Nomograms for Prediction of Molecular Phenotypes in Colorectal Cancer.

BACKGROUND: Colorectal cancer (CRC) patients with different molecular phenotypes, including microsatellite instability (MSI), CpG island methylator phenotype (CIMP), and somatic mutations in BRAF and KRAS gene, vary in treatment response and prognosis. However, molecular phenotyping under adequate quality control in a community-based setting may be difficult. We aimed to build the nomograms based on easily accessible clinicopathological characteristics to predict molecular phenotypes.
METHODS: Three hundred and six patients with pathologically confirmed stage I-IV CRC were included in the cohort. The assays for MSI, CIMP, and mutations in BRAF and KRAS gene were performed using resected tumor samples. The candidate predictors were identified from clinicopathological variables using multivariate Logistic regression analyses to construct the nomograms that could predict each molecular phenotype.
RESULTS: The incidences of MSI, CIMP, BRAF mutation and KRAS mutation were 25.3% (72/285), 2.5% (7/270), 3.4% (10/293), and 34.8% (96/276) respectively. In the multivariate Logistic analysis, poor differentiation and high neutrophil/lymphocyte ratio (NLR) were independently associated with MSI; poor differentiation, high NLR and high carcinoembryonic antigen/tumor size ratio (CSR) were independently associated with CIMP; poor differentiation, lymphovascular invasion and high CSR were independently associated with BRAF mutation; poor differentiation, proximal tumor, mucinous tumor and high NLR were independently associated with KRAS mutation. Four nomograms for MSI, CIMP, BRAF mutation and KRAS mutation were developed based on these independent predictors, the C-indexes of which were 61.22% (95% CI: 60.28-62.16%), 95.57% (95% CI: 95.20-95.94%), 83.56% (95% CI: 81.54-85.58%), and 69.12% (95% CI: 68.30-69.94%) respectively.
CONCLUSION: We established four nomograms using easily accessible variables that could well predict the presence of MSI, CIMP, BRAF mutation and KRAS mutation in CRC patients.

Introduction

Colorectal cancer (CRC) is one of the most prevalent and fatal cancers worldwide.1,2 CRC is widely recognized as a result of gradual accumulations of genetic and epigenetic changes involving in different genes and pathways, and thus it is considered as a disease with high heterogeneity.3 This heterogeneous nature confers the variation of CRC patients in treatment response and prognosis. Several molecular phenotypes have been studied to investigate CRC heterogeneity in past decades. Among them, microsatellite instability (MSI), CpG island methylator phenotype (CIMP), and somatic mutations in BRAF and KRAS exons were most widely used in clinical decision-making.4,5

It has been suggested that 5-fluorouracil (5-FU) is an effective chemotherapeutic agent to markedly improve CRC survival in past decades.6 The regimen incorporating irinotecan and capecitabine is a well-established option for use as first-line, second-line and sequential treatment of CRC.7 However, adverse effects on survival were found when oxaliplatin or adjuvant treatment with 5-FU was applied in patients with MSI, while they had a special sensitivity to irinotecan.8–10 Moreover, several studies have shown that a CIMP (+) phenotype might improve the therapeutic effect of 5-FU treatment.11,12

The molecular phenotyping can guide the targeted therapy and immune-checkpoint treatments as well. The response to anti-epidermal growth factor receptor (EGFR) therapy, including cetuximab and panitumumab, also varies in molecular subtypes. It has been well documented that the patients with KRAS mutations would be resistant to anti-EGFR therapies, and thus anti-EGFR agents should be avoided in this subgroup of patients.3 In BRAF-mutant CRCs with advanced stages, the FOLFOXIRI regimen (irinotecan, oxaliplatin, 5-FU and leucovorin) and bevacizumab were considered as a favourable treatment, but they can benefit from oxaliplatin as well as patients with MSI does.8 An anti-EGFR may also be resisted in CIMP-high CRCs that display extensive DNA promoter hypermethylation and tumor suppressor gene repression. In addition, DNA methylation inhibition may be an efficient treatment for tumors with CIMP.13 Of note, MSI has become one of the most popular biomarkers in CRC and other cancers for treatment response to immune checkpoint blockades.8–10 BRAF mutation and CIMP have also been considered as promising prognostic markers in CRC.14

