Dynamic phospho-proteogenomic analysis of gastric cancer cells suggests host immunity provides survival benefit

Comprehensive proteogenomic profiles of GC cell lines with phenotypic drug response

We started by characterizing GC cell lines in terms of fifty percent growth inhibitory concentration (GI₅₀) for different chemotherapeutic agents, GC-associated mutations^17,18,19, and RNA and protein expression profiles (Fig. 1). Although 5FU resistance was observed for both 45FU and 74FU cell lines¹⁵, 74FU also had cisplatin (CIS) and etoposide (ETP) resistance, indicating cross resistance among DNA-damaging drugs (Fig. 1A; Supplementary Fig. 1A and 1B). Neither cell line showed apparent resistance to the non-DNA damaging drug docetaxel (DTX) that function as a mitotic inhibitor. As expected from gene alterations reported for GC, all cell lines tested carried TP53 mutations (Fig. 1B). The cell line MKN1 carried a PIK3CA mutation, one of the key determinants of the Epstein-Barr virus (EBV)-positive GC subtype (Fig. 1B), but none of the eight cell lines tested showed EBV gene expression (Supplementary Fig. 1C). Copy number variation (CNV) analysis revealed a positive correlation between the number of CNV loci and CIS sensitivity (Fig. 1C). This correlation resembles CIS hypersensitivity of cells derived from Fanconi anemia patients as a phenotypic consequence of genomic instability²⁰. CIS sensitivity was also associated with a lack of vimentin silencing^21,22 (Fig. 1D). As reflected by their differential drug responses, 45FU and 74FU apparently do not share similarly expressed genes and proteins, except for CDH1, which encodes E-cadherin (Fig. 1E). Indeed, both lines retained most of the genetic programs of their parental cells (Supplementary Fig. 1D). Thus, we further examined reasons for the correlation between CIS sensitivity and increased number of CNV loci, and then pursued unbiased RPPA analysis to identify shared signaling outcomes between 45FU and 74FU.

• Download figure
• Open in new tab

Figure 1.

Integrative multiplatform analyses of GC cell lines. (A−E) GI₅₀ heatmap for each drug tested (A), cell line summary (B), copy number variation (CNV) landscape (C), and mRNA and protein expression heatmaps (D and E) are depicted. Cell lines are ordered by CIS sensitivity. Relevant genes for cell line summary and expression profiles were selected based on previous studies of GC subtypes^17,18.

CIS-sensitive GC cells predominantly demonstrate CIN

To examine reasons for the correlation between CIS sensitivity and increased number of CNV loci, we segregated CIS-sensitive and -resistant cell lines based on their GI₅₀ CIS profiles (Fig. 2A). As expected from Fig. 1, the same cell lines showed CIS and ETP sensitivity and resistance. This observation is further supported by the PRISM repurposing primary screen²³, showing a positive correlation between CIS and ETP sensitivity for both gastric and colon cancer cell lines (Fig. 2B and 2C; Supplementary Table 1). Although no difference in CNV gain is observed, a 3-fold higher level of CNV loss was detected in dual CIS/ETP-sensitive cell lines compared to their resistant counterparts (Fig. 2D). Genes that were frequently lost included c-kit (KIT), kinase insert domain-containing receptor (KDR/VEGFR2), and adhesion G protein-coupled receptor L3 (ADGRL3), which are all located on chromosome at 4q12–q13 (Fig. 2E). Moreover, scores for CIS and ETP sensitivity for gastrointestinal cell lines from the PRISM repurposing primary screen were negatively correlated with KIT copy number (Fig. 2F). Thus, CIS sensitivity in GC cell lines could be defined by chromosomal instability (CIN), represented by loss of proximal 4q12–q13, which could be related to the KIT copy number.

• Download figure
• Open in new tab

Figure 2.

CNV and mesenchymal gene expression in CIS-sensitive GC cells. (A) GI₅₀ profiles of GC cell lines to define sensitivity and resistance to CIS and ETP. (B) Scatter plot showing correlation between area under the dose-response curve (AUC) values for CIS and ETP in gastrointestinal cancer cell lines (right). Each dot indicates an individual cell line (n = 17) shared between the AUC data sets. (C) Heatmaps represent pairwise Pearson correlations between AUC values of CIS and ETP assessed in the PRISM repurposing secondary screen²³. (D) Number of CNV loci detected. Each dot indicates an individual cell line. The two-tailed P values were obtained with a nonparametric Mann-Whitney U test. (E) Binary matrix representing genes that commonly lost its copy number in cell lines with dual CIS/ETP sensitivity. The two-tailed P values were obtained with Fisher’s exact test. (F) Scatter plot showing correlation between AUC values of CIS or ETP and KIT copy number in CCLE gastrointestinal cancer cell lines. Each dot indicates an individual cell line. (G) Hallmark GSEA signatures from RNA-seq data ranked by Normalized Enrichment Score (NES) for CIS/ETP-sensitive vs. CIS/ETP-resistant cell lines (left). GSEA plots showing positive and negative enrichment of “EMT” and “DNA repair” gene sets in CIS/ETP-sensitive cell lines (right). FDRq, false discovery rate (q value). (H) Time course RPPA data showing changes in cleaved PARP levels after CIS treatment. Error bars represent s.e.m. (I) Caspase-focused RPPA analysis of dual CIS/ETP-sensitive (mean, n = 5) and -resistant (mean, n = 3) cell lines (left) and a schematic of CIS-activated signaling outcomes in dual CIS/ETP-sensitive cell lines (right). c-PARP, cleaved PARP; c-C3−9, cleaved caspase-3−9.

