Nomogram Integrating Genomics with Clinicopathologic Features Improves Prognosis Prediction for Colorectal Cancer

Colorectal carcinoma is the third most commonly diagnosed malignant disease and the second leading cause of cancer death worldwide (1). Despite advances in colorectal cancer screening, diagnosis, and curative resection, its prognosis is not entirely satisfactory because optimal management and individual therapy strategies still present great challenges, as colorectal cancer is a well-recognized heterogeneous disease. Currently, treatment decisions and prognoses for patients with colorectal cancer are primarily driven by assessment of the tumor–node–metastasis (TNM) staging system, which is based merely on anatomical information (2). Previous studies revealed that patients with stage I colorectal cancer have a 5-year survival rate of approximately 93%, which decreases to approximately 80% for patients with stage II disease and to 60% for patients with stage III disease (3). However, discrepancies in the survival outcomes of patients at the same stage and receiving similar treatments are commonly observed.

More importantly, according to the current TNM stage, adjuvant therapy is generally recommended for all patients with stage III disease (4). However, patients with T1–2N1M0 tumors (stage IIIA) have significantly higher survival rates than those with stage IIB tumors (3), suggesting that stratifying patients at high risk of recurrence, who are most likely to benefit from adjuvant therapy, is critical. Therefore, substantial effort has been made to discover new clinicopathologic indicators to optimize the current staging system. To date, some clinicopathologic features, such as emergency presentation, adjacent organ involvement (T4), intestinal perforation or obstruction, high tumor grade, inadequate sampling of lymph nodes, and lymphatic/vascular invasion, with prognostic value to classify colorectal cancer patients who are at a high risk of recurrence have been applied in clinical practice. However, these factors are all relatively weak and insufficient to identify colorectal cancer patients who may benefit from adjuvant therapy, which leads to potential under-treatment or over-treatment (5).

Colorectal cancer biological behavior (e.g., recurrence, metastasis, drug resistance) is a tightly regulated process that requires the aberrant expression of related genes to empower carcinoma cells with their corresponding abilities. Accordingly, before clinicopathologically detectable changes occur, the underlying alterations necessary for recurrence have already occurred in the primary colorectal cancer, providing the possibility to develop robust prognostic tools by using multiple genes in combination (6). During the last decade, gene signatures have shown great promise in predicting the long-term outcomes and treatment responses of individual patients (7). Of note, because of the outstanding ability to predict the prognoses of patients with breast cancer, multigene assays, such as Oncotype DX and MammaPrint, have been successfully approved by the US Food and Drug Administration and are available in routine clinical practice (8). Actually, genomic signatures to predict the prognosis of colorectal cancer have also been continually developed in the past 10 years, but none are commercially used in the clinic (9). For example, Agesen and colleagues established a 13-gene expression classifier, ColoGuideEx, for prognosis predictions specific to patients with stage II colorectal cancer (10). Based on the essential role of lipid metabolism in carcinogenesis, Teodoro and colleagues identified a group of 4 genes that predict survival in intermediate-stage colon cancer patients, allowing the delineation of a high-risk group that may benefit from adjuvant therapy. In addition, Smith (11), Chen (12), Oh (13), and Popovici and colleagues (14) also published their own signatures based on different mechanisms. Despite a large number of studies, no signatures have remained credible in either meta-analyses or prospective trials (15). However, these published signatures clearly show low prediction accuracies but moderate clinical usefulness (15). Moreover, when confined to a specific colorectal cancer stage, promising results regarding risk stratification have also been reported. Therefore, investigating the predictive ability of these published signatures on comprehensive large-scale datasets and identifying whether any can be used to clinically guide treatment decisions is necessary.

Currently, while doubts about the predictive value of clinicopathologic features are increasing, they still provide the most reliable guidelines for the prognostication and treatment of colorectal cancer (16). Thus, we hypothesized that integrating genomic signatures with clinicopathologic features in a model would yield a predictive accuracy exceeding that of the currently available prognostic system. Nomogram is a statistical prediction model that combines multiple prognostic factors to make intuitive graphical and individualized predictions (17). Here, we aimed to apply a systematic approach to evaluate the clinical usefulness of colorectal cancer–related signatures and then construct a composite clinicopathologic–genomic nomogram by integrating factors with potential prognostic value in a training set. Moreover, using another independent set, the capacity of the nomogram to stratify colorectal cancer patients most likely to benefit from chemotherapy was further validated.