Given the values of these subtypes in distinguishing prognosis and response to therapies, molecular phenotyping is deserved in clinical decision-making. Unfortunately, testing tumor samples for molecular subtype under adequate quality control in a community-based setting sometimes may be difficult due to cost and technique limit, but clinicopathological information is easy-to-get in almost all clinical settings. Therefore, understanding the clinicopathological factors that could predict the presence of MSI, CIMP, and mutations in BRAF and KRAS is becoming crucial to provide crude molecular information for primary care physicians and assist molecular phenotyping for pathologists. Several studies have revealed the specific clinicopathological features associated with each molecular subtype.15–18 For example, CRCs with right-side location or poor differentiation have been shown to be associated with MSI-high, CIMP (+) and BRAF mutation. In addition, CIMP (+), BRAF mutation and KRAS mutation were more frequent in elderly female patients. Our study, therefore, aimed to conduct a comprehensive association analysis of clinicopathological variables with MSI, CIMP, and mutations in BRAF and KRAS, and establish nomograms using these easily accessible predictors for each molecular phenotype to make them be well used in clinical practice.

Materials and Methods

Patients

The eligible patients were identified from the prospectively collected tissue bank of our institute from 2009 to 2012. Three hundred and six patients with pathologically confirmed stage I-IV CRC were included. As shown in Figure 1, the patients with multiple primary cancers, inflammatory bowel disease, tumor samples having extensive DNA degradation and missing medical records, Lynch syndrome, familial adenomatous polyposis, and other hereditary cancer syndromes were excluded. To avoid the potential influence of chemo/radiotherapy on CIMP test and clonal selection of other molecular phenotypes, the patients receiving chemo/radiotherapy before sample collecting were excluded. All the patients were treated and followed according to the NCCN guideline-based protocols in our institute.19,20 The demographic and clinicopathological information of included patients were collected from the medical record. The tumors located in ascending and transverse colon were defined as proximal tumor, while the distal tumor includes the tumors located in descending colon, sigmoid colon, and rectum.21 This study was approved by the Institutional Review Board of the Six Affiliated Hospital of Sun Yat-sen University and conducted in compliance with the Declaration of Helsinki. The written informed consent was obtained from the patients included in this study.

Figure 1

Flow diagram for patient disposition and molecular assays to construct the nomograms for prediction of molecular phenotypes.

Mutational Analysis for KRAS and BRAF

The KRAS exon 2 and BRAFV600E mutation status of resected tumor samples were determined by Sanger sequencing. These mutation analyses were performed at the Molecular Laboratory of our institute under a high-quality control as previously described.22

CIMP Assay

To determine the CpG Island Methylator Phenotype (CIMP) in tumor samples, DNA was extracted (Qiagen, 51306) and bisulfite-treated (Zymo Research, D5002) according to the manufacturer’s protocol. The assay panel, including promoters of five genes (CACNA1G, IGF2, NEUROG1, RUNX3, SOCS1),23 was exploited to assess CIMP using quantitative methylation-specific PCR (qMSP) as previously described.24,25

Microsatellite Analysis

Microsatellite status was analyzed based on five commonly used markers of microsatellite sequence: BAT-25, BAT-26, NR-21, NR-22, and NR-24 using a fluorescence-based pentaplex polymerase chain reaction technique and capillary electrophoresis.26,27

Statistical Analysis

The potential predictive variables, including albumin (≤40 vs >40 g/L), total protein (≤60 vs >60 g/L), platelet counts (≤300×109/L vs >300×109/L), hemoglobin (≤110 vs >110 g/L), MCH (≤27 vs >27 pg), MCHC (≤320 vs >320 g/L), CEA (>5 vs ≤5 ng/mL), AFP (>25 vs ≤25 ng/mL), CA19-9 (≤37 vs >37 kU/L), CA125 (≤35 vs >35 kU/L), and CA153 (≤25 vs >25 kU/L), were preoperatively determined and categorized according to previous studies.20,28,29 BMI (<18.5 vs 18.5–24 vs ≥24 kg/m2) was categorized according to the reference standard in Chinese populations.30 The preoperative CEA/tumor size ratio (CSR), defined as the ratio of CEA level and the maximum tumor diameter, was exploited to evaluate the CEA level per tumor volume as we previously described.31 We used the preoperative neutrophil/lymphocyte ratio (NLR) and platelet/lymphocyte ratio (PLR) to determine the baseline systemic inflammation status in patients,32,33 and receiver operating characteristic curve (ROC) analysis was used to identify the optimum cutoff point for these variables ( and ).