Lack of vimentin silencing is associated with CIS/ETP sensitivity induced by CIN

We next investigated the positive correlation between CIS sensitivity and the lack of vimentin silencing. As expected, gene set enrichment analysis (GSEA) showed that “epithelial-mesenchymal transition (EMT)” was the most enriched upregulated gene set in cell lines having dual CIS/ETP-sensitivity (Fig. 2G, left). RNA-seq gene expression and mass spectrometry-based proteomic data from the Cancer Cell Line Encyclopedia (CCLE) also showed a positive correlation between both vimentin gene and protein expression with CIS sensitivity across gastrointestinal cell lines (Supplementary Fig. S2A and S2B), which complements our RPPA-based proteomic data. TWIST, one of the master transcription factors of the vimentin gene, is directly regulated by hypoxia-inducible factor 1α (HIF1A), a key regulator of glycolytic genes, and promotes metastasis²⁴. Supporting this transcriptional mechanism, “Glycolysis” and “Hypoxia” genes are concurrently enriched for cell lines having dual CIS/ETP sensitivity (Fig. 2G, left). These cell lines also showed negative enrichment of “interferon (IFN)α response” and “IFNγ response” pathways. Interestingly, recent work showed that CIN-induced chronic activation of cyclic GMP– AMP synthase (cGAS)-STING, which is a driver of cancer metastasis, led to STING depletion, thereby reducing IFN responsiveness²⁵. Therefore, the lack of vimentin silencing in dual CIS/ETP-sensitive cell lines could be one of the consequences of chronic cGAS-STING activation induced by CIN.

CIS/ETP-induced cell death pathway shares caspase effectors

We also explored central effectors of cell death pathways that are activated by CIS and ETP in GC cells. Consistent with the correlation between CIS/ETP sensitivity and CIN, downregulated genes are significantly enriched for the “DNA repair” pathway (Fig. 2G). Pairwise comparisons of time course protein expression profiles between dual CIS/ETP-sensitive and -resistant cell lines revealed consistently low levels of cleaved PARP in dual CIS/ETP-resistant cell lines regardless of CIS or ETP treatments (Fig. 2H; Supplementary Fig. S2C). Interestingly, PARP^−/− mice were reported to exhibit chromosomal aberrations, including gain and loss in regions of chromosomes 4, 5, and 14²⁶. PARP is a well-established target of caspase cleavage. Subsequent caspase-focused analysis highlighted lack of cleaved caspase-3, -7, and -6 in dual CIS/ETP-resistant cell lines following CIS treatment (Fig. 2I). Both cleaved caspase-3 and -7 cleave PARP to inhibit its enzymatic activity²⁷, whereas cleaved caspase-6 creates a positive feedback loop with caspase-8, -7, and -3^28,29. Therefore, caspase-6 cleavage appears to be a key event in the CIS response in GC cells. Meanwhile, levels of cleaved caspase-9 and X-linked inhibitor of apoptosis protein (XIAP) concurrently increased in response to ETP treatment, reflecting the early plateau of PARP cleavage (Supplementary Fig. S2D). Together, these results suggest that caspase-3 and -7 are the central effectors of cell death pathways activated by CIS and ETP in GC cells.

