A transcriptomic signature that predicts cancer recurrence following hepatectomy in patients with colorectal liver metastases

Abstract

Background:

Cancer recurrence is an important predictor of survival outcomes in patients with colorectal cancer associated liver metastasis (CRLM), who undergo radical hepatectomy. Therefore, identification of patients with the greatest risk of recurrence is critical for developing a precision oncology strategy that might include frequent surveillance (in low-risk patients) or a more aggressive treatment approach (in high-risk patients). We performed genome-wide expression profiling, to identify and develop a transcriptomic signature for predicting recurrence in patients with CRLM.

Methods:

We analyzed a total of 383 CRLM patients, including 63 patients from a publicly available dataset (GSE81423) and 320 patients from whom surgical specimens were collected for independent training (n=169) and validation (n=151) of identified biomarkers. Using Cox’s proportional hazard regression analysis, we evaluated the clinical significance of the identified gene signature by comparing its performance with several key clinical factors.

Results:

We identified a 6-gene panel that robustly categorized patients with recurrence in the discovery (AUC=0.90). We showed that the panel was a significant predictor of recurrence in the clinical training (AUC=0.83) and validation cohorts (AUC=0.81). By combining our panel with key clinical factors, we established a risk-stratification model that emerged as an independent predictor of recurrence (AUC=0.85; Univariate: HR=4.34, 95% CI=2.71–6.93, P<0.001; Multivariate: HR=3.40, 95% CI=1.76–6.56, P<0.001). The stratification model revealed recurrence prediction in 89% of high-risk group and non-recurrence in 62% of low-risk group.

Conclusions:

We established a novel transcriptomic signature that robustly predicts recurrence, which has significant implications for the management of patients with CRLM.

INTRODUCTION

Colorectal cancer (CRC) is the second most common cause of cancer-related deaths worldwide [1, 2]. Management of patients with distant metastases is a major challenge: 20–30% of patients with CRC have distant metastatic disease at the time of diagnosis, and 5-year overall survival (OS) in these patients with metastatic CRC is less than 10% [3–5]. Approximately 50% of patients with CRC develop liver metastases after treatment of the primary tumor, and liver metastases is the most frequent cause of mortality in these patients [6–10]. Hepatectomy is the current treatment standard for colorectal liver metastases (CRLM) and can offer lead to prolonged survival, with a 5-year OS of 30–50% [11–14]. However, in spite of advances in surgical techniques, imaging modalities, and post-operative management, 50–70% of patients still experience cancer recurrence following hepatectomy, and most such recurrences occur within first 2 years of surgery [8, 15–18]. Consequently, due to these frequent recurrence events, the post-operative treatment options available to these patients are often complicated. The National Comprehensive Cancer Network guidelines suggest several selections of the treatment including surveillance with or without adjuvant chemotherapy in patients with resectable CRLM. Unfortunately, there are no randomized studies available that compared recurrence in patients of surgery or surgery with chemotherapy; underscoring the need for precise biomarkers that can enable a more personalized treatment strategy.

Several clinical trials have demonstrated that the chemotherapy improves survival outcomes after hepatectomy in patients with CRLM. For example, perioperative FOLFOX (oxaliplatin, leucovorin, and fluorouracil) chemotherapy led to a 7% improvement in progression-free survival (PFS) when followed after hepatectomy for CRLM [19, 20]. Likewise, other studies evaluating the addition of targeted therapies such as anti-EGFR antibodies in metastatic CRC also noticed improved outcomes [21, 22]. On the same lines, inclusion of anti-VEGF antibodies to a modified FOLFOX or FOLFIXIRI regimen (fluorouracil, leucovorin, oxaliplatin, and irinotecan) associated with improved response and resection rates and prolonged PFS in patients with CRLM [23–25]. These findings indicate that patients with CRLM who have a higher risk of recurrence benefit from such interventions and should be offered such intensive treatments following hepatectomy. Regarding post-surgical surveillance alone, serological carcinoembryonic antigen (CEA) level measurements, and the use of computed tomography (CT) imaging of the chest, abdomen, and pelvis every 3–6 months for the first 2 years and thereafter every 6 months for 3 more years are recommended. However, these criteria are often inadequate in high-risk patients, given that some patients need more aggressive surveillance and chemotherapy considering that more than half of patients experience tumor recurrence in 2 years of surgical treatment.