Descriptive statistics were used to summarize baseline characteristics between patients with different molecular phenotypes, and the variables were compared using the Chi-square test or Rank-sum test according to their distributions. To estimate the predictive value of variables for each molecular phenotype, univariate Logistic regression analysis was used, and the odds ratio (OR) and the 95% confidence intervals (95% CI) were calculated. The variables considered significant in the univariate logistic regression analysis were further entered into the backward stepwise multivariable logistic regression analysis, based on which nomograms were constructed to predict the status of CIMP, MSI, KRAS mutation and BRAF mutation. The C-index was acquired for each nomogram, and internal validation using the bootstrap method was performed to determine the adjusted C-index. Calibration curves of the nomograms were generated to show the relationship between the predicted and observed outcomes. The SPSS (23.0) and R (3.6.0) were used for all analyses. The significant values were 2-tailed, and all variables were considered statistically significant if P values were less than 0.05.

Results

Three hundred and six patients meeting the inclusion and exclusion criteria were finally included in this study. Among them, the assays for MSI, CIMP, BRAF mutation, and KRAS mutation are available in 285, 270, 293, 276 patients, respectively (Figure 1), the incidences of which were 25.3% (72/285), 2.5% (7/270), 3.4% (10/293), and 34.8% (96/276) respectively. In consistent with previous studies, patients with CIMP (+) are tightly associated with the status of BRAF mutation (83.3% vs 1.7%, P < 0.001, Table 1). In addition, patients with MSI had significantly higher CIMP (+) frequency (6.9% vs 0.5%, P = 0.004, Table 1), and patients with KRAS mutation had significantly higher BRAF mutation rate (5.6% vs 0%, P = 0.017, Table 1). Patients’ baseline characteristics are summarized in Table 1.

Table 1

Baseline Characteristics of Included CRC Patients with Different Molecular Phenotypes