5FU-induced MAPK signaling pathways converge on phosphorylation of MEK1

Next, we pursued proteogenomic analysis to identify shared signaling outcomes in the two 5FU-resistant cell lines, 45FU and 74FU (Fig. 3A). Pairwise comparison of targeted pharmacogenomic and comprehensive cancer panel sequencing-based mutation profiles between 5FU-resistant and matched parental cell lines identified a missense mutation in neurofibromatosis type 1 (NF1), c.5461G>T (V1821F) that is unique to the 45FU cell line (Fig. 3B and 3C). NF1 encodes a GTPase-activating protein (GAP) that negatively regulates Ras pathway activity by accelerating the hydrolysis of Ras-bound GTP, thereby acting as a tumor suppressor³⁰. However, NF1^V1182F is not listed in the Genome Aggregation Database (gnomAD) (https://gnomad.broadinstitute.org/) as a pathogenic mutation. Previous studies showed that a missense mutation in dihydropyrimidine dehydrogenase (DPYD), c.85T>C (C29R), confers gain-of-function activity^31,32. Despite the three-to five-fold difference in 5FU sensitivity between the 5FU-resistant and matched parental cell lines, both pairs harbor the same DPYD^C29R mutation. Consistent with previous reports^15,16, both 5FU-resistant cell lines showed a two-to four-fold increase in TYMS expression compared to their matched parental cell lines in the presence of 5FU, suggesting that enhanced 5FU metabolism contributes in part to the phenotypic 5FU response (Supplementary Fig. 3A). Although we detected no unique mutations in 74FU cells, the parental cell line MKN74 has a unique mutation in fibroblast growth factor receptor 2 (FGFR2), c.1199G>A (R400Q), suggesting that this cell line contains a minor subpopulation with wild-type FGFR2 that confers a survival advantage in the presence of 5FU. In contrast, 74FU cells harbor a missense mutation in Tet methylcytosine dioxygenase 2 (TET2), c.367C>T (R123C), which was not detected in the parental MKN74 cells. Unlike dual CIS/ETP sensitivity determined by the loss of proximal 4q12–q13, pairwise comparison of CNV profiles between the 5FU-resistant and matched parental cell lines identified no copy number changes shared by 45FU and 74FU cell lines (Fig. 3D). To identify proteins associated with 5FU resistance, we thus determined the correlation between protein level and GI₅₀ 5FU using Spearman’s rank correlation analyses (Fig. 3E). At baseline, signal transducer and activator of transcription (STAT1)-pY⁷⁰¹, MAP-extracellular signal-regulated kinase (MEK) 1-pS²⁹⁸, and Rac1 p21-activated kinase (PAK) 1-pT⁴²³ are the top three proteins that positively correlated with GI₅₀ 5FU. At 24 h after 5FU treatment, MEK1-pS²⁹⁸, but not STAT1-pY⁷⁰¹ and PAK1-pT⁴²³, remained as highly expressed, suggesting that MEK1-pS²⁹⁸ is a key driver molecule of 5FU resistance.

• Download figure
• Open in new tab

Figure 3.

5FU-induced MAPK signaling pathways. (A) GI₅₀ profiles of GC cell lines to define 5FU sensitivity and resistance. Two pairs of 5FU resistant and parental cell lines are highlighted. (B) Pharmacogenomics panel shows drug metabolizing genes harboring single-nucleotide variants (SNVs) in each cell line. DPYD is highlighted as a canonical 5FU metabolism pathway gene. (C) Comprehensive cancer panel. TP53, TET2, FGFR2, and NF1 are highlighted for common, MKN74-unique, and 45FU-unique mutations. (D) CNV analysis showing no shared copy number changes between 74FU and 45FU. Copy number changes were calculated by subtracting the copy number of each chromosomal position in matched parental cell lines from that in resistant cell lines (ΔCNV). Chromosomal positions with maximum and minimum ΔCNVs are highlighted. (E) Spearman’s rank correlation analysis to determine the correlation between GI₅₀ 5FU and protein expression at baseline (top panel) and 24 h after 5FU treatment (bottom panel). (F) Temporal proteomic changes in eight cell lines within 24h of 5FU, CIS, ETP, or DTX treatment. Seven clusters were determined by K-means clustering and further grouped for early, intermediate, late, and no response based on their kinetics. (G) Proportions of protein expression time courses segregated into three doses for each drug. Two-tailed P values were obtained with Fisher’s exact test. (H) Proportions of protein expression time courses from high-dose conditions segregated into six signaling pathways including DNA damage response (DDR), MAPK, PI3K, STAT, NFκB, and WNT pathways. (I) MAPK pathway-focused RPPA analysis of 45FU and parental MKN45 cells. (J) Temporal changes in MEK1-pS²⁹⁸ levels after 5FU treatment. (K) MAPK pathway-focused RPPA analysis of 74FU and parental MKN74 cells. (L) Temporal changes in MEK1-pS²⁹⁸ levels after 5FU treatment. (M) 5FU-activated MAPK signaling pathways in 45FU cells. (N) 5FU-activated MAPK signaling pathways in 74FU cells. Error bars represent s.d. (A) or s.e.m. (J and L).

Dynamic phosphoproteomics identifies MEK1 as a key 5FU-resistance pathway molecule

To see how 5FU affects signaling pathway activity in 5FU-resistant and -sensitive cell lines, we divided the protein expression time courses into four groups, early (peaked by 6 h), intermediate (peaked by 12 h), late (peaked by 24 h), and no response (flat), and then determined the proportions of these dynamics in different signaling pathways including DNA damage response (DDR), mitogen-activated protein kinase (MAPK), phosphoinositide-3-kinase (PI3K), STAT, NFκB, and WNT (Fig. 3F–3H). Compared to a non-DNA damaging drug DTX, DNA-damaging drugs including 5FU, CIS, and ETP reduced the proportions of early, intermediate, and late protein changes in a dose-dependent manner, indicating depletion of protein dynamics (Fig. 3G). The segregation of protein expression time courses by sensitive and resistant cell lines revealed that 5FU-resistant cell lines had lower changes in global protein expression dynamics compared to their sensitive counterparts, whereas the cell lines that are resistant to other drugs showed increased protein expression dynamics in response to each drug (Fig. 3H). Among the signaling pathways tested, the MAPK pathway appears to remain active in the 5FU-resistant cells having global protein dynamics depletion (Fig. 3H), which is consistent with the positive correlation between MEK1-pS²⁹⁸ levels and GI₅₀ 5FU (Fig. 3E).