Recently, several clinical factors have been identified in patients with CRLM that predict the risk of cancer recurrence following hepatectomy [26–30]. Fong score, which combines CEA levels, tumor number, tumor size, lymph node positivity, and recurrence-free survival (RFS), is often used for predicting the recurrence after hepatectomy in patients with CRLM [31]. However, instead of or in addition to clinical factors, unravelling the molecular properties that characterize tumors may be pivotal for tailoring individualized therapies based on molecular predictors of survival outcomes. To this end, various groups have reported molecular subtypes in cancers with distinct biological properties and corresponding predictive and prognostic values [18, 26–30, 32–35]. Genetic and morphological evaluation scores, which combine clinicopathological and clinically-available biological indicators, including KRAS mutation status, have been reported to be a preoperative prognostic tool that can be used for treatment selection [36]. There is a growing consensus that molecular markers can improve the selection of patient for chemotherapy and other treatments in addition to hepatectomy or surveillance. The ability to analyze tumors at the RNA level promises to revolutionize the understanding of malignant disease, and we hope this will usher in a new biomarker.

Therefore, in this study, we conducted a comprehensive discovery effort to distinguish gene expression signatures for predicting cancer recurrence after hepatectomy in CRLM patients. We initially identified this signature in a genome-wide expression profiling dataset. Our risk-stratification model allows robust identification of patients with CRLM who are at high risk of cancer recurrence, which could improve the management of patients by recommending more intensive treatment selection in high-risk patients and sparing other low-risk patients from the unnecessary toxicity associated with adjuvant treatments.

METHODS

Gene biomarker discovery

To perform a comprehensive genome-wide biomarker discovery, gene expression profiling results from a dataset (GSE81423) were analyzed to identify and establish a biomarker panel for predicting recurrence following radical hepatectomy in patients with CRLM. The expression data was obtained from 63 patients with CRLM (33 recurrence and 30 non-recurrence), as illustrated in Fig. S1. We evaluated the prediction of recurrence by using selected biomarkers and subsequently determined area under the curve (AUC) values for each of the receiver operator characteristic (ROC) plots [37, 38].

Patient cohorts

To perform clinical training and validation of the identified gene panel, a total of 320 formalin-fixed paraffin-embedded (FFPE) specimens from two independent cohorts of resectable patients with CRLM were analyzed. This included a training cohort (n = 169; 119 with recurrence and 50 non-recurrence) of patients enrolled at the Tokushima and Kyushu University, and a validation cohort (n = 151; 102 with recurrence and 49 non-recurrence) of patients enrolled at the Kumamoto University, Japan. All patients underwent an initial radical hepatectomy between January 1995 and December 2017 in the training cohort, and between January 2001 and December 2016 in the validation cohort. The specimens were diagnosed as CRLM by pathologists. The follow-up protocol included abdominal ultrasonography, CT, and colonoscopy, evaluation of carbohydrate antigen 19-9 (CA19-9) and CEA levels, as well as a physical examination. Only regions confirmed by imaging modalities were diagnosed as tumor recurrence sites. The current study was conducted in accordance with the Declaration of Helsinki and approved by the institutional review boards of all participating institutions. A written informed consent was obtained from all patients.

RT-qPCR and RNA extraction

RNA was extracted from 10 μm sections of FFPE surgical resected tissues. Synthesis of complementary DNA (cDNA) was conducted with 500 ng of RNA. Real-time quantitative reverse transcription polymerase chain reaction (RT-qPCR) analysis was performed as we previously reported [39]. Normalized expression values were log₁₀-transformed [40]. The primers used in the current study are shown in Table S1.