Variablea		Microsatellite Status				CIMP				BRAF				KRAS
MSS	MSI	MSI%	P	–	+	+%	P	Wild	Mut	Mut%	P	Wild	Mut	Mut%	P
Age	≤ 62	118	33	21.9%	0.761	158	1	0.6%	0.043	150	4	2.6%	0.626	95	52	35.4%	0.826
	>62	95	29	29.1%		141	6	4.1%		133	6	4.3%		85	44	34.1%
Gender	Male	113	44	28.0%	0.235	172	1	0.6%	0.058	161	4	2.4%	0.463	102	51	33.3%	0.573
	Female	100	28	21.9%		127	6	4.5%		122	6	4.7%		78	45	36.6%
BMI	≤18.5	23	4	14.8%	0.419	27	0	0%	0.697	25	1	2.8%	0.513	19	6	24.0%	0.250
	18.5-24	112	40	26.3%		160	3	1.8%		153	3	1.9%		88	58	39.7%
	≥24	68	25	28.1%		97	1	1.1%		92	4	4.3%		63	29	30.7%
Family history of CRC	Yes	3	1	25.0%	0.507	5	0	0%	0.246	5	0	0%	1	5	0	0%	0.330
	No	230	71	23.6%		293	7	2.4%		289	10	3.3%		204	96	32.0%
Tumor location	Proximal	33	19	36.5%	0.039	51	3	5.6%	0.108	51	2	3.8%	0.796	27	25	48.1%	0.025
	Distal	180	53	22.7%		248	4	1.6%		232	8	3.3%		153	71	31.7%
Tumor length	>4.00	97	37	27.6%	0.427	136	7	4.9%	0.014	132	7	5.0%	0.272	89	40	31.0%	0.187
	≤ 4.00	114	35	23.5%		160	0	0%		148	3	2.0%		89	56	38.6%
Mucinous tumor	Yes	21	9	30%	0.528	266	6	2.2%	0.786	30	1	3.2%	0.643	13	17	56.7%	0.008
	No	192	63	24.7%		33	1	2.9%		253	9	3.4%		167	79	32.1%
Differentiation	Poor	29	19	39.6%	0.012	50	6	10.7%	<0.001	44	7	13.7%	<0.001	30	18	37.5%	0.637
	Moderate-well	180	52	22.4%		242	1	0.4%		233	2	1.3%		148	76	33.9%
Lymphovascular invasion	+	12	9	42.9%	0.057	20	3	13.0%	0.004	17	5	22.7%	<0.001	15	4	21.1%	0.189
	-	199	63	24.0%		276	4	1.4%		264	5	1.9%		164	92	35.9%
Perineural invasion	+	15	9	37.5%	0.259	270	7	2.5%	0.687	23	2	8.0%	0.407	17	7	29.2%	0.632
	-	198	63	24.1%		26	0	0%		260	8	3.0%		163	89	35.3%
TNM staging	I	44	16	26.6%	0.965	64	0	0%	0.515	59	0	0%	0.346	34	22	39.2%	0.474
	II	94	30	24.2%		123	4	3.1%		123	4	3.1%		78	46	37.1%
	III	66	24	26.7%		101	3	2.9%		90	6	6.3%		60	25	29.4%
	IV	7	2	22.2%		9	0	0%		9	0	0%		7	2	22.2%
CEA (ng/mL)	>5	46	16	25.8%	0.994	68	4	5.6%	0.124	61	6	9.0%	0.022	37	23	38.3%	0.622
	≤ 5	152	53	25.9%		212	3	1.4%		204	4	1.9%		129	69	34.8%
CA19-9 (kU/L)	>37	28	9	24.3%	0.787	40	3	7.0%	0.127	34	5	12.8%	0.005	23	14	37.8%	0.726
	≤ 37	167	60	26.4%		235	4	1.7%		226	5	2.2%		142	76	34.9%
AFP (ng/mL)	>25	0	1	100%	0.257	1	0	0%	1	1	0	0%	1	0	1	100%	0.359
	≤ 25	191	65	25.4%		266	7	2.6%		253	9	3.4%		159	88	35.6%
CA125 (kU/L)	>35	12	7	36.8%	0.280	17	3	15.0%	0.003	16	4	20%	0.001	12	7	36.8%	0.904
	≤ 35	181	62	25.5%		255	4	1.5%		243	6	2.4%		151	83	35.5%
CA153 (kU/L)	>25	4	0	0%	0.516	4	0	0%	1	4	0	0%	1	3	1	25.0%	0.977
	≤ 25	163	61	27.2%		224	6	2.6%		215	9	4.0%		133	83	38.4%
Albumin (g/L)	≤ 40	78	24	23.5%	0.394	107	7	6.1%	0.001	103	7	6.4%	0.163	65	34	34.3%	0.999
	>40	122	48	28.2%		179	0	0%		168	4	2.3%		109	57	34.3%
Total protein (g/L)	≤ 60	15	5	25.0%	0.693	19	3	13.6%	0.010	18	3	14.3%	0.045	11	9	45.0%	0.472
	>60	126	52	29.2%		188	3	1.6%		179	5	2.7%		110	64	36.7%
Platelet counts (109/L)	≤ 300	155	57	26.9%	0.830	229	3	1.3%	0.035	216	6	2.7%	0.210	129	79	37.9%	0.018
	>300	41	14	25.5%		51	4	7.3%		50	4	7.4%		42	11	20.8%
MCH (pg)	≤ 27	51	17	25.0%	0.590	65	4	5.8%	0.131	61	5	7.6%	0.069	46	18	28.1%	0.227
	>27	131	52	28.4%		199	3	1.5%		188	5	2.6%		115	66	36.4%
MCHC (g/L)	≤ 320	60	20	25.0%	0.545	82	5	5.7%	0.064	76	6	7.3%	0.103	51	25	32.9%	0.758
	> 320	122	49	28.7%		182	2	1.1%		174	4	2.2%		110	59	34.9%
Hemoglobin (g/L)	≤ 110	69	21	23.3%	0.419	87	6	6.5%	0.007	84	5	5.6%	0.370	54	33	37.9%	0.480
	> 110	129	50	27.9%		195	1	0.5%		184	5	2.6%		117	59	33.5%
NLR (median = 2.05)	> 2.05	102	32	23.9%	0.404	143	1	0.7%	0.126	137	4	2.8%	0.704	90	42	31.8%	0.260
	≤ 2.05	96	38	28.4%		138	6	4.2%		130	6	4.4%		80	50	38.5%
PLR (median = 127.34)	≤ 127.34	94	36	27.7%	0.618	142	1	0.7%	0.126	135	3	2.2%	0.328	85	44	34.1%	0.865
	>127.34	104	34	25.0%		137	6	4.2%		130	7	5.1%		85	46	35.1%
CSR (median = 0.64)	≤ 0.64	100	36	26.5%	0.869	139	3	2.1%	1	133	4	2.9%	0.729	87	44	33.6%	0.422
	>0.64	96	33	25.6%		138	4	2.8%		129	6	4.4%		77	48	38.4%
Microsatellite status	MSS	-	-	-	-	212	1	0.5%	0.004	204	5	2.4%	0.131	136	67	33.0%	0.367
	MSI	-	-	-	-	67	5	6.9%		67	5	6.9%		44	28	38.9%
CIMP	-	212	67	24.0%	0.004	-	-	-	-	282	5	1.7%	< 0.001	174	96	35.6%	0.095
	+	1	5	83.3%		-	-	-	-	1	5	83.3%		6	0	0%
BRAF	Wild	204	67	24.7%	0.131	282	1	0.4%	< 0.001	-	-	-	-	170	96	36.1%	0.017
	Mutation	5	5	50%		5	5	50.5%		-	-	-	-	10	0	0%
KRAS	Wild	136	44	24.4%	0.367	174	6	3.3%	0.095	170	10	5.6%	0.017	-	-	-	-
	Mutation	67	28	29.5%		96	0	0%		96	0	0%		-	-	-	-