To identify gene expression data arrayed using the Affymetrix platform with clinically annotated data, we systematically searched Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/), ArrayExpress (http://www.ebi.ac.uk/arrayexpress/) and related literature with the keywords “colorectal cancer,” “colorectal cancer,” “colon cancer,” “survival,” “relapse,” “recurrence,” “prognostic,” “prognosis,” and “outcomes” published through August 1, 2017. For some datasets whose clinical data did not accompany their gene expression profiles, we either searched the supplements or contacted one or more of the investigators to obtain the missing information. Moreover, datasets with small sample sizes and duplications were excluded. Raw microarray data and the corresponding clinical data of these datasets were retrieved and manually organized when available. Only patients diagnosed with colorectal cancer having clinicopathologic and survival information available were included. Patients with follow-up or survival times of less than 1 month as well as patients with missing or insufficient data on age, local invasion, lymph node metastasis, distant metastasis, and TNM stage were excluded from subsequent processing. Eventually, all patients satisfying the inclusion criteria were combined and summarized in Supplementary Tables S1 and S2.

For raw CEL files available from Affymetrix microarrays, the data were normalized and annotated using a MAS5 algorithm and the corresponding annotation files from R Bioconductor to obtain summarized values for each probe set; otherwise, preprocessed data as provided by the contributors were used. For each sample in every data set, measurements without a gene annotation were excluded, and multiple probe sets corresponding to a single gene were summarized into a gene symbol by taking the most variable probe set measured by the interquartile range (IQR).

Gene signatures potentially related to colorectal cancer prognosis were systematically retrieved from PubMed. The search was restricted to recent papers to increase validity (from January 2004 to August 2017). The selection criteria are detailed in Supplementary Fig. S1. Articles that provided a list of differentially expressed genes in primary tumor samples associated with colorectal cancer prognosis were included in our study. Studies based on tissue microarray and those that were exclusively focused on differences between stages or between primary tumors and metastases were excluded. The signatures finally included in our analysis are described in Table 1 (detailed descriptions provided in Supplementary Table S3). In addition, the probe sets or genes of those signatures were re-annotated using the SOURCE web tool to address the retired gene symbols and their differences in the tested platforms. The re-annotated genes were then subjected to biological function enrichment analysis, and the online analytical tool DAVID (Database for Annotation, Visualization and Integrated Discovery; ref. 18) was used to enrich gene ontology (GO) functions and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. GO terms and KEGG pathways with significant enrichment false discovery rate (FDR) values less than 0.05 were selected for further analysis. In addition, genes from the abovementioned signatures were mapped and imported into the Retrieval of Interacting Genes/Proteins (STRING) 9.1 database, which queried the human protein–protein interaction network for interactions between effective linkers and seeds to construct a functional subnetwork. Cluster analyses were performed using correlation distance metrics and the average linkage agglomeration algorithm (R package nclust version 1.9.0). Nonnegative matrix factorization (NMF) was done with the NMF package (version 0.20.5) and standard strategies.

The re-preprocessed microarray dataset, which represents the genomic features of the individual, was classified with the prognostic signatures identified above using the nearest template prediction (NTP) method (19) as implemented in Gene Pattern software (Broad Institute of Harvard and MIT, Boston, MA; ref. 20). NTP requires only a list of prespecified template signature genes and a dataset to be tested and not a corresponding training dataset to capture good and poor gene expression patterns in each sample. Briefly, a template containing representative expression patterns of the signature genes was defined based on published gene signatures from their respective studies. The proximity of the signature gene expression patterns of the sample to the template was evaluated by calculating the cosine distance. The FDR was used to correct the P values for multiple hypothesis testing, and prediction analysis was performed separately for each dataset. A prediction of good or poor for the related signature was determined based on FDR predictions <0.05, and the remaining samples with intermediate expression levels of the correlated genes in the signature were classified as uncertain.

Concordance among these predictions was evaluated using unsupervised clustering according to the Cramer V coefficient of the paired prediction overlap as previously described (21). Cramer V statistic values range from 0 to 1, with values between 0.36 and 0.49 indicating substantial correlation, and values higher than 0.5 indicating strong correlation.