Statistics

Clinicopathologic characteristics are shown in Table 1. OS and RFS were calculated from the date of surgery to the date of death from any cause or recurrence, or last follow-up date. Hazard ratios (HR) were calculated with 95% confidence intervals (CI) as we previously described [39]. Statistical analyses were performed using JMP Pro V 13.0 statistical software (SAS Institute Japan, Tokyo, Japan), Medcalc statistical software V.17.4.0 (Medcalc Software bvba, Ostend, Belgium).

Table 1:

Clinicopathological characteristics of clinical cohorts

Characteristics		Training cohort (n=169), n (%)	Validation cohort (n=151), n (%)	P
Age (years)	Median (range)	65 (33–92)	64 (25–94)	0.113
Gender	Male	114 (67.5)	101 (66.9)
	Female	55 (32.5)	50 (33.1)	0.914
Recurrence status	Recurrence	119 (70.4)	102 (67.5)
	Non-recurrence	50 (29.6)	49 (32.5)	0.58
Syn / Meta	Synchronous	96 (56.8)	98 (64.9)
	Metachronous	73 (43.2)	53 (35.1)	0.139
Tumor number	Solitary	83 (49.1)	62 (41.1)
	Multiple	85 (50.3)	89 (58.9)
	Missing	1 (0.6)	0 (0)	0.135
Tumor size (mm)	Median (range)	25 (5–100)	28 (3–160)	0.781
CEA (ng/ml)	Median (range)	9.8 (0.5–1430)	7.1 (0.9–2061)	0.201
CA19-9 (U/ml)	Median (range)	19.0 (0.6–6140)	19.5 (0.1–1756)	0.732
Fong score	High score (≥3)	44 (26.0)	42 (27.8)
	Low score (<3)	125 (74.0)	109 (72.2)	0.72
Neoadjuvant chemotherapy	Yes	92 (54.5)	83 (55.0)
	No	76 (44.9)	68 (45.0)
	Missing	1 (0.6)	0 (0)	0.971
Adjuvant chemotherapy	Yes	111 (65.7)	86 (57.0)
	No	57 (33.7)	63 (41.7)
	Missing	1 (0.6)	2 (1.3)	0.126
Tumor location of primary site	Right side	33 (19.5)	31 (20.5)
	Left side	69 (40.8)	64 (42.4)
	Rectum	67 (39.7)	56 (37.1)	0.895
Differentiation of primary site	Well	72 (42.6)	66 (43.7)
	Moderately	79 (46.7)	74 (49.0)
	Poor	6 (3.6)	1 (0.7)
	Others	3 (1.8)	0 (0)
	Missing	9 (5.3)	10 (6.6)	0.121

CA19-9, Carbohydrate antigen 19-9; CEA, Carcinoembryonic antigen;

RESULTS

Genome-wide expression profiling identifies a 6-gene panel that predicts cancer recurrence

To identify a gene panel that can predict cancer recurrence in CRLM patients, we first performed a genome-wide discovery in a public gene dataset from a cohort of patients with CRLM (GSE81423). By comparing gene profiles between patients with and without recurrence after radical hepatectomy, we identified a panel of 18 candidate genes differentially expressed (P < 0.05, |log₂ fold-change| >1). Next, we performed lasso regression analysis with cross validation, which reduced the candidates to a panel of 6 genes: COX6A1, ERN1, IFITM2, S100P, STK24, and TMTC3. Applying the 6-gene panel to the GSE81423 data resulted in an AUC of 0.90 (95% CI = 0.79–0.96: Fig. S2A), highlighting the diagnostic potential of this panel for identifying recurrence in patients with CRLM (Fig. S2B).

