Introduction

Colorectal cancer (CRC) is the third most common cancer worldwide [1, 2]. Despite advances in early detection and treatment of CRC patients, CRC remains the second most deadly cancer worldwide, accounting for more than 900,000 deaths every year [1, 2]. Currently, the tumour-node-metastasis (TNM) staging system remains the most important clinically used factor to predict patient prognosis [3, 4]. However, even within the same TNM stage, the prognosis of patients may differ significantly, most likely due to heterogeneity in patient and CRC characteristics [3-6]. It has been suggested that CRCs represent a heterogeneous group of tumours that develop via several distinct molecular pathways involving different genetic and epigenetic alterations [7-10]. One of the most frequently activated molecular pathways is the PI3K/AKT/mTOR signalling pathway, which is regulated by the tumour suppressor PTEN [11, 12]. It has been proposed that the PI3K/AKT/mTOR signalling pathway is involved in ‘rewiring’ cancer metabolism from oxidative phosphorylation towards aerobic glycolysis through the regulation of three transcription factors: HIF-1α, MYC, and p53 [13-18].

Aerobic glycolysis, also known as the ‘Warburg effect’, is characterised by increased glucose uptake and lactate secretion, even in the presence of oxygen [16, 17, 19]. First, glucose uptake by tumour cells is stimulated by upregulation of the expression of glucose transporter 1 (GLUT1) [17, 20]. Then, glycolytic flux and lactate production are increased by upregulation of the expression of pyruvate kinase M2 (PKM2), pyruvate dehydrogenase kinase 1 (PDK1) and lactate dehydrogenase A (LDHA) [16, 17]. Finally, the expression of monocarboxylate transporter 4 (MCT4) is increased to promote lactate secretion and prevent cytoplasmic acidification [17, 20].

The Warburg effect is thought to increase the malignant potential of tumour cells [17] and may contribute to therapy resistance [21]. Glycolysis-related proteins are therefore considered to have potential prognostic value [22]. Numerous studies have investigated the prognostic potential of key glycolytic enzymes and transcriptional regulators in various types of cancer, including CRC, but results remain inconsistent (see [22] for meta-analysis). However, most previous studies focused on investigating a single protein involved in the Warburg effect, while this pathway is much more complicated. Therefore, there may not be one single protein driving the Warburg effect, but rather a combination of proteins.

In the present study, we therefore attempted to capture the Warburg effect by ensuring that the different steps of the pathway were represented by at least one protein. These steps include: (1) upstream regulation of the Warburg effect (PTEN, p53), (2) more glucose entering the pathway (GLUT1), (3) enhanced glycolysis (PKM2), (4) increased lactate production (LDHA), and (4) enhanced lactate secretion (MCT4). The expression levels of these six proteins (PTEN, p53, GLUT1, PKM2, LDHA, MCT4) were combined into a sum score. Based on the sum score, patients were divided into three subgroups representing low, moderate or high likelihood of the presence of the Warburg effect, hereafter referred to as the “Warburg-subtypes” (Warburg-low, -moderate, and -high, respectively). We then investigated the relationship between these Warburg-subtypes and patient survival in a large population-based series of CRC patients. We hypothesised that patients with Warburg-high CRC have a worse prognosis compared to patients with Warburg-low CRC.

Methods Design and study population

The population-based series of CRC patients was derived from the prospective Netherlands Cohort Study (NLCS), which has been described in detail previously [23]. In short, the NLCS was initiated in September 1986 (baseline) when 120,852 men and women, aged 55–69 years, completed a mailed, self-administered questionnaire on diet and other cancer risk factors [23]. The NLCS was approved by the institutional review boards of the TNO Quality of Life Research Institute (Zeist, the Netherlands) and Maastricht University (Maastricht, the Netherlands) (METC number 85-012). All cohort members consented to participation by completing the questionnaire. Follow-up for cancer incidence was established by annual record linkage with the Netherlands Cancer Registry and PALGA, the nationwide Dutch Pathology Registry [24, 25], covering 20.3 years of follow-up since study initiation (17 September 1986 until 1 January 2007). The estimated completeness of cancer incidence follow-up was >96% [26]. After excluding patients who reported a history of cancer (excluding non-melanoma skin cancer) at baseline, 4,597 incident CRC patients were available (Figure 1).