All the patients included in this study were randomly separated into training and validation sets. Disease-free survival (DFS) was defined as the interval between the day that surgery was performed and the day that recurrence was first detected. If recurrence was not diagnosed, the date of death due to colorectal cancer or last follow-up was used. Overall survival (OS) was calculated from the date of diagnosis or surgery to the date of death or last follow-up. Continuous variables were expressed as median (IQR) or median (range) values, and group comparisons were performed by the Student t test or the Wilcoxon rank sum test. Categorical variables were expressed as percentages, and group comparisons were performed by Pearson χ² test or the Fisher exact test. Median follow-up was calculated using the reverse Kaplan–Meier method (22). In the training set, survival curves for different variable values were generated using Kaplan–Meier estimates and compared using the log-rank test. Variables that achieved significance at P values less than 0.05 were subjected to multivariate analyses via the Cox regression model. Then, statistical analyses to identify independent prognostic factors were conducted. Based on the multivariate analysis results, a prognostic nomogram was established using a backward stepwise Cox proportional hazard model. The entire population was divided into three risk groups (high, intermediate, and low) according to the tertiles of the total scores given by the established nomogram in the training set. Concordance index (c-index) values were used to measure the discrimination performance. Finally, for external validation, the total scores of each patient in the validation set were calculated according to the proposed nomogram to verify its generalization. Three risk groups were determined by the tertiles defined in the training set, and the respective Kaplan–Meier survival curves were delineated. All statistical tests performed were two-sided, and P values less than 0.05 were considered statistically significant.

To comprehensively evaluate their clinical usefulness, we retrieved all of the published signatures. Accordingly, 81 signatures published as valid prognostic tools in colorectal cancer were obtained from 507 potentially related articles. Among these signatures, those with genes not clearly described in their respective studies or investigating the prognostic value based on tissue array, immune cells, circulation blood, or RT-PCR results were further excluded (see flowchart in Supplementary Fig. S1 for details). Finally, a total of 27 signatures from 26 studies met all of the inclusion criteria and were retained for subsequent analysis (Supplementary Table S3). Among these signatures, almost all were obtained from experiments based on colorectal cancer tissues except for one, which was derived from an animal experiment (23). The 27 signatures contained 1274 total genes, among which 1041 were unique, and the signature sizes ranged from 4 to 264 (Fig. 1A). The top overlapping genes in the above signatures were FN1 (10 times), CYP1B1 and POSTN (6 times), 7 genes (5 times), 7 genes (4 times), 29 genes (3 times), and 107 genes (twice). As is clearly shown in Fig. 1A, none of these genes appeared in all signatures. In addition, some repeated genes in different signatures even presented opposite prognostic values. This lower-than-expected overlap, at least in part, accounted for the low reproducibility of the signatures and the failure in clinical practice.

Accumulating evidence strongly suggests that the prognosis of colorectal cancer patients, such as recurrence and metastasis, is tightly regulated by aberrantly expressed genes via specific biological processes (16). To gain insight into the underlying biological meanings of the gene set obtained from the above signatures, enrichment (GO and KEGG) and protein–protein interaction (PPI) analyses were conducted (Supplementary Table S4). Figure 1B–D shows that specific GO categories closely related to colorectal cancer prognosis, such as cell cycle (FDR < 0.0001), regulation of cell proliferation (FDR < 0.0001), extracellular matrix (FDR < 0.0001), and chemokine activity (FDR < 0.0001), were significantly enriched. Additionally, some KEGG pathways well known to be involved in colorectal cancer metastasis and recurrence were enriched, including colorectal cancer metastasis signaling (FDR < 0.0001), wnt-β catenin signaling (FDR = 0.0002), VEGF signaling (FDR = 0.0008), and chemokine signaling (FDR = 0.0021; Fig. 1E). More importantly, a visualized network of these gene interactions revealed that the abovementioned pathways not only form highly integrated modules but also play essential roles in the entire network (Fig. 1F). These findings suggested that notwithstanding the low reproducibility in the system test and clinical validation, genes obtained from colorectal cancer signatures actually have potential prognostic value.