Clinical cohort established the transcriptomic panel that predicts cancer recurrence

To confirm the predictive our discovered transcriptomic panel, we evaluated the performance in two clinical cohorts of patients with CRLM who underwent radical hepatectomy. The training cohort consisted of 169 patients, which included 119 patients with recurrence (70.4%), and the validation cohort consisted of 151 patients, which included 102 patients with recurrence (67.5%) with a median age of 64 years. According to the chemotherapy regimen, 92 patients in the training and 83 patients in the validation cohort received neoadjuvant chemotherapy, and 111 patients in the training and 86 patients in the validation cohort received adjuvant chemotherapy (Table 1). According to the detailed chemotherapy regimen, perioperative chemotherapy was performed in 254 of the 320 patients with CRLM. The details of these treatment are as follows: FOLFOX, 80 cases; XELOX, 51 cases; LV/UFT, 34 cases; FOLFIRI, 28 cases; oral 5FU, 18 cases; oral capecitabine, 10 cases; IRIS, 9 cases; SOX, 7 cases; LV/5FU, 7 cases and others, 10 cases.

Next, we interrogated the diagnostic performance of our panel to predict recurrence in the training cohort (119 with recurrence vs. 50 with non-recurrence). We trained the panel that robustly identified recurrence in patients with CRLM (AUC = 0.83, 95% CI = 0.76–0.88, Fig. 1A, B). Thereafter, we developed a transcriptomic panel by using parameters: Logit (P) = (COX6A1*0.1442) + (ERN1*−0.4654) + (IFITM2*−0.2096) + (S100P*−0.2334) + (STK24*−0.1346) + (TMTC3*−0.2320).

Figure 1.

The transcriptomic panel for identifying cancer recurrence. A) ROC curve for the panel in the training cohort of 169 patients (with recurrence = 119, non-recurrence = 50, AUC = 0.83). B) Risk score plot in the training cohort. C) ROC curve for the panel in the validation cohort of 151 patients (with recurrence = 102, non-recurrence = 49, AUC = 0.81). D) Risk score plot in the validation cohort.

We evaluated the accuracy of 6-gene transcriptomic panel by using same statistical model (i.e., same coefficients and cutoff values) to another validation cohort of patients with CRLM (102 with recurrence, and 49 non-recurrence). The diagnostic performance of our panel in the validation cohort was comparable in the training cohort (AUC = 0.81, 95% CI = 0.74–0.87: Fig. 1C, D).

Recurrence prediction correlates with survival outcomes

The recurrence is correlated to poor survival in CRLM patients. Thus, to determine the prognostic potential of the transcriptomic panel, we analyzed the survival outcomes (OS and PFS) in the clinical training and validation cohorts. The median follow-up period after hepatectomy were 39.7 months (95% CI = 34.5–44.4) in training and 37.1 months (95% CI = 26.9–47.8) in validation cohort. It was demonstrated that the patients with high-risk exhibited significantly poor prognosis in the training [P < 0.001]; OS [P = 0.029]; Fig. 2A, B), and validation cohorts (RFS [P < 0.001]; OS [P = 0.006]; Fig. 2C, D). We performed an univariate Cox’s proportional hazard regression analysis, which identified that our panel was an independent predictor of recurrence, compared to each clinical factors (training: HR = 3.18, 95% CI = 2.03–4.98, P < 0.001; validation: HR = 2.60, 95% CI = 1.73–3.92, P < 0.001; Table 2).

Figure 2.

The transcriptomic panel of a prognostic potential. (A) RFS and (B) OS between low- and high-risk groups of the transcriptomic panel in the training cohort. (C) RFS and (D) OS between low- and high-risk groups of the transcriptomic panel in the validation cohort.

Table 2.