Flow diagram of the number of CRC patients available for analyses in the Netherlands Cohort Study (NLCS), 1986–2006.

Tissue collection and TMA construction

Formalin-fixed paraffin-embedded (FFPE) tissue blocks from CRC patients were collected as part of the Rainbow-Tissue MicroArray (TMA) project during 2012–2017 [27] (supplementary material, Table S1). Tumour and normal tissue FFPE blocks were requested for 3,872 incident CRC patients, selected based on available linkage to PALGA record (which provides access to pathology laboratories) and surgical or endoscopic resection specimen with pathology report, excluding those who received neoadjuvant therapy. FFPE blocks from 3,021 CRC patients were retrieved from 43 pathology laboratories throughout the Netherlands (78% retrieval rate), after excluding patients without approval of donor pathology labs, without pathology report or FFPE blocks. For TMA construction, H&E-stained sections were reviewed by pathologists and areas with the highest tumour density were marked. Three 0.6 mm diameter cores with tumour and three cores with normal epithelium were sampled per patient (TMA-Grandmaster, 3D-Histech, Hungary). After excluding patients with unusable FFPE blocks, tumour cores from 2,694 CRC patients were successfully assembled in 78 TMA blocks (Figure 1).

Immunohistochemistry

Five micrometre thick serial sections were cut from all 78 TMA blocks and subjected to immunohistochemistry (IHC). Details of the primary antibodies and staining protocols are shown in supplementary material, Table S2. After IHC, TMA sections were scanned using the Aperio scanner (Leica Microsystems, Milton Keynes, UK) at 40× magnification at the University of Leeds (UK) Scanning Facility.

Three non-pathologist observers (GEF: histology technician; KO: PhD student; JCAJ: PhD student) were trained by a senior histopathologist (HIG) in recognising adenocarcinoma and IHC scoring. Presence of adenocarcinoma was confirmed for every individual core by reviewing the H&E-stained TMA sections in combination with the pan-cytokeratin stained sections if necessary. Requiring at least one tumour core per patient, tumour cores of 2,497 CRC patients passed quality control (Figure 1).

After quality control, all tumour cores were scored by at least two observers, independently and blinded for patient characteristics (see supplementary material, Table S3 for percentage of slides evaluated per observer). Scoring protocols for all proteins and kappa values for inter- and intra-observer agreement are shown in supplementary material, Table S4. In brief, the expression of p53 in the tumour cells was scored as negative; 1–10% positive nuclei; 11–50% positive nuclei; 51–90% positive nuclei; and 91–100% positive nuclei. PTEN expression was scored as negative (no staining in tumour cytoplasm); weak (staining of tumour cytoplasm weaker than adjacent stroma); moderate (similar staining intensity in tumour cytoplasm and adjacent stroma); or strong (staining of tumour cytoplasm stronger than adjacent stroma), according to the protocol of Richman et al [28]. GLUT1 and MCT4 expression were scored as negative; 1–10% tumour cells with membranous staining; 11–50% positive tumour cells; and >50% positive tumour cells. LDHA was evaluated according to the protocol of Koukourakis et al [29], with minor adaptations. LDHA expression was scored in the tumour cells as negative/weak cytoplasmic staining; 1–50% tumour cells strong cytoplasmic staining; >50% tumour cells strong cytoplasmic staining. PKM2 expression was scored in the tumour cells as negative/weak cytoplasmic staining; moderate cytoplasmic staining; 1–50% tumour cells strong cytoplasmic staining; and >50% tumour cells strong cytoplasmic staining.