To identify MAPK signaling outcomes that are shared between 45FU and 74FU, we examined a phosphoproteomic analysis involving RPPA that was focused on the MAPK pathway. Among the three major MAPKs, c-Jun N-terminal kinase (JNK), p38, and extracellular signal-regulated kinase (ERK), JNK and p38 are activated by apoptosis signal-regulating kinase 1 (ASK1) in response to a diverse array of stresses such as oxidative stress, endoplasmic reticulum stress and calcium influx³³, but no established stress-responsive upstream kinases have been reported for ERK. Interestingly, a previous study suggested that Rac1 GTPase, a PAK activator, plays an essential role in activation of gamma-irradiation-induced ERK1/2 signaling in the breast cancer cell line MCF7³⁴. PAK has been demonstrated to phosphorylate S298 of MEK1 and S338 of C-Raf^35,36. Therefore, we included PAK as an upstream kinase of Raf and MEK that can respond to 5FU. As expected from the genetic profiles, the effects of 5FU on signaling pathways differed between 45FU and 74FU cells (Fig. 3I–3L). The 45FU cells showed higher amplitude of ASK1 phosphorylation following 5FU treatment than did the parental cells, whereas phosphorylation levels of both ASK1 and its downstream kinase JNK were largely similar in both 74FU and its parental cell line (Fig. 3I and 3K). In addition to ASK1, phosphorylation of A-Raf at S299 also increased in 45FU cells in response to 5FU, with baseline activity also being higher than in parental cells, potentially due to the loss-of-function NF1 mutation in the cell line (Fig. 3C). Meanwhile, phosphorylation of C-Raf at S338 was instead depleted in 45FU cells, along with an increase in inhibitory phosphorylation at S259 (Fig. 3I). For 74FU cells, smaller or no changes in A-Raf phosphorylation were seen, and only phosphorylation of PAK being observed among kinases upstream of MEK (Fig. 3K). Notably, both 45FU and 74FU cells showed a sharp increase in MEK1 phosphorylation at S298 in response to 5FU (Fig. 3J and 3L). However, phosphorylation of the downstream kinase ERK did not reflect the amplitude and timing of MEK1-pS²⁹⁸ peaks (Supplementary Fig. 3B and 3C). Thus, regardless of their distinct genomic alterations, 5FU-induced MAPK signaling pathways in 5FU-resistant cells converges on phosphorylation of MEK1 either via PAK alone or both PAK and A-Raf (Fig. 3M and 3N).

5FU stimulates divergent signaling cascades in 5FU-resistant GC cells

To support the results of the RPPA-based analysis, we also used GSEA for pathway profiling of 45FU and 74FU cells (Fig. 4A–4D). Unexpectedly, “G2/M checkpoint” was the top upregulated pathway in 45FU cells, but was the top downregulated pathway in 74FU cells, and vice versa for “IFNα response” (Fig. 4A and 4C). To visualize the distinct signaling outcomes, we analyzed RPPA datasets focused on DDR for activation of “G2/M checkpoint” and STAT for “IFNα response” (Fig. 4E–4H). Although the DDR pathway is known to be upregulated upon 5FU treatment, to our knowledge, established 5FU-responsive regulators in the STAT pathway have not been reported. The DNA sensing cGAS-STING pathway is known to activate IRF3 and NFκB in the presence of damaged DNA¹¹. Both IRF3 and transcriptional targets of NFκB like IRF1 and IRF7, can upregulate type I IFNs expression that in turn activates the STAT pathway. Thus, we included proteins and phosphoproteins in the NFκB pathway to assess potential IFNα responses. Consistent with the GSEA data, 45FU cells showed increased DDR pathway activity in response to 5FU, as indicated by higher amplitudes of ATR-pS⁴²⁸, p53-S¹⁵, and p21 expression compared to their parental cells, although levels of ATR-pS⁴²⁸ were high at baseline (Fig. 4E). Of note, 45FU cells are characterized by decrease in cleaved PARP levels, suggesting that intact PARP could detect DNA lesions and recruit repair proteins to the DNA damage³⁷. In contrast to 45FU cells, 74FU cells exhibit little or no changes in DDR pathway protein levels except for increased levels of cleaved caspases (Fig. 4F). As indicated by increased levels of myeloid differentiation factor 88 (MYD88), TAK1-binding protein 2 (TAB2)-pS³⁷² and IκBα-pS^32/36, the NFκB pathway was activated by 5FU in 45FU cells (Fig. 4G). Meanwhile, 74FU cells display decreased levels of MYD88 and TAB2-pS³⁷² at baseline (Fig. 4H), and instead have higher amplitudes of STAT1-pY⁷⁰¹ and STAT3-pS⁷²⁷ expression compared to 45FU cells (Fig. 4G and 4H). This result is in good agreement with GSEA data showing “IFNα response” and “IFNγ response” among the most upregulated pathways (Fig. 4C).