Univariate and multivariate analysis for the recurrence free survival

	Univariate analysis			Multivariate analysis
Factors	HR	95% CI	p	HR	95% CI	p
Training cohort (n=169)
Age
(≥65 vs. <65)	0.85	0.59–1.21	0.365
Gender
(Male vs. Female)	0.95	0.64–1.39	0.783
Syn / Meta
(Syn vs. Meta)	1.47	1.01–2.12	0.042	1.13	0.77–1.65	0.554
Tumor number
(Multi vs. Solitary)	1.99	1.38–2.88	0.005	1.92	1.33–2.79	0.001
Tumor size
(≥25mm vs. 25mm)	1.62	0.97–2.71	0.066
CEA
(≥5.0ng/mL vs. <5.0ng/mL)	1.33	0.88–2.00	0.171
CA19-9
(≥37U/mL) vs. <37U/mL)	1.30	0.89–1.91	0.177
Fong score
(≥3 vs. <3)	1.60	0.96–2.80	0.071
Neoadjuvant chemotherapy
(Yes vs. No)	1.18	0.73–1.91	0.492
Adjuvant chemotherapy
(Yes vs. No)	1.55	1.03–2.31	0.034	1.22	0.80–1.85	0.355
RNA panel
(High risk vs. Low risk)	3.18	2.03–4.98	<0.001	2.87	1.82–4.53	<0.001
Validation cohort (n=151)
Age
(≥65 vs. <65)	0.99	0.97–1.04	0.392
Gender
(Male vs. Female)	0.83	0.55–1.24	0.364
Syn / Meta
(Syn vs. Meta)	1.78	1.17–2.72	0.008	1.28	0.81–2.01	0.291
Tumor number
(Multi vs. Solitary)	1.96	1.29–2.98	0.002	1.09	0.70–1.72	0.699
Tumor size
(≥25mm vs. 25mm)	1.42	0.95–2.13	0.085
CEA
(≥5.0ng/mL vs. <5.0ng/mL)	1.60	1.05–2.42	0.027	2.24	1.43–3.51	<0.001
CA19-9
(≥37U/mL) vs. <37U/mL)	1.59	1.05–2.39	0.027	0.97	0.62–1.51	0.882
Fong score
(≥3 vs. <3)	1.45	0.96–2.32	0.054
Neoadjuvant chemotherapy
(Yes vs. No)	1.39	0.94–2.17	0.061
Adjuvant chemotherapy
(Yes vs. No)	0.78	0.52–1.16	0.223
RNA panel
(High risk vs. Low risk)	2.60	1.73–3.92	<0.001	1.85	1.17–2.93	0.009
Risk model
(High risk vs. Low risk)	4.34	2.71–6.93	<0.001	3.65	2.02–6.59	<0.001

CA19-9, Carbohydrate antigen 19-9; CEA, Carcinoembryonic antigen; Meta, Metachronous; Syn, Synchronous

Risk model by combining the panel and clinical factors improves to predict recurrence

We evaluated whether the stratification model that couples known risk factors with the panel improve the accuracy in predicting recurrence. Indeed, once we incorporated four conventional clinicopathological factors (i.e., tumor number, synchronous lesions, CA19-9, and CEA,) into the stratification model, the model revealed a markedly better diagnostic performance compared with the only panel or other risk factors (AUC = 0.85, Fig. 3A).

Figure 3.

The risk-stratification model in the validation cohort. A) ROC curves for the model by combining the panel and clinical factors, vs. the transcriptomic panel or indicated risk factors in the validation cohort. B–C) Forest plot showing HRs of indicated clinicopathological factors, the panel, and the stratification model in univariate (B) and multivariate (C) analysis. D) RFS between high- (Red) and low- (Blue) risk groups by the risk-stratification model. E) Bar graphs shows 61.9% of low-risk patients with CRLM who would have experienced non-recurrence, and 88.6% of high-risk patients with CRLM who would have experienced recurrence.

We evaluated the diagnostic accuracy for our panel in validation. Therefore, we revealed that the sensitivity, specificity, positive predictive value (PPV), and negative predictive values (NPV) were 59.8%, 83.7%, 88.4%, and 50.0%, respectively (Table 3). When we performed the stratification model that also included risk factors, the model was significantly superior to the transcriptomic panel alone in predicting sensitivity, specificity, PPV, and NPV, with values of 77.2%, 81.6%, 89.7%, and 63.5%, respectively. Herein, our novel model by univariate and multivariate Cox’s regression analysis was better than the panel and various clinicopathological factors for predicting recurrence (Univariate: HR = 4.34, 95% CI = 2.71–6.93, P < 0.001, Fig. 3B; Multivariate: HR = 3.65, 95% CI = 2.02–6.59, P < 0.001, Fig. 3C) in validation (Table 2). These data highlight the potential significance of the model for identifying recurrence in CRC patients with liver metastasis.