Supplementary material, Figure S1 shows a flow diagram of the process of combining multiple core-level scores into patient-level Warburg-subtypes. Scores from individual observers were combined into a ‘combination score’ if the same score was given by at least two observers. If the score was discrepant between observers, cores were either reviewed jointly by the two initial observers to agree on a final score, or an experienced pathologist (IS) with special interest in gastrointestinal pathology determined the final score. To obtain patient-level data for every protein, the scores of all available tumour cores (1–3 tumour cores per patient) were averaged and the value was rounded to the nearest scoring category. The average score was categorised to achieve three approximately equal-sized groups, representing low, moderate, and high protein expression. Cut-offs for PTEN and p53 were based on published literature [28, 30], cut-offs for other proteins were determined based on the distribution of patients (supplementary material, Table S4 shows cut-offs per protein).

Establishing Warburg-subtypes

Warburg-subtypes were created using a sum score, where high protein expression for p53, GLUT1, LDHA, MCT4 or PKM2 added a score of 2 per protein, moderate expression a score of 1 per protein, and low expression a score of 0 per protein. Since PTEN is inversely associated with the Warburg-effect [31], its score was reversed, i.e. 2 = low expression, 1 = moderate expression, 0 = high expression. The resulting sum score ranged from 0 to 12, where a higher sum score indicated a higher probability of the presence of the Warburg effect. Patients with missing data for one or more of the proteins were excluded from further analyses, resulting in 2,399 CRC patients for which a Warburg-subtype could be determined (Figure 1). Based on the sum score, CRC patients were categorised into the ‘Warburg-low’ subtype (sum score 0–3, n = 698), ‘Warburg-moderate’ subtype (sum score 4–5, n = 859) or ‘Warburg-high’ subtype (sum score 6–12, n = 842) (supplementary material, Figure S1).

DNA mismatch repair status

DNA mismatch repair (MMR) status was assessed by IHC for MLH1 and MSH2 proteins (see supplementary material, Table S2 for primary antibodies and staining protocols), and immunostaining was evaluated according to the protocol of Richman et al [28]. Nuclear immunostaining of stromal cells or lymphocytes adjacent to the tumour served as internal positive controls. Tumours with loss of either MLH1 or MSH2 expression, in the presence of internal positive controls, were considered MMR deficient (dMMR). Tumours that expressed both MLH1 and MSH2 were considered MMR proficient (pMMR).

Clinical characteristics and follow-up

Information on patient and tumour characteristics, such as incidence date, TNM stage, tumour location, tumour differentiation grade, and initial treatment information was retrieved from the cancer registry or PALGA histopathology reports. Patients who were diagnosed at autopsy were excluded (n = 5) because the date of diagnosis was not known, leaving 2,394 CRC patients for analyses (Figure 1). Follow-up for vital status of the CRC patients was carried out through linkage with the Central Bureau of Genealogy and the municipal population registries until 31 December 2012. Cause of death was retrieved from Statistics Netherlands. CRC-specific deaths included those with an underlying cause attributed to malignant neoplasms of the colon, rectosigmoid junction, and rectum (ICD-10 codes C18–C20). Vital status was available for 2,393 patients, and information regarding CRC-specific death was available for 2,356 patients.

Statistical analyses

Descriptive statistics and frequency distributions were calculated for clinical characteristics. Differences between Warburg-subtypes were evaluated using Chi-square for categorical variables and Kruskal–Wallis tests for continuous variables. The primary endpoints of the current study were CRC-specific survival, defined as the time from CRC diagnosis to CRC-related death or end of follow-up, and overall survival, defined as the time from CRC diagnosis to death from any cause or end of follow-up. Because of the limited number of events in the later period with follow-up of more than 10 years (CRC-specific deaths: n = 33, 3.2%; overall deaths: n = 275, 15.1%), survival analyses were restricted to 10 years of follow-up. Analyses were stratified for TNM stage and tumour location. The relationship between Warburg-subtypes and CRC-specific or overall survival was estimated using Kaplan–Meier curves and Wilcoxon tests. Hazard ratios (HRs) and 95% CIs were estimated using Cox proportional hazards regression.