Furthermore, to assess the global performance of these signature and construct a composite nomogram, publicly available datasets with whole-genome profiles and corresponding clinical information in colorectal cancer patients were screened and downloaded (see flowchart in Supplementary Fig. S1 for details). Ultimately, 1710 colorectal cancer patients pertaining to 9 cohorts (Supplementary Table S5) met the set criteria and were retained. Although the profiles were obtained using different technologies and based on fresh or formalin-fixed and paraffin-embedded (FFPE) specimens as source material, they showed no evidence of cohort-bias clustering by principal component analysis (PCA; Supplementary Fig. S2A–S2C) or hierarchical clustering (Supplementary Fig. S2D), suggesting that patients from different cohorts in the present study could be blended together. Moreover, to balance the baseline characteristics, all eligible patients were randomly divided into a training set (n = 855) and a validation set (n = 855). The demographic and major clinicopathologic characteristics of the patients are summarized in Supplementary Table S2. As we expected, no significant differences were observed between the statistical properties of the training and validation sets.

To determine the correlations of signatures and colorectal cancer patient outcomes, the prognostic performances of the 27 signatures were evaluated in the training set using a modified NTP method as previously described (19). Among the 27 signatures evaluated, all except one (Teodoro's (24) signature had the only poor prognostic gene) were able to confidently stratify patients (FDR < 0.05) into good and poor subgroups. Table 1 and Fig. 2A summarize the prediction results obtained for each of the 855 patients. Popovici signature (34) was the most prevalent prediction in the training set (83.6%; 615 of 855), whereas Agesen signature (42) was securely identified in only 19.9% (167 of 855) of colorectal cancer patients. Of note, conflicting prognostic outcomes were commonly observed among the training set, and only 34% patients had consistent results (good or poor prognosis) in more than 5 signatures. However, as demonstrated in Fig. 2A, the overall tendency of the NTP analysis was consistent and roughly consistent with previous studies (25). In addition, global Cramer V coefficients were calculated for each signature to gain insight into their concordances. Unsupervised clustering based on these coefficients indicated a substantial association among these signatures (Fig. 2B). Obviously, 5 signatures obtained from the studies of Oh (13), Minh (11), Chang (9), Jiang (26), and Fritzmann and colleagues (27) were at least moderately correlated with each other. Moreover, another 4 signatures (Toshiaki's, Aziz's, Chen's, Xu's) were also clustered with similar coefficients. These findings are largely consistent with the initial purpose of each signature, which further confirms the potential prognostic value of these signatures.

We then applied Cox regression analysis to assess whether any signatures were statistically significant for colorectal cancer–related recurrence/progression and independent of clinical features in the training set. The univariate and clinical factor-adjusted multivariate analysis results are displayed in Supplementary Table S6 and Fig. 2C. Univariate analyses revealed that age (HR,1.12; 95% CI, 1.021–1.269, P = 0.048), local invasion (T stage; HR, 2.21; 95% CI, 1.139–7.029, P = 0.027), lymph node metastasis (N stage; HR, 2.28; 95% CI, 1.189–3.784, P = 0.021), distant metastasis (M stage; HR, 2.46; 95% CI, 1.363–4.453, P = 0.003), and TNM stage (HR, 2.15; 95% CI, 1.165–3.971, P = 0.014) were confidently associated with poor prognosis. However, with respect to genomic factors, only 8 signatures were validated to have significant effects on DFS (Fig. 2C). Next, we used multivariate Cox analysis with the covariates, including age, local invasion, lymph node metastasis, distant metastasis, and TNM stage, to evaluate the predictive values of the 8 signatures. As shown in Fig. 4C, after each covariate was adjusted, Popovici signature (HR, 1.92; 95% CI, 1.298–2.841, P = 0.001) and Fritzmann signature (HR, 2.16; 95% CI, 1.144–4.069, P < 0.0001) remained powerful enough to indicate prognosis in the training set for DFS, which revealed that these two signatures are independent prognostic factors.