Table 3:

Model performance for the risk of cancer recurrence

Variable	Value (95% CI)
	GSE81423	Training cohort	Validation cohort (RNA panel)	Validation cohort (Risk model)
Cutoff value	-	1.44	1.44	2.13
Sensitivity, %	90.9 (75.7–98.1)	79.0 (70.6–85.9)	59.8 (49.6–69.4)	77.2 (67.8–85.0)
Specificity, %	76.7 (57.7–90.1)	72.0 (57.5–83.8)	83.7 (70.3–92.7)	81.6 (68.0–91.2)
AUC, %	89.5 (79.2–95.8)	82.5 (75.9–87.9)	81.1 (74.0–87.0)	85.0 (78.2–90.3)
PPV, %	81.1 (68.9–89.2)	87.0 (81.0–91.4)	88.4 (79.9–93.6)	89.7 (82.6–94.0)
NPV, %	88.5 (71.9–95.8)	59.0 (49.4–68.0)	50.0 (43.4–56.6)	63.5 (54.3–71.8)

AUC, area under the curve; CI, confidence interval;

NPV, negative predictive value; PPV, positive predictive value

Risk-stratification model robustly identifies CRLM patients at high risk for cancer recurrence

To determine the clinically usefulness of our model, we applied our risk-stratification model to dichotomize the patients within the validation cohort into high- and low-risk groups of patients for risk of cancer recurrence. We observed that the RFS was significantly lower in the low- vs. high-risk patients by our model (P < 0.001; Fig. 3D). Of the 88 cases who were classified as high-risk patients, 78 cases had cancer recurrence in 5 years (88.6%) (Fig. 3E). Of the 63 cases who were classified as low-risk patients, 39 cases did not have tumor recurrence in 5 years (61.9%), highlighting that these low-risk patients could avoid expense and toxicity from unnecessary treatment. These data highlight the clinical potential of our model to inform personalize treatment recommendations in patients with low- and high-risk CRLM.

DISCUSSION

Accumulating studies continue to emphasize the importance of identifying clinical predictive markers for recurrence in CRC patients with liver metastasis (CRLM) [8, 15–18]. However, the specific clinical risk factors that can robustly predict cancer recurrence and long-term survival outcomes in patients with CRLM remains unclear. Clinically, CRLM rarely spreads via micro-metastases adjacent to the liver metastases. Thus, the width of a negative surgical margin following hepatectomy in patients with CRLM does not affect the risk of marginal recurrence or impact survival [41–43]. Furthermore, a recent study suggested that parenchyma-sparing hepatectomy did not increase recurrence in the remnant liver and improved salvageability for recurrent patients [44]. Considering this evidence in the context of our findings about cancer outcomes, our transcriptomic panel could identify high-risk CRC patients who require close follow-up surveillance and intensive chemotherapy to prevent subsequent recurrence.

Response to chemotherapy is recognized as one of the strong predictor of survival outcome after surgery, and provides better prognosis than traditional clinicopathological features [45]. However, several issues complicate the inclusion of response to chemotherapy in a clinical scenario. The best way to assess treatment response remains unclear. Histopathological scoring of responses to chemotherapy clearly predicts treatment outcome, but such assessments can only be made after surgery on resected specimens, and different groups use different scoring systems [46]. Previous data indicated that CA19-9, CEA, synchronous lesions, and tumor number were correlated to the tumor characteristics, a higher risk of tumor recurrence, and poor survival [47–50]. In the present study, our novel model revealed better diagnostic performance to predict cancer recurrence (AUC = 0.85), compared with clinical risk factors (i.e., synchronous, CA19-9, CEA, and tumor number; AUC = 0.68 [training] and 0.68 [validation]; Fig. S3). In addition, our model robustly stratified patients into low-and high-risk patients, indicating that 88.6% cases of high-risk patients had cancer recurrence and 61.9% cases of low-risk patients did not have cancer recurrence. This highlights dramatically better performance, compared with clinicopathological risk factors and the potential importance of our findings on clinical translation for identification of recurrence and improved survival in CRLM patients.