The proportional hazards assumption was tested using the scaled Schoenfeld residuals [32], by evaluating −log–log transformed survival curves and by introducing time-covariate interactions into the models. HRs were adjusted for a set of a priori selected prognostic factors: age at diagnosis, sex, tumour location, TNM stage, differentiation grade, MMR status and adjuvant therapy. A separate category (‘unknown’) was used for patients with unknown clinical information regarding TNM stage, differentiation grade, adjuvant therapy, or MMR status to enable inclusion of these patients in the Cox proportional hazards models.

Cancer stage was based on the pathological TNM classification, according to the edition that was valid at the time of cancer diagnosis (supplementary material, Table S5). Hence, five different TNM versions have been used during the follow-up period (TNM versions 3–6). However, the main TNM stage groupings (I/II/III/IV) have remained essentially unchanged [33]. Year of diagnosis (1986–2006) and TNM version were considered as potential confounders to account for potential differences in clinical practice over the years. Both variables were not included in our final models because they did not introduce a ≥10% change in HRs.

In sensitivity analyses, we repeated analyses after excluding CRC patients who died within 30 days after diagnosis (n = 93). Furthermore, analyses were repeated after excluding CRC patients with unknown clinical information regarding TNM stage, differentiation grade, adjuvant therapy or MMR status (n = 265).

All analyses were conducted in Stata Statistical Software: Release 16 (StataCorp., College Station, TX, USA). P values <0.05 were considered significant.

Results

In total, 2,394 CRC patients were available for analyses and classified as Warburg-low (n = 695, 29.0%), Warburg-moderate (n = 858, 35.8%) or Warburg-high (n = 841, 35.1%), based on the combined protein expression levels of LDHA, GLUT1, MCT4, PKM2, p53, and PTEN (supplementary material, Table S6).

Clinical characteristics

Clinical characteristics of the 2,394 included CRC patients are shown in Table 1. Warburg-subtypes differed significantly with respect to tumour location, TNM stage, tumour extension (pT), lymph node involvement (pN), differentiation grade, and MMR status, but did not differ with respect to age at diagnosis, sex, and adjuvant therapy status. The Warburg-high subtype was more often observed in tumours located in the colon, whereas the Warburg-low and -moderate subtypes were more often observed in tumours located in the rectum or rectosigmoid (p = 0.027). Furthermore, the Warburg-high subtype was more common in TNM stage IV tumours, whereas Warburg-low and -moderate subtypes were more common in TNM stage I tumours (p = 0.001). The Warburg-high subtype was more frequently observed in tumours with a higher primary tumour extension (pT, p = 0.007) and tumours with lymph node involvement (pN, p = 0.006). Lastly, Warburg-high tumours were more often poorly differentiated (p < 0.001) and MMR deficient (p < 0.001) compared to Warburg-low and -moderate tumours.

Table 1. Clinical characteristics of the colorectal cancer patients within the Netherlands Cohort Study (NLCS,1986–2006, total n = 2394) according to Warburg-subtypes. Clinical characteristics Total CRC patients (n = 2394) Warburg-subtypes Warburg-low (n = 695) Warburg-moderate (n = 858) Warburg-high (n = 841) P value* Year of diagnosis 1986–1988 109 (4.6) 38 (5.5) 35 (4.1) 36 (4.3) 0.211 1989–1991 206 (8.6) 60 (8.6) 81 (9.4) 65 (7.7) 1992–1994 306 (12.8) 80 (11.5) 107 (12.5) 119 (14.2) 1995–1997 426 (17.8) 128 (18.4) 161 (18.8) 137 (16.3) 1998–2000 444 (18.6) 146 (21.0) 152 (17.7) 146 (17.4) 2001–2003 442 (18.5) 109 (15.7) 159 (18.5) 174 (20.7) 2004–2006 461 (19.3) 134 (19.3) 163 (19.0) 164 (19.5) Age at diagnosis in years, median (range) 74.0 (55.0–89.0) 74.0 (55.0–89.0) 74.0 (56.0–88.0) 74.0 (56.0–89.0)