Accurately predicting DFS in patients with colorectal cancer is important for counseling, follow-up planning, and selection of appropriate adjuvant therapy. Thus, to provide the clinician with a quantitative and user-friendly method to generate individualized predictions, we constructed a nomogram that integrated the genomic signatures identified above with clinicopathologic features (Fig. 3A). Calibration plots revealed excellent agreement between the nomogram-predicted probabilities and the actual observations of 1-, 3-, and 5-year DFS (Fig. 3B). The nomogram demonstrated that distant metastasis made the largest contribution to prognosis, followed by the TNM stage and lymph node metastasis. Age and Popovici signature had moderate effects on survival rate. Each category within these variables was assigned a point on the top scale based on the coefficients from Cox regression. By summing all of the assigned points for the seven variables and drawing a vertical line between total points and the survival probability axis, we easily obtained the estimated probabilities of 1-, 3-, and 5-year DFS. The risk score cutoff values (≤10.6, 10.6–21.2, and ≥21.2) were selected in terms of total points to stratify patients into roughly equal tertiles in the training set, which accurately divided patients into low- (reference), intermediate- (HR, 2.227; 95% CI, 1.507–3.291, P < 0.001), and high-risk (HR, 6.787; 95% CI, 4.786–9.624, P < 0.001) subgroups with significantly different DFS rates (Fig. 3C and D).

Although the TNM stage in combination with other clinical factors is a well-recognized prediction system for colorectal cancer prognosis (28), its effectiveness urgently needs to be increased. To determine whether the genomic signatures added additional prognostic value to the current system, time-dependent receiver operating characteristic (ROC) analysis was applied to compare the performances between the nomogram, clinicopathologic factors, and genomic signatures (Fig. 3E). Expectedly, the nomogram achieved the greatest area under the ROC curve (AUCs at 5-year DFS, 78.07, 71.40, 71.34, and 60.83, P < 0.05), suggesting that the integrated clinicopathologic genomic nomogram had a prognostic performance superior to those of clinicopathologic and genomic information by themselves.

To validate the correlation between the nomogram score (total points) and DFS statuses of patients with colorectal cancer, Kaplan–Meier analysis and log-rank tests based on the same cutoff value were conducted on the validation set. As shown in Fig. 4A and B, applying the clinicopathologic–genomic nomogram stratified patients into three distinct risk subgroups with significantly different DFS rates. In addition, the prognostic accuracy of the nomogram was remarkably better than those of both clinicopathologic and genomic information by themselves (Fig. 4C), which was consistent with the results obtained from the training set. Disagreements on potentially appropriate candidates for adjuvant therapy prevail, especially for patients belonging to AJCC stages II and III. In the present study, the survival rates predicted by the nomogram were significantly distinct between the Kaplan–Meier curves (Fig. 4D and E) in patients categorized by the abovementioned major clinicopathologic features (AJCC stages II and III).

Moreover, to ensure the effectiveness of statistical power, patients with available chemotherapy information were integrated together regardless of the set to which they belonged (training set or validation set). Intriguingly, we noted that adjuvant chemotherapy did not enhance OS in all 869 patients (HR, 0.98; 95% CI, 0.769–1.258, P = 0.873; Fig. 5A) or in patients who were defined as low risk by the nomogram (HR, 0.93; 95% CI, 0.515–1.695, P = 0.824; Fig. 5B). However, patients in the nomogram defined as intermediate risk (HR, 0.98; 95% CI, 0.255–0.679, P < 0.001; Fig. 5C) or high risk (HR, 0.67; 95% CI, 0.469–0.957, P = 0.028; Fig. 5D) had favorable responses to adjuvant chemotherapy. Taken together, these results suggested that the clinicopathologic–genomic nomogram has the discriminatory power to identify colorectal cancer patients who are suitable candidates for adjuvant chemotherapy.