Note: aAll the laboratory variables were preoperatively determined.

Abbreviations: MSS, microsatellite stability; MSI, microsatellite instability; CIMP, CpG island methylator phenotype; BMI, body mass index; CEA, carcinoembryonic antigen; CSR, CEA/tumor size ratio; NLR, neutrophil/lymphocyte ratio; PLR, platelet/lymphocyte ratio.

Predictive Variables for MSI

In our cohort, the characteristics of patients with MSI and microsatellite stability (MSS) are similar except for tumor differentiation and location. MSI was more frequent in poorly differentiated CRCs [39.6% (19/48) vs 22.4% (52/232), P=0.013] and proximal CRCs [36.5% (19/52) vs 22.7% (53/233), P=0.039] (Table 1). Next, we performed logistic regression analyses to identify the clinicopathological variables that predict MSI in CRC. In the univariate analysis, tumor differentiation, location and NLR were significantly associated with MSI (Table 2). These factors were entered into a multivariate analysis, in which poor differentiation (OR=2.392, 95% CI: 1.213–4.715; P=0.012) and high NLR (OR=3.893, 95% CI: 1.14–13.293; p=0.030) were independently associated with MSI status (Table 3).

Table 2

Predictive Factors for Molecular Phenotypes in Univariate Logistic Regression Analysis

Molecular Subtypes	Variablea		P	OR	CI 95%
MSI	Tumor location	Proximal	0.041	1.955	1.029–3.717
		Non-proximal		1
	Differentiation	Poor	0.014	2.268	1.177–4.369
		Moderate-well		1
	NLR	High	0.026	3.988	1.177–13.510
	Low		1
CIMP	Differentiation	Poor	0.002	29.040	3.421–246.524
		Moderate-well		1
	Lymphovascular invasion	+	0.003	10.350	2.166–49.463
		-		1
	Platelet (109/L)	>300	0.022	5.987	1.300–27.577
		≤ 300		1
	NLR	High	0.008	17.746	2.100–149.938
		Low		1
	PLR	High	0.050	5.250	0.999–27.582
		Low		1
	CSR	High	0.015	6.696	1.450–30.923
	Low		1
BRAF	Lymphovascular invasion	+	<0.001	15.529	4.095–58.899
		-		1
	Differentiation	Poor	<0.001	12.356	3.077–49.625
		Moderate-well		1
	CEA(ng/mL)	≥ 5	0.015	5.016	1.371–18.353
		<5		1
	PLR	High	0.042	4.175	1.055–16.524
		Low		1
	CSR	High	0.002	8.325	2.248–30.829
	Low		1
KRAS	Differentiation	Poor	0.637	1.168	0.612–2.230
		Moderate-well		1
	Tumor location	Proximal	0.027	1.995	1.081–3.681
		Distal		1
	Histology	Mucinous	0.027	2.371	1.103–5.098
		Non-mucinous		1
	NLR	High	0.013	1.937	1.149–3.267
	Low		1

Notes: aAll the laboratory variables were preoperatively determined. Only predictive factors with statistical significance were presented in this table. The cutoff of each variable determined by ROC can be found in .

