach in gastric cancer cell lines so that we could carry out experimental validation. We measured ten GC cell lines and obtained a quantitate proteome of 8,973 proteins and a phosphoproteome of 10,332 phosphorylation sites (see Methods, Tables S15 and S16, Figures S11 and S12). We then measured IC₅₀ values of Silmitasertib in the ten GC cell lines. The activity of CSNK2A1 for each cell line was calculated by the mean value method based on the phosphoproteomics data. It was evident that the computed CSNK2A1 activity in cancer cell lines was negatively correlated with the IC₅₀ value of Silmitasertib (Spearman, R = −0.47, p = 0.04, Figure 5). We also found that the abundance of CSNK2A1 in the ten gastric cancer cell lines was also negatively correlated with IC₅₀ value of the Silmitasertib (Spearman, R = 0.69, p = 0.013, Figure 5). It seems that both the activity and abundance of CSNK2A1 could predict Silmitasertib sensitivity of GC cell lines.

Figure 5.

A Patient-Specific Strategy for Nomination of Kinases

(A) Patient-specific kinase inhibitors for DGC.

(B) The relationship of CSNK2A1 activity with the IC₅₀ of Silmitasertib.

Also see Figures S11 and S12.

Discussion

We measured phosphoproteomes of 83 DGC patients to paint the phosphoproteomic landscape of DGC. Using a global phosphoproteomics data analysis workflow that we developed recently (Tong et al., 2018), we demonstrated that large-scale phosphoproteomics alone could be used to classify DGCs into subtypes that are associated with distinct clinical outcomes as well as nomination of kinase targets for further clinical evaluation.

We previously subtyped the same DGC patients with protein profiling data (PX1–3) (Ge et al., 2018). Our current study suggests that subtyping with phosphoproteomics data may be more accurate, as the Ph1 group that is associated with best OS retrieved some patients assigned to the PX2 and PX3 groups but with better OS. It remains to be tested whether the accuracy of prediction for OS and chemosensitivity can be further improved by using both whole proteomics and phosphoproteomics data. It will be a significant challenge to translate the current findings into clinical practice.

Protein kinases have become a major class of drug targets, and kinase inhibitors have demonstrated their efficacy in the treatment of many cancers (Wu et al., 2015, Wu et al., 2016, Fabbro et al., 2015). In this study, we investigated kinase activation in a systematic manner to handle large-scale phosphoproteomics data. Our analyses successfully dysregulated kinase pathways in DGC that conformed to the general conceptual framework of cancer hallmarks. In addition to the finding that the MTOR signaling network was indeed the key for DGC signaling as expected, the finding that SRPK2 kinase was activated in DGC on both protein abundance and kinase activity suggests that RNA splicing may play a significant role in DGC. It is necessary to delineate the exact signaling pathways to pinpoint the substrates of SRPK2 and what downstream biochemical pathways they regulate.

Noticeably, clinical trials of kinase inhibitors of HER2, EGFR, AKT1, and MTOR have been successfully used in other cancer treatment but failed in DGC (Choi et al., 2016). We speculate that one of the major causes might be the selection of patients in the clinical trials. Take MTOR as an example: among the 25 DGC patients whose MTOR kinase activity was among the top three most activated in our cohort, only eight were the number one most activated potential therapeutic kinases. Stratifying patients with activated MTOR activity might increase the success rate for MTOR inhibition to work in DGC patients.

It is exciting to find about 50% of the DGC patients with hyperactivation of CSNK2A1 and an actionable inhibitor, Silmitasertib, is available. This suggests the possibility of translating this finding into clinical actions. Moreover, we validated such a finding in 10 gastric cancer cell lines that calculated CSNK2A1 kinase activity from our bioinformatics workflow could predict the sensitivity to CSNK2A1 inhibition with Silmitasertib. Recently, a more potent CSNK2A1 inhibitor was reported to have anticancer activity in cell line and mouse models (Oshima et al., 2019). Given that DGC patients lack any targeted therapy options, it is worthy trying in a preclinical setting to test whether CSNK2A1 inhibition would demonstrate efficacy.

The goal of individualized treatment in cancer is paradoxical in practice. It is more accurate and elegant to perform analysis of each individual tumor to find the druggable targets, but it is almost cost prohibitive to do so in clinical practice. The best compromise would be to find activated kinases with high frequency in a cohort of cancer, demonstrate their preclinical efficacy, and then develop a practical companion test for each high-frequency kinase so that a clinical trial can be carried out selecting the right patients. This work is the first step in this line of approach, as we have identified and nominated a collection of kinases that activated with high frequency and are correlated with OS in DGC.

Limitations of the Study

Our current bioinformatics workflow only used less than 20% of the data due to the limitation of the kinase-substrate database, in which we could only analyze phosphorylation data included in the database. How to make use of the rest of the 80% of data would further strengthen the utility of the method as well as enhance our understanding of DGC and hold the promise to find novel pathways and drug targets. We note that it is necessary to further develop the current method to use in a clinical setting. Biopsy samples instead of the resected samples, for example, from stomach endoscopy may be used to acquire the phosphoproteomics data. An improvement of the current methodology is to make use of tiny amount of the biopsy samples to acquire data and be able to analyze the data for cancer molecular subtyping and nomination of actionable kinase targets.

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.

Published: December 20, 2019