• Download figure
• Open in new tab

Figure 4.

5FU elicits PD-L1 expression in 5FU-resistant GC cells. (A) Hallmark GSEA signatures from RNA-seq data ranked by NES for 45FU vs. parental MKN45 cells. (B) GSEA plots showing positive and negative enrichment of “G2/M checkpoint” and “IFNα response” gene sets in 45FU cells. (C) Hallmark GSEA signatures from RNA-seq data ranked by NES for 74FU vs. parental MKN74 cells. (D) GSEA plots showing positive and negative enrichment of “IFNα response” and “G2/M checkpoint” gene sets in 74FU cells. (E) DNA damage response (DDR) pathway-focused RPPA analysis of 45FU and parental MKN45 cells (left) and a schematic summarizing signaling outcomes in 45FU cells (right). (F) DDR pathway-focused RPPA analysis of 74FU and parental MKN74 cells (left) and a schematic summarizing signaling outcomes in 74FU cells (right). c-PARP, cleaved PARP; c-C3−9, cleaved caspase-3−9. (G) STAT and NFκB (proinflammatory) pathway-focused RPPA analysis of 45FU and parental MKN45 cells (left) and a schematic summarizing signaling outcomes in 45FU cells (right). (H) Proinflammatory pathway-focused RPPA analysis of 74FU and parental MKN74 cells (left) and a schematic summarizing signaling outcomes in 74FU cells (right). (I) Spearman’s rank correlation analysis to determine the correlation between STAT1-pY⁷⁰¹ expression and expression of other proteins (left) and a representative scatter plot showing correlation between STAT1-pY⁷⁰¹ and PD-L1 (right). (J) Spearman’s rank correlation analysis to determine the correlation between IκBα-pS^32/36 expression and expression of other proteins (left) and a representative scatter plot showing correlation between IκBα-pS^32/36 and PD-L1 (right). FDRq, false discovery rate (q value) (B and D).

5FU specifically elicits PD-L1 expression via STAT1 phosphorylation

Notably, both 45FU and 74FU cells demonstrated increased expression of programmed death ligand-1 (PD-L1) in response to 5FU (Fig. 4G and 4H). A previous study reported that surface expression of PD-L1 and phosphorylation of STAT1 in pancreatic cancer cell lines was elicited by treatment with conventional chemotherapeutic drugs including 5FU, gemcitabine, and paclitaxel, although the correlation with phenotypic drug response was not clear³⁸. To clarify whether non-DNA damaging drugs like paclitaxel and DTX have similar effects as DNA damaging drugs on IFN responses, we used GSEA to compare expression of IFN-responsive genes in GC cell lines with and without CIS, ETP, 5FU, or DTX treatment. Interestingly, expression of IFNα and IFNγ pathway genes was upregulated by treatment with CIS, ETP, or 5FU, but not with DTX, whereas NFκB pathway genes were upregulated by all drugs tested (Supplementary Fig. 4). To examine the significance of PD-L1 expression among the proteins that are coexpressed with STAT1-pY⁷⁰¹, we determined the correlation between expression levels of STAT1-pY⁷⁰¹ and the other proteins or phosphoproteins using Spearman’s rank correlation analyses (Fig. 4I). In line with the scenario, PD-L1 correlation is ranked on top along with PAK1-pT⁴²³/PAK2-pT⁴⁰², MEK1-pS²⁹⁸, and IκBα-pS^32/36. The correlation between expression levels of IκBα-pS^32/36 and these other proteins further emphasized an association between STAT1 and PD-L1 (Fig. 4J). Collectively, these results suggest that 5FU elicits not only divergent signaling cascades in 5FU-resistant cells but also convergent PD-L1 expression.