0.645†

Sex, n (%) Men 1333 (55.7) 406 (58.4) 485 (56.5) 442 (52.6) 0.058 Women 1061 (44.3) 289 (41.6) 373 (43.5) 399 (47.4) Tumour location, n (%) Colon 1703 (71.1) 467 (67.2) 608 (70.9) 628 (74.7) 0.027 Rectosigmoid 234 (9.8) 81 (11.7) 81 (9.4) 72 (8.6) Rectum 457 (19.1) 147 (21.2) 169 (19.7)

141 (16.8)

pTNM stage, n (%) I 468 (19.6) 170 (24.5) 172 (20.1) 126 (15.0) 0.001 II 909 (38.0) 260 (37.4) 309 (36.0) 340 (40.4) III 625 (26.1) 163 (23.5) 233 (27.2) 229 (27.2) IV 335 (14.0) 82 (11.8) 123 (14.3) 130 (15.5) Unknown 57 (2.4) 20 (2.9) 21 (2.5)

16 (1.9)

Tumour extension (pT), n(%) T1 101 (4.2) 39 (5.6) 35 (4.1) 27 (3.2) 0.007 T2 454 (19.0) 152 (21.9) 174 (20.3) 128 (15.2) T3 1535 (64.1) 421 (60.6) 542 (63.2) 572 (68.0) T4 239 (10.0) 62 (8.9) 84 (9.8) 93 (11.1) Unknown 65 (2.7) 21 (3.0) 23 (2.7) 21 (2.5) Lymph node involvement (pN), n (%) N0 1247 (52.1) 377 (54.2) 450 (52.5) 420 (49.9) 0.006 N+ 870 (36.3) 220 (31.7) 314 (36.6) 336 (40.0) Unknown 277 (11.6) 98 (14.1) 94 (11.0) 85 (10.1) Differentiation grade, n (%) Well 205 (8.6) 80 (11.5) 76 (8.9) 49 (5.8) <0.001 Moderate 1571 (65.6) 463 (66.6) 565 (65.9) 543 (64.6) Poor/undifferentiated 415 (17.3) 89 (12.8) 139 (16.2) 187 (22.2) Unknown 203 (8.5) 63 (9.1) 78 (9.1) 62 (7.4) Adjuvant therapy, n (%) No 1874 (78.3) 547 (78.7) 668 (77.9) 659 (78.4) 0.181 Yes 499 (20.8) 137 (19.7) 185 (21.6) 177 (21.1) Unknown 21 (0.9) 11 (1.6) 5 (0.6) 5 (0.6) dMMR, n (%) No 2116 (88.4) 628 (90.4) 775 (90.3) 713 (84.8) 0.001 Yes 254 (10.6) 58 (8.4) 79 (9.2) 117 (13.9) Unknown 24 (1.0) 9 (1.3) 4 (0.5) 11 (1.3) * P value for the Chi-square test, unless otherwise specified. † P value for the Kruskal–Wallis test. Survival

The median (range) follow-up time since diagnosis was 4.86 years (0.0027–25.99 years). Survival analyses were restricted to 10 years of follow-up. During these first 10 years of follow-up, 1,551 (64.8%) deaths were observed, of which 986 (63.6%) were CRC-related deaths.

Univariable Kaplan–Meier curves showed differences between Warburg-subtypes for CRC-specific survival (p = 0.0037) and overall survival (p = 0.0004) (Figure 2). Patients with Warburg-high tumours had a significantly worse CRC-specific survival (HR 1.30; 95% CI 1.11–1.52) and overall survival (HR 1.26; 95% CI 1.12–1.43) compared to patients with Warburg-low tumours in univariable analyses (Table 2). The Warburg-high subtype remained a significant predictor of prognosis in multivariable-adjusted analyses (HR 1.17; 95% CI 1.00–1.38, and HR 1.19; 95% 1.05–1.35 respectively) (Table 2). No significant difference in survival was observed for CRC patients with Warburg-moderate compared to Warburg-low tumours.