Although surgery remains the mainstay of curative treatment, the prognosis of patients with colorectal cancer, especially for long-term outcome, depends more on pre- or postoperative individualized therapies (29), including various strategies of chemotherapy and chemoradiation. However, as a well-recognized heterogeneous cancer, its prognosis varies significantly, and optimal management presents challenges. Currently, adjuvant chemotherapy is regarded as standard treatment for patients with stage III colon cancer (4). However, patients with T1–2N1M0 tumors (stage IIIA) have significantly higher survival rates than those with stage IIB tumors (3). Moreover, for some patients, long-term survival continues to be jeopardized by the high risk of recurrence even after chemotherapy, and personalized therapy needs to further optimized. With respect to stage II patients, disagreement regarding which patients are potentially appropriate candidates for adjuvant therapy prevails. The MOSAIC trial (30) has shown only a 20% benefit from adjuvant therapy at 3 years for patients with stage II disease and a 28% benefit for patients with recognized clinical risk factors. Thus, for approximately 75% of stage II patients who receive curative surgery, the administration of adjuvant chemotherapy is unnecessary and harmful. Unfortunately, low-risk stage II patients have not been accurately defined. In addition, patients with microsatellite instability (MSI colorectal cancer) have a better prognosis (31) and do not benefit from 5-fluorouracil (5-FU)–based adjuvant chemotherapy (32), similar to patients with CpG island methylation (33). All the abovementioned studies highlight that the rational application of therapeutic strategies requires accurate prognosis prediction and risk stratification for colorectal cancer patients.

Considering the inherent deficiency of TNM stages, substantial effort has been placed on their optimization (2), and they have been continuously developed over the past decade. Substantial changes in colorectal cancer classifications, based mainly on the analysis of Surveillance, Epidemiology, and End Results (SEER) data, were made in the seventh edition (34). However, a comparative study demonstrated that the seventh TNM edition did not provide a greater accuracy for predicting colorectal cancer patient prognoses but rather resulted in a more complex classification for daily clinical use (35). The eighth edition of the colorectal cancer staging system, containing major advances, including the introduction of molecular markers, such as MSI, KRAS, and BRAF mutations, to strengthen its discriminatory power, were implemented worldwide on January 1, 2018 (2). Nonetheless, the forthcoming edition may not provide substantial improvement because the prognostic value of clinicopathologic factors has peaked. Moreover, the current models still lack genomic information, which may directly reflect the heterogeneity of colorectal cancer and present substantial potential prognostic value.

Genomic signatures predicting the prognosis of colorectal cancer have been continually developed in the past 10 years. In the present study, we retrieved 81 articles related to the development of signatures for predicting colorectal cancer prognosis, including OS, DFS, disease-specific survival (DSS), and various types of recurrence. For instance, based on the association between the aberrantly expressed gene and the duration of individual DFS, Zhang and colleagues (36) constructed a six RNA-based classifier, which was further validated to have reliable predictive power for detecting recurrence in patients with stage II colon cancer. Similarly, Anna and colleagues (23) developed a genomic profile specific for adult intestinal stem cells (ISC), which was highly enriched in genes that positively predicted the risk of recurrence. Patients bearing primary colorectal cancers with high average expression levels of ISC genes had a relative risk of relapse that was 10-fold higher than that for patients with low levels (23). Moreover, recent evidence (37) suggests that gene signatures that closely reflect specific biological processes or oncogenic pathway statuses also have potential prognostic values for stratifying colorectal cancer patients. Even though almost all of these signatures were proven to have prognostic value in their respective publications, none are routinely used in clinical practice.

Potential explanations for the unsatisfactory results must be considered. Thus, in the current study, we exhaustively reviewed and analyzed the published multigene signatures in colorectal cancer. To eliminate the potential confounding effects, only data derived from the Affymetrix human genome platform having clearly described genes were included. Of the 81 published signatures, 27 signatures from 26 publications met all of the inclusion criteria and were retained (Table 1; Supplementary Table S3). The 27 signatures contained 1,274 total genes, among which 1,041 were unique, and the signature sizes ranged from 4 to 264 (Fig. 1A). The top overlapping genes in the above signatures were FN1 (10 times), CYP1B1 and POSTN (6 times), 7 genes (5 times), 7 genes (4 times), 29 genes (3 times), and 107 genes (twice). As is clearly shown in Fig. 2A, none of these genes appeared in all signatures. These results were expected because they were previously reported (15) and in agreement with the systematic analyses of lung cancer and breast cancer prognostic signatures (38). Obviously, the slight overlap of these signatures in gene identity was perplexing but sufficiently explained the low reproducibility. Moreover, genetic heterogeneity was recently shown to exist not only between but also within tumors, meaning that biological statuses, including oncogenic and carcinostatic states, might regularly vary in the same patient (39). Our NTP analysis revealed that some patients concurrently harbored more than one signature, which further confirmed the abovementioned results and indicated that their tumors were genomic instability at the individual level. Taken together, the different signatures might reflect diverse biological processes and be active in different or identical individual tumors, which partially accounts for the wide nonoverlapping among signatures constructed in different colorectal cancer studies.