Lymphocyte count predicts a survival benefit with adjuvant chemotherapy

PD-L1, also known as B7-H1 or CD274, is a ligand for programmed cell death-1 (PD-1) that inhibits T cell activation by binding to PD-1 on T cells³⁹. Anti-PD-L1/PD-1 antibodies together with an S-1 adjuvant chemotherapy containing the 5FU prodrug tegafur³, has become a standard treatment for locally advanced GC in Asia⁴⁰. Based on the observation that the two 5FU-resistant cell lines showed increased levels of PD-L1 compared to their parental cell lines, we examined whether PD-L1 could be a prognostic factor for GC patients who underwent gastrectomy followed by S-1 adjuvant chemotherapy using the Northern Japan Gastric Cancer Study patient cohort^41,42,43. Unexpectedly, neither overall survival (OS) nor the relapse-free survival (RFS) rate differed between surgery and S-1 adjuvant chemotherapy in PD-L1⁻ patients (Fig. 5A; Supplementary Fig. 5A). The most remarkable effect was seen when patients were segregated into two groups based on median total lymphocyte count (TLC), where S-1 for TLC⁺ patients demonstrated higher 5-year OS and RFS rates in GC patients than that of surgery alone (5-year OS: 80.7% vs. 67.2%; hazard ratio: 0.46; 95% confidence interval, 0.23‒0.88; P = 0.020; 5-year RFS: 76.5% vs. 63.3%; hazard ratio: 0.49; 95% confidence interval, 0.26‒0.8.9; P = 0.020) (Fig. 5B and 5C; Supplementary Fig. 5B–5D). These observations indicate that host could contribute to relapse after S-1 adjuvant chemotherapy in addition to tumor cell-intrinsic 5FU-resistance or immunosuppression mechanisms (i.e., activation of PAK-MEK1 pathways or PD-L1 expression). In line with this scenario, H. pylori infection, which can recruit T lymphocytes and macrophages to gastric mucosa⁴⁴, might confer a potential survival benefit in patients with advanced GC even after gastrectomy⁴⁵. However, despite a strong correlation with inflammation scores for gastric mucosa based on the Sydney System⁴⁶, the presence of H. pylori does not seem to be directly related to TLC (Fig. 5D and 5E; Supplementary Fig. 5E and 5F), suggesting that TLC may reflect an event that is distinct from that of H. pylori-induced immunity. Whether H. pylori-induced or not, lymphocytes can nonetheless secrete antitumor cytokines such as TNFα/β, and IFNγ⁴⁷. To examine whether lymphocyte-derived cytokines are associated with survival of patients with GC, we compared the hazard ratios for patient survival outcomes based on expression levels of genes regulated by NFκB and STAT1, which can be activated by TNFα/β and IFNγ signaling inputs, with an independent cohort¹⁸. Remarkably, expression levels of NFκB/STAT1 targets including IRF1, CD274, IRF9, and STAT1 are associated with better OS in the cohort (Fig. 5F and 5G). These expression levels are in good agreement with the correlation between TLC and survival outcomes in our cohort. In contrast, other NFκB targets such as NFKBIA, IRF7, and IL6 appear to drive opposite functions and outcomes.

• Download figure
• Open in new tab

Figure 5.

Survival curves for patients with stage II/III GC stratified by potential confounding factors. (A) Kaplan-Meier curves for overall survival (OS) in PD-L1⁻ (left) or PD-L1⁺ (right) patients treated with or without S-1. (B) Kaplan-Meier curves for OS in patients with low total lymphocyte count (TLC⁻) or high total lymphocyte count (TLC⁺) who did or did not receive S-1 treatment. (C) Subgroup analysis of OS in PD-L1⁻ (n = 188) or PD-L1⁺ (n = 78) patients (left) and TLC⁻ (n = 128) or TLC⁺ (n = 138) patients (right) evaluated for surgery alone and surgery followed by S-1 adjuvant chemotherapy. (D) Positive correlation between inflammation score and H. pylori positivity (left). Two-tailed P values were obtained with the nonparametric Mann-Whitney U test. Subgroup analysis of OS in H. pylori⁻ (n = 131) or H. pylori⁺ (n = 66) patients evaluated for surgery alone and surgery followed by S-1 adjuvant chemotherapy (right). (E) Kaplan-Meier curves for OS in H. pylori⁻ (left) or H. pylori ⁺ (right) patients who did or did not receive S-1. (F) Hazards for OS were evaluated for the expression of genes regulated by GC cell-intrinsic signaling in response to lymphocyte cytokines. (G) Kaplan-Meier curves for OS in ACRG GC patients (n = 283) stratified based on tumor IRF1 (left) and NFKBIA (right) expression levels. Cox proportional hazards model was used to determine the hazard ratio of each mRNA expression level (C, D, and F). Error bars represent 95% confidence intervals. The P values were obtained with log-rank test (A, B, E, and G). Stratification strategies are shown in Supplementary Fig. 5G. ACRG, Asian Cancer Research Group¹⁸.