Abbreviations: MSI, microsatellite instability; CIMP, CpG island methylator phenotype; CEA, carcinoembryonic antigen; CSR, CEA/tumor size ratio; NLR, neutrophil/lymphocyte ratio; PLR, platelet/lymphocyte ratio.

Table 3

Predictive Factors for Molecular Phenotypes in Multivariate Logistic Regression Analysis

Molecular Subtypes	Variablea		P	OR	CI 95%
MSI	Differentiation	Poor	0.012	2.392	1.213–4.715
		Moderate-well		1
	NLR	High	0.030	3.893	1.140–13.293
	Low		1
CIMP	Differentiation	Poor	0.004	28.373	2.961–271.921
		Moderate-well		1
	NLR	High	0.020	14.518	1.526–138.108
		Low		1
	CSR	High	0.047	6.230	1.023–37.959
	Low		1
BRAF	Differentiation	Poor	0.005	9.447	1.937–46.071
		Moderate-well		1
	Lymphovascular invasion	+	0.005	10.861	2.043–57.727
		-		1
	CSR	High	0.002	14.350	2.718–75.753
	Low		1
KRAS	Differentiation	Poor	0.022	0.164	0.035–0.771
		Moderate-well		1
	Tumor location	Proximal	0.013	2.351	1.202–4.598
		Distal		1
	Histology	Mucinous	0.005	11.651	2.119–64.074
		Non-mucinous		1
	NLR	High	0.015	1.983	1.144–3.438
	Low		1

Notes: aAll the laboratory variables were preoperatively determined. The cutoff of each variable determined by ROC can be found in .

Abbreviations: MSI, microsatellite instability; CIMP, CpG island methylator phenotype; CSR, carcinoembryonic antigen/tumor size ratio; NLR, neutrophil/lymphocyte ratio.

Predictive Variables for CIMP

A CIMP (+) status was more frequent in CRCs characterized as older patients [4.1%(6/147) vs 0.6%(1/159), P=0.043], larger size [4.9%(7/143) vs 0.0%(0/160), P=0.014], poor differentiation [10.7%(6/56) vs 0.4%(1/243), P<0.001], lymphovascular invasion [13.0%(3/23) vs 1.4%(4/280), P=0.004], and elevated CA125 [15.0%(3/20) vs 1.5%(4/239), P=0.003] (Table 1). To identify the clinicopathological predictors for CIMP (+) in CRC, we performed Logistic regression analyses. A CIMP (+) status was found to be associated with poor differentiation, lymphovascular invasion, platelet counts, NLR, PLR and CSR in the univariate analysis (Table 2), while only the association with poor differentiation (OR=28.373, 95% CI: 2.961–271.921; P=0.004), NLR (OR 14.518, 95% CI: 1.526–138.108; P=0.020), and CSR (OR=14.350, 95% CI: 2.718–75.753; P=0.047) were still significant in multivariate analysis (Table 3).

Predictive Variables for BRAF Mutation

BRAF mutation was more frequent in CRCs with poor differentiation [13.7% (7/51) vs 1.3% (2/235), P<0.001], lymphovascular invasion [22.7% (5/22) vs 1.9% (5/269), P<0.001], elevated CEA [9.0% (6/67) vs 1.9% (4/208), P=0.022], elevated CA19-9 [12.8% (5/39) vs 2.2% (5/231), P=0.005], and elevated CA125 level [20.0% (4/20) vs 2.4% (6/249), P=0.001] (Table 1). Next, we performed Logistic regression analyses to identify predictors for BRAF mutation from clinicopathological variables. The predictors that was significant in the univariate analysis, including poor differentiation, lymphovascular invasion, CEA, NLR, PLR and CSR (Table 2), were entered into a multivariate analysis, in which poor differentiation (OR=9.447, 95% CI: 1.937–46.071; P=0.005), lymphovascular invasion (OR=10.861, 95% CI: 2.043–57.727; P=0.005), and high CSR (OR=14.350, 95% CI: 2.718–75.752; P=0.002) were independently associated with BRAF mutation (Table 3).