In terms of global performance, in the training set, 8 of 27 signatures showed a significant association with prognosis and could reasonably predict the outcome. However, after multivariate adjustment, only 2 signatures remained powerful enough to indicate prognosis for DFS, which directly demonstrated the low reproducibility of colorectal cancer genomic signatures. More importantly, Popovici signature, Fritzmann signature and TNM stage had remarkable prognostic abilities and were independent of each other, indicating that these signatures may be used to refine the current prognostic model and facilitate further stratification of patients with colorectal cancer at the same TNM stage.

Our primary goal was to construct a nomogram that could integrate genomic signatures with clinicopathologic variables in a large cohort of colorectal cancer patients to add additional prognostic value to the current staging system. Our multivariate analysis in the training set revealed that age, local invasion, lymph node metastasis, distant metastasis, TNM stage, Popovici signature and Fritzmann signature were independent prognostic factors (Supplementary Table S6), which was highly consistent with previous studies on colorectal cancer risk factors (40, 41). Accordingly, we cautiously built a clinicopathologic–genomic nomogram using the abovementioned significant factors. The AUC of the nomogram showed a superior prediction accuracy compared with that of the combined clinical factors in the training and validation sets (Figs. 3E and 4C). Additionally, the calibration plots showed optimal agreement between the expected and observed survival probabilities (Fig. 3B), ensuring the reliability of our nomogram. Moreover, the nomogram successfully stratified colorectal cancer patients into three risk groups with remarkably different DFS values based on total points, which was further confirmed using the same cutoff value in the validation set.

Intriguingly, the TNM staging system is the most important tool for guiding treatment in clinical practice. However, disagreement prevails on potentially appropriate candidates for adjuvant therapy, especially for stage II and III colorectal cancer patients. Therefore, the identification of a high-risk subgroup among these patients is of urgent clinical needed. Our findings suggest that the clinicopathologic–genomic nomogram could be a promising tool to stratify stage II and III patients into distinct risk subgroup and guide individualized therapy. Moreover, in the current study, we clearly revealed that chemotherapy provides a survival benefit to patients classified as intermediate- and high risk by the nomogram (Fig. 5).

Despite the promising results, there were several shortcomings in this work. First, the NTP method uses only a list of signature genes to make class predictions in each patient's expression data, allowing the method to be less sensitive to differences in experimental and analytical conditions and applicable to every patient without optimization of the analysis parameters. In the current study, we applied the NTP method for all prediction models, which generally consisted of two major components: a gene signature and a prediction algorithm. Obviously, the intrinsic limitation of the NTP method is that it inevitably causes the algorithm lose its prognostic value. Second, the forthcoming 8th edition of the colorectal cancer TNM staging system may better predict prognosis, but data used in the present study were from the current 7th edition or the 6th edition. Third, the prognostic signatures of colorectal cancer included in our study were not exhaustive, and some promising signatures were excluded because of insufficient information. In addition, some clinicopathologic factors, even those well recognized to have prognostic value, were not included in the composite nomogram because of the low availability of information.

In summary, based on analyzing a large-scale colorectal cancer cohort and published multigene prognostic signatures, we confirmed herein for the first time that genomic information in combination with clinical data can improve patient prognostic evaluations and should be considered as an independent and complementary approach to the current clinicopathologic prognostic model. Furthermore, considering the wide promise of our nomogram, which integrates genomic signatures with clinicopathologic features to improve the prognosis prediction of colorectal cancer, further systematic research is needed.

No potential conflicts of interest were disclosed.

Conception and design: Y. Xiong, Z. Fu

Development of methodology: W. You

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y. Xiong, M. Hou

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y. Xiong, M. Hou, L. Peng, H. Zhou

Writing, review, and/or revision of the manuscript: Y. Xiong, W. You, L. Peng

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Z. Fu

This work was financially supported by a grant from the National Natural Science Foundation of China (grant no. 81572319; project recipient: Z. Fu).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.