We next validated the finding that NFKBIA expression was associated with poor survival in patients with GC in an immunohistochemical study in our cohort and further examined whether IκBα expression could be a surrogate for low TLC (Supplementary Fig. 6). The number of patients who are TLC⁻ was significantly lower in IκBα⁺ patients than IκBα⁻ patients (Supplementary Fig. 5G). By excluding IκBα⁻ patients, S-1 for TLC⁺ patients demonstrated higher 5-year OS and RFS rates than for patients with GC who had surgery alone (5-year OS: 89.9% vs. 70.6%; hazard ratio: 0.26; 95% confidence interval, 0.06‒0.82; P = 0.026; 5-year RFS: 90.0% vs. 67.7%; hazard ratio: 0.22; 95% confidence interval, 0.05‒0.65; P = 0.008) (Fig. 6A– 6D). Exclusion of IκBα⁻ patients enabled more accurate outcome prediction compared to the method based on only TLC status (Fig. 5B; Supplementary Fig. 5B). Subgroup analysis showed the potential for substantial reduction in the risk of GC-related death and relapse after S-1 adjuvant chemotherapy in patients who were TLC-IκBα double-negative (Fig. 6C and 6E). In addition, S-1 adjuvant chemotherapy provided no survival benefits for IκBα⁺ patients and notably the survival rate was even worse than surgery alone for TLC⁻ patients (Fig. 6F–6J).

• Download figure
• Open in new tab

Figure 6.

Survival curves for patients with stage II/III GC stratified by IκBα level and potential confounding factors. (A) Representative immunohistochemical staining in a IκBα⁻ tumor region. (B) Kaplan-Meier curves for OS in IκBα⁻ patients stratified by treatment (i.e., S-1 and Surgery) divided into TLC⁻ and TLC⁺ groups. (C) Subgroup analysis stratified by interaction of TLC⁻ (n = 53) or TLC⁺ (n = 64) based on the hazard for OS was evaluated by treatment. (D) Kaplan-Meier curves for RFS in IκBα⁻ patients stratified by treatment divided into TLC⁻ and TLC⁺ groups. (E) Subgroup analysis stratified by interaction of TLC⁻ (n = 53) or TLC⁺ (n = 64) based on the hazard for OS was evaluated by treatment. (F) Representative immunohistochemical staining in a IκBα⁺ tumor region. (G) Kaplan-Meier curves for OS in IκBα⁺ patients stratified by the treatment divided into TLC⁻ and TLC⁺ groups. (H) Subgroup analysis stratified by interaction of TLC⁻ (n = 58) or TLC⁺ (n = 49) based on the hazard for OS was evaluated by treatment. (I) Kaplan-Meier curves for RFS in IκBα⁺ patients stratified by treatment divided into TLC⁻ and TLC⁺ groups. (J) Subgroup analysis stratified by interaction of TLC⁻ (n = 58) or TLC⁺ (n = 49) based on the hazard for RFS was evaluated by treatment. Scale bar, 50 μm (A and F). Risk for survival was evaluated using Cox proportional hazards models (C, E, H and J). Error bars represent 95% confidence intervals. The P values were obtained with a log-rank test (B, D, G and I).

Collectively, TLC predicts a survival benefit for patients with advanced GC who received postoperative S-1 adjuvant chemotherapy. Thus, drug-resistance mechanisms involving NFκB signaling may allow GC cells to respond to lymphocyte-derived cytokines such as TNFα/β and IFNγ in the patients’ body.

Cell culture

GCIY, GSS, MKN1, MKN45, and MKN74 cells were obtained from the RIKEN Cell Bank. 5FU-resistant lines 45FU and 74FU were established from MKN45 and MKN74 cells, respectively, as described previously¹⁵. IWT-1 is a cell line that was established in our laboratory from a male Japanese patient with GC who had relapsed with peritonitis carcinomatosa⁵⁴. Use of the IWT-1 cell line was approved by the Iwate Medical University Institutional Review Board (H25-116, and HG H25-15), and written informed consent was obtained from the family of donor patient, who had died at the time the cell line was established. The IWT-1 cell line is now distributed via the RIKEN BRC Cell Bank (https://cell.brc.riken.jp/en/). All eight cell lines were grown in RPMI 1640 media (Life Technologies, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (FBS) (Life Technologies), and cultured at 37°C in a humidified incubator supplied with 5% CO₂.

Drugs

5FU and DTX were purchased from Kyowa-Hakko Bio (Tokyo, Japan), and Sanofi KK (Tokyo, Japan), respectively. CIS and ETP were purchased from Nippon Kayaku (Tokyo, Japan).

Growth suppression assay

Cells were plated in a 96-well plate at 1.0 to 4.0 × 10⁴ cells/well. Cell viability was determined using CCK-8 (Dojindo, Kumamoto, Japan) and a TriStar LB 941 microplate reader (Berthold Technologies, Bad Wildbad, Germany). GI₅₀ values were calculated using GraphPad Prism software version 7.0 (GraphPad software Inc., San Diego, CA, USA).

Targeted gene sequencing

Genotyping using the Ion AmpliSeq Pharmacogenomics Research Panel was conducted according to the manufacturer’s instructions using the Ion Torrent platform (Thermo Fisher Scientific, Waltham, MA, USA). This panel analyzes genetic variants including SNPs, insertion/deletions, and CNVs that are associated with 40 drug metabolizing enzymes. Gene mutation profiling was also performed using the Ion AmpliSeq Comprehensive Cancer Panel that covers nearly all coding regions of 409 genes that are known to be related to cancer (Thermo Fisher Scientific).