Predictive Variables for KRAS Mutation

KRAS mutation was more frequent in patients with proximal tumors [48.1% (35/52) vs 31.7% (71/224), P=0.025], mucinous tumor [56.7% (17/30) vs 32.1% (79/46), P=0.008], and high platelet counts [0.0% (0/4) vs 38.7% (79/204), P=0.015], while other characteristics were similar between KRAS wild-type and mutant patients (Table 1). Next, we performed logistic regression analyses to identify the clinicopathological predictors for KRAS mutation in CRC. In the univariate analysis, poor differentiation, proximal tumor, mucinous tumor and NLR were significant predictors for harboring KRAS mutation (Table 2). The further multivariate analysis showed all these variables, including poor differentiation (OR=0.164, 95% CI: 0.035–0.771; P=0.022), proximal tumor (OR=2.351, 95% CI: 1.202–4.598; P=0.013), mucinous tumor (OR=11.651, 95% CI: 2.119–64.074; P=0.005), and high NLR (OR=1.983, 95% CI: 1.144–4.438; P=0.015), were independently associated with KRAS mutation (Table 3).

Predictive Nomograms Established for MSI, CIMP, BRAF and KRAS Mutation

Four Nomograms were developed based on the independently significant factors in the multivariate logistic regression analysis (Figure 2, left). The nomogram for predicting MSI status was a model in which NLR weighted more than differentiation. Tumor differentiation weighted most, and NLR and CSR were followed in the nomogram for predicting CIMP (+). The nomogram for predicting BRAF mutation included predictors similar to that for CIMP (+), except for NLR replaced by lymphovascular invasion. These three predictors weighted similar in this model. In the nomogram for predicting KRAS mutation, the histological features of differentiation and mucinous tumor showed a superior impact on the prediction over proximal location and NLR. Using these nomograms, we could easily calculate the probability of MSI, CIMP (+), BRAF mutation and KRAS mutation based on clinicopathological information.

Figure 2

Nomograms and calibration curves for predicting the probability of (A) MSI, (B) CIMP (+), (C) BRAF mutation and (D) KRAS mutation. The predicted and observed probabilities of MSI, CIMP (+), BRAF mutation and KRAS mutation were shown in the calibration curves.

Abbreviations: NLR, neutrophil/lymphocyte ratio; PLR, platelet/lymphocyte ratio; CSR, CEA/tumor size ratio.

We further used 1000 bootstrap resamples to compute adjusted C-indexes. The C-indexes of MSI, CIMP (+),BRAF mutation and KRAS mutation were 61.22% (95% CI: 60.28–62.16%), 95.57% (95% CI: 95.20–95.94%), 83.56% (95% CI: 81.54–85.58%), and 69.12% (95% CI: 68.30–69.94%) respectively. Calibration curves between predicted and actual observations by internal validation demonstrated that these nomograms showed good statistical performance for predicting the probability of each phenotype, except for the nomograms for MSI and CIMP (+), in which the probability of MSI would be overestimated when the probability was less than 0.2 (Figure 2, right).

Discussion

In this study, we identified the independent predictors for MSI, CIMP (+), BRAF mutation and KRAS mutation. Among these predictors, NLR and PLR as the systemic inflammation markers, and CSR as a tumor size-corrected CEA indicator have not been reported to be associated with any of molecular phenotypes so far. To the best of our knowledge, this is the first study exploiting them in models to predict molecular phenotypes. We constructed four nomograms using these independent predictors, and their internal validations showed good statistical performance to predict molecular phenotypes. Considering the significance of MSI, CIMP (+), BRAF mutation and KRAS mutation in currently clinical decision-making, the nomograms we generated that could predict molecular phenotypes using easily accessible clinicopathological variables would be widely used in clinical practice.