The various genes in the panel have been shown to be genuine candidates for the pathogenesis of cancer. For instance, IFITM2 is a member of IFN-inducible transmembrane gene family, which plays key roles in modulating the immune response and thus may control tumor development [51, 52]. IFITM2 has been reported to be highly specific to human colorectal carcinogenesis and overexpressed in CRC [53]. S100P is a member of the S100 family of proteins containing 2 EF-hand calcium-binding motifs, which regulates many intracellular and extracellular activities [54, 55]. S100P is overexpressed in CRC and regulates the invasion and metastasis of CRC by promoting epithelial–mesenchymal transition [56]. Furthermore, STK24, encoding a serine/threonine protein kinase, is a member of serine/threonine kinase family that functions upstream of mitogen-activated protein kinase (MAPK) signaling and positively regulates the cell cycle, cell growth, migration, and synapse development [57]. Similarly, the STK24 activity was increased in the immunoblot analyses of CRC, which revealed an increased amount of STK24 kinase in CRC [58].

We acknowledge that our study has some potential limitations. At first, current retrospective study might have resulted in a selection bias. So, to confirm the accuracy of our stratification model, a constructed prospective clinical trial is required. Next, our research used Japanese cohort in the training and validation, who showed relatively homogenous clinicopathologic characteristics; these could vary among populations from other countries. It is important to validate the selected biomarkers and the stratification model in patient cohorts in other countries to further enhance the generalization of findings. Third, unfortunately, we did not have the mutation data for the KRAS and BRAF genes for predicting the recurrence, future studies will be needed to explore these associations.

In conclusion, we established a new stratification model that was validated in clinical cohort for the identification of high-risk CRLM patients. Our current findings reveal the clinical impact of our stratification model on accurate risk-stratification of patients, which will improve personalized management of patients with CRLM.

Supplementary Material

1

Supplemental Figure S1. Overview of the study.

Supplemental Figure S2. Genome-wide discovery analysis of a gene panel to predict cancer recurrence. A) ROC curve for the diagnostic accuracy of a 6-gene panel for classifying patients with cancer recurrence in GSE81423 (AUC = 0.90). B) Risk score distribution plot in GSE81423.

Supplemental Figure S3. Prediction of recurrence in training and validation cohorts without the transcriptomic panel. A) ROC curve for combined clinical risk factors for recurrence (CA19-9, CEA, synchronous, and tumor number) without the transcriptomic panel in the training cohort (AUC = 0.68). B) ROC curve for the combined clinical risk factors for recurrence without the panel in the validation cohort (AUC = 0.68).

Highlights.

• Cancer recurrence is an important predictor of survival outcomes in CRLM.

• Currently, there are no available biomarkers to predict the risk of CRLM.

• We identified a gene signature by a genome-wide and comprehensive discovery.

• Our novel signature could predict the high-risk patients of cancer recurrence.

Abbreviations:

AUC

Area under the curve

CA19-9

Carbohydrate antigen 19-9

CEA

Carcinoembryonic antigen

CI

Confidence intervals

CRC

Colorectal cancer

CRLM

Colorectal liver metastases

CT

Computed tomography

FFPE

Formalin-fixed paraffin-embedded

HR

Hazard ratio

NPV

Negative predictive value

OS

Overall survival

PFS

Progression-free survival

PPV

Positive predictive value

RFS

Recurrence-free survival

ROC

Receiver operator characteristic

RT-qPCR

Real-time quantitative reverse transcription polymerase chain reaction