Kaplan–Meier curves according to metabolic subtypes (i.e. ‘Warburg-low’, ‘Warburg-moderate’, ‘Warburg-high’) in colorectal cancer patients within the Netherlands Cohort Study (NLCS, 1986–2006, total n = 2394), showing (A) CRC-specific survival and (B) overall survival.

Table 2. Univariable and multivariable-adjusted Hazard Ratios (HRs) for associations between Warburg-subtypes and survival within the Netherlands Cohort Study (NLCS, 1986–2006). CRC-specific survival Overall survival HR (95% CI) HR (95% CI) n CRC deaths (%) Univariable Multivariable-adjusted* Deaths (%) Univariable Multivariable-adjusted* Warburg subtypes Warburg-low 695 258 (37.1) 1.00 (ref) 1.00 (ref) 424 (61.0) 1.00 (ref) 1.00 (ref) Warburg-moderate 858 360 (42.0) 1.16 (0.99–1.37) 1.05 (0.89–1.23) 548 (63.9) 1.08 (0.95–1.23) 1.02 (0.90–1.16) Warburg-high 841 368 (43.8) 1.30 (1.11–1.52) 1.17 (1.00–1.38) 579 (68.8) 1.26 (1.12–1.43) 1.19 (1.05–1.35) * Adjusted for age at diagnosis, sex (men/women), tumour location (colon/rectosigmoid/rectum), TNM stage (I/II/III/IV/unknown), differentiation grade (well/moderate/poor/unknown), adjuvant therapy (yes/no/unknown) and MMR deficiency (yes/no/unknown).

Univariable and multivariable-adjusted HRs for other relevant prognostic factors included in the model are shown in supplementary material, Table S7. In multivariable-adjusted analyses, age at diagnosis (per year), TNM stage, and tumour differentiation grade were associated with a significantly worse CRC-specific and overall survival, while adjuvant therapy and MMR deficiency were associated with better survival. Moreover, women had improved overall survival, but not CRC-specific survival. No significant associations were found between tumour location and CRC-specific or overall survival.

Next, we stratified CRC patients by TNM stage to assess the disease stage-dependent prognostic value of Warburg-subtypes. Univariable Kaplan–Meier curves showed that CRC-specific survival differed between Warburg-subtypes in TNM stage III (p = 0.0011), but not in the other TNM stages (supplementary material, Figure S2). Univariable Cox regression analyses revealed that patients with Warburg-high tumours had a significantly poorer CRC-specific (HR 1.40; 95% CI 1.06–1.85) and overall survival (HR 1.42; 95% CI 1.12–1.80) compared to patients with Warburg-low tumours in TNM stage III (Table 3). After multivariable adjustment, the Warburg-high subtype remained a significant predictor of CRC-specific (HR 1.45; 95% CI 1.10–1.92) and overall mortality (HR 1.47; 95% CI 1.15–1.87) in TNM stage III (Table 3).

Table 3. TNM stage-specific univariable and multivariable-adjusted Hazard Ratios (HRs) for associations between Warburg-subtypes and survival within the Netherlands Cohort Study (NLCS, 1986–2006). CRC-specific survival Overall survival HR (95% CI) HR (95% CI) n CRC deaths (%) Univariable Multivariable-adjusted* Deaths (%) Univariable Multivariable-adjusted* TNM stage I Warburg-low 170 27 (15.9) 1.00 (ref) 1.00 (ref) 80 (47.1) 1.00 (ref) 1.00 (ref) Warburg-moderate 172 31 (18.0) 1.14 (0.68–1.91) 1.12 (0.67–1.89) 74 (43.0) 0.92 (0.67–1.27)