RNA-seq

Total RNA was extracted using an RNeasy Kit (Qiagen), according to the manufacturer’s instructions. RNA-Seq libraries were constructed using TruSeq Stranded mRNA HT Sample Prep Kit (Illumina Inc, San Diego, CA, USA). Sequencing was carried out on the Illumina HiSeq3000 platform with a 36-base-single end run. Quality control of RNA-seq reads was performed using FastQC version 0.11.5. Raw sequence reads were mapped to the UCSC genome for human (hg19) using HISAT2 version 2.1.0, without novel splice variant discovery. Fragments per kilobase of exon per million (FPKM) mapped reads were calculated from the mapped reads using Ballgown version 2.6.0. FPKM data was later used in gene set enrichment analysis (GSEA) performed with GSEA Software (version 4.1.0) and WebGestalt⁵⁵.

qRT-PCR

Total RNA was extracted from GC cell lines using an RNeasy kit (QIAGEN). cDNA was synthesized from 500 ng total RNA in a 10 μl reaction volume using PrimeScript RT Master Mix (TaKaRa Bio, Otsu, Japan). Quantitative RT-PCR (qRT-PCR) was performed using the LightCycler Nano System (Roche, Mannheim, Germany). Primer sequences are shown in Supplementary Table 4.

Reverse-Phase Protein Array (RPPA)

For sample collection, cells were exposed to anticancer drugs (5FU, CIS, ETP, and DTX) in a 96-well plate at low, medium, and high concentrations (Supplementary Table 2). Each concentration was applied at five time points over 24 h. Cells were collected by trypsinization followed by pipetting and centrifugation at 1,700 g for 2 min at 4 °C. The resulting cell pellet was stored at - 80 °C until further RPPA analysis as previously described⁵⁶. For RPPA assays, cell pellets were processed to obtain cell lysates according to previously published protocols⁵⁶. An individual RPPA contains 2,000 dots for 12 conditions ×6 time courses ×8 cell lines, and 800 dots for control MIX samples^57,58. The dots were printed on nitrocellulose-embedded glass slides (Grace BioLabs, Bend, OR, USA) with an Aushon 2470 Microarrayer (Quanterix, Billerica, MA, USA). The Mix sample contains a variety of sample lysates from all possible drug administration conditions. Each sample lysate was spotted as nine serial 2-fold dilutions in tetraplicate for quantitative analysis and subsequently probed with individual primary antibodies (Supplementary Table 3) for which the specificity was verified by strip Western blotting^57,58. To block nonspecific interactions, SuperG (Grace Biolabs), iBlock (Life Technologies), or standard BSA solutions were used depending on the antibody. Signals were obtained using a TSA kit (Thermo Fisher Scientific) and subsequently quantified using a TissueScope 4000 Scanner (Huron Technologies, Waterloo, ON).

Cohorts (Discovery and Validation)

GSE62254 data set (n = 300) was used to determine lymphocyte cytokine-responsive genes that are associated with survival of GC patients. The Northern Japan Gastric Cancer Study patient cohort (NCT01905969, n = 658) was used to examine the effects of TLC, tumor PD-L1 and IκBα expression, and H. pylori positivity on treatment outcomes in advanced GC patients.

Immunohistochemistry

A tissue microarray (TMA) including 658 GC specimens was made for immunohistochemistry (IHC) staining. The primary antibodies used were: rabbit anti-IκBα polyclonal antibody (E-AB-70089, Elabscience, Houston, TX, USA), mouse anti-PD-L1 monoclonal antibody (22C3, Dako, Santa Clara, CA, USA), and CD4/CD8 (4B12/SP16; Biocare Medical, Concord, Pacheco, CA, USA). For IκBα scoring, cases in which >30% of all tumor cells were stained were defined as positive. The staining evaluation was focused on the cytoplasm in the epithelial component. For staining evaluation, the MKN1 cell line was used as a positive control, and cases having equal or greater positivity relative to the positive control were considered positive. All scoring was performed by an independent pathologist (A.Y-A.), who was blinded to the clinical outcomes. The scoring algorithms and staining evaluation for PD-L1 and CD4/CD8 have been described previously⁴².

Statistics

Data are shown as mean ± s.d. unless otherwise stated. Statistical analysis was performed using GraphPad Prism (GraphPad Software). Statistical significance of differences between two groups was evaluated using an unpaired two-tailed t test, or as indicated in the figure legends. For the correlation analysis, Pearson correlation coefficient r and P value in linear regression were used to evaluate the strength and statistical significance of the correlation. Differences having a P value of < 0.05 were considered statistically significant for screening purposes. For heatmap visualization, genes or proteins were ordered by average-linkage clustering, complete-linkage clustering, or from highest to lowest values in a column, unless otherwise stated. For survival analysis, patients were divided into two groups based on the median lymphocyte count or IHC positivity.