The missense mutations in KRAS occur in approximately 37.5–38% CRCs in Chinese populations.22,34 A similar sequencing result was found in our cohort, in which KRAS mutation presented in 34.8% (96/276) patients with CRC. KRAS mutation has been found to be more likely to present in female, older patients, and tumors with right-side location, poor differentiation, elevated CEA or CA19-9, and high albumin/globular protein.17,28 In our study, we found similar results in the association analysis with poor differentiation and proximal tumor. We also identified high systemic inflammation status (high NLR) as an independent predictor for KRAS mutation. The preference to developing KRAS mutations in high-NLR CRC supports the recent findings that inflammatory signaling plays a critical role in promoting KRAS-driven oncogenesis through the interaction with autophagy and MAPK signaling.35

It has been reported that BRAF mutation presented in approximately 10–15% CRCs in Western cohort.36 However, several studies showed that BRAF mutation was only found in 2.8–4.4% CRCs in Chinese population.22,34 In our study, BRAF mutation presented in 3.4% (10/293) cases, which is accordant to the reported mutation rate in Chinese population. These results showed that there may exist a distinct nature of CRC between populations. The previous studies have reported various predictors for BRAF mutation, including elderly female patients and tumors characterized as right-sided, mucinous and poor differentiation.17,22,37 In our study, poor differentiation, lymphovascular invasion and high CSR were independent predictors for BRAF mutation. The distinct BRAF-mutation epidemiology and genetic basis between our population and previous cohort may contribute to the variation in predictors. The developed nomogram using these variables showed a high predictive accuracy up to 83.56%. As shown in the calibration curve, nomogram-predicted probability of status also fitted well with actual molecular status. This nomogram showed good statistical performance for predicting the probability of BRAF mutation.

It has been shown in both our cohort (Table 1) and previous report16,38 that CIMP (+) is tightly associated with BRAF mutation. Since CIMP (+) was reported to represent about 15% of CRCs in western population,39 it is not surprising that CIMP (+) incidence in our study, similar to BRAF mutation frequency, is lower than that in the previous report (2.7% versus 15%). Some retrospective studies have described the clinical features associated with CIMP (+) CRCs, including proximal tumor, elderly females, poor differentiation and mucinous tumor.16 In consistent with this study, poor differentiation was also independently associated with CIMP (+) status in our study. Moreover, high NLR and high CSR were independent predictors for CIMP (+) status as well. We built a nomogram showing good statistical performance for predicting CIMP (+) using these three independent predictors. However, this nomogram could only predict tumor with low risk of CIMP (+). This might result from low CIMP (+) incidence in our cohort.

Approximately 5% to 25% of sporadic CRCs develop with the defects in DNA mismatch repair (MMR) system.39–41 Similarly, MSI presented in 25.3% (72/285) patients in our cohort. MMR deficiency leads to MSI in cancer cells, which is the second most common pathway for CRC development. According to previous studies, the CRCs with MSI have distinct features, including right-sided tumor, poor differentiation, abundant tumor-infiltrating lymphocytes and less aggressive clinical course.18,34,42 It has been demonstrated that MSI has high sensitivity as the screening test to identify individuals with Lynch syndrome.43 Our nomogram for MSI, thus, may provide useful information for primary physicians to identify this subgroup of hereditary cancers. Models for predicting the presence of MSI-H status has been built. Jenkins et al developed the MsPath model in 2007.15 However, this model is only applied to patients diagnosed before the age of 60 years. In addition, Angela Hyde et al developed a histology-based model for predicting MSI in 2010.18 Unfortunately, popular use of this model would be limited by its predictors that need to be evaluated by experienced pathologists. In current study, we identified NLR as an independent predictor for MSI, which could be easily used and provided valuable information in practice. However, there were only two independent predictors in this model, and the generated nomogram using differentiation and NLR did not perform well for the prediction.

The robustness of this study includes the high quality-control in molecular assays, strict patient selection to eliminate the confounding influence on molecular phenotyping, and reliable statistical workflow to construct nomograms using continuous and categorized variables. However, this study has some limitations. First, the statistical power of the results in CIMP and BRAF mutation was limited by their low incidences in our population. Second, the sample size of stage-IV patients was small, and thus the nomograms need to be further trained and validated in a cohort with sufficient stage-IV cases to make them can be applied to stage-IV CRC. Moreover, patients included in our study were from a single institution. As a result, there may exist a variation of predictive ability of models among institutions, and an external validation set would be useful to validate our predictive models.

In conclusion, we established four models with easily obtained variables to predict the probability of MSI, CIMP (+), BRAF mutation and KRAS mutation. The nomograms should not replace the molecular laboratory tests of CRC, but it could allow physicians to speculate molecular subtypes of CRCs, then better estimate patients’ prognosis where genetic testing is not available or reimbursed because of infrastructure limits.