Predicting tumor repopulation through the gene panel derived from radiation resistant colorectal cancer cells

In order to acquire the necessary data for this study, we used the public functional genomics data repository, Gene Expression Ominibus. By setting “colorectal cancer”, “radiation” and “resistance” as three keywords for obtaining desired project, we enrolled series GSE97543 as our mainly analyzing dataset eventually [15]. In this project, gene expression of radiation resistant colorectal cell line and its control group were described, and all GEO data were downloaded through package “GEOquery” of R software [16].

To further get gene expression and clinical data of patients, part of cases originated in TCGA (The Cancer Genome Atlas) program were involved. We collected gene expression data of colon, rectum and rectosigmoid junction adenocarcinoma (TCGA-COAD, TCGA-READ) cases, including 644 tumor samples and 51 normal samples to search target genes with GSE97543. Meanwhile, radiation treated TCGA-CO/READ patients (n = 35) were set as training set. Patients who received external radiotherapy as only adjuvant therapy in BRCA (breast carcinoma, n = 38), LUAD (lung adenocarcinoma, n = 36), CESC (cervical squamous cell carcinoma and endocervical adenocarcinoma, n = 20), HNSC (head and neck squamous cell carcinoma, n = 95) and ESCA (esophageal carcinoma, n = 17) were introduced to be validating parts. All TCGA data, mRNA expression and clinical details were manually downloaded from the website and organized by R software.

Two R packages “limma” and “edgeR” were adopted to identify the differential expression genes (DEGs) in both databases [17, 18]. |Log2 fold change|≥ 1 and P. value ≤ 0.05 were set as the selection criterion, which defined the scope of DEGs in two databases for further analysis. And volcano plots were draw by R packages “limma” and “ggplot2”. A gene annotation and analysis resource website, “Metascape” (https://​www.​metascape.​org/​), was utilized to cluster gene function and to enrich pathways related to DEGs [19].

After acquiring 252 upregulated co-expression genes, which were candidates to construct radiation resistance gene panel, between GSE97543 and TCGA-CO/READ, we involved univariate Cox regression to pick out the most relevant genes with PFS (progression-free survival) of radiation treated colorectal patients with P. value < 0.06, which was set to prevent missing out clinically significant genes that closely approached the conventional significance level [20, 21].

Then, we developed a proportional hazards regression analysis (multivariate Cox model) by the following formula using the result of univariate Cox regression.

In this equation, risk score is the mRNA expression of each key predictor genes (xi) multiplied by the coefficient (Coefficienti), the latter came from multivariate Cox regression analysis. Using the median value of risk score that calculated by the formula before, the 35 TCGA patients in the training set were divided into two groups, high risk group and low risk group. Based on this grouping routine, we observed the expression levels of key genes in the two groups and established a survival analysis between PFS and risk score. ROC (receiver operating character) curves were implemented to evaluate the predictive ability of radiation resistance gene panel using R package “survivalROC”, and the AUC (area under the curve) values were calculated to visualize inspection capabilities.

In order to test whether risk score was an independent prognostic factor, we incorporated the age, gender, stage, T, N, M status and risk score of 35 TCGA patients into the multivariate Cox regression model, so as to exclude confounding caused by other clinical features. Furthermore, R package “rms” was used to make a nomogram model, showing above clinical information.

Besides TCGA-CO/READ, we further chose patients who received radiotherapy as only adjuvant therapy from other TCGA projects to check validity of radiation resistance gene panel. Logistic regression models were built by R package “glmnet”, and different gene combination models were listed. Several best models and their AUC values of different cancer types were displayed respectively.

The R package, “WGCNA” (Weighted Gene Co-expression Network Analysis), was utilized to build a co-expression network targeting DEGs. This method aims to find co-expressed gene modules, and to explore the relationship between gene networks and the phenotype of interest, as well as the core genes in the network [22]. It is divided into two parts: expression cluster analysis and phenotypic correlation.

The specific method was that selecting patients who received radiotherapy in TCGA-COREAD, and incorporating the top ten thousand DEGs with the largest variance of these patients into the cluster analysis of gene expression. After clustering the samples and eliminating outliers, we determined the soft-thresholding power according to the algorithm, which was used to construct a scale-free co-expression network and determine gene modules. Each color represented a module, and each module contained genes with similar expression patterns. Dynamic tree cut analysis was developed to represent the classification of genes, and modules with high similarity were fused to construct merged dynamic clusters. In addition, different modules were established the correlation to two clinical traits, PFS and risk score, and modules that contained key radiation resistant genes were selected to carry out GO analysis.

Immortalized human colorectal tumor cell lines (HCT 116 and HT-29) and 293 T cells used in this research were bought from cells bank of the Chinese Academy of Science (Shanghai, China). HCT 116, HT-29 and 293 T cells were seeded in completed DMEM medium (Wisent, China). 10% Fetal Bovine Serum (Wisent, China) and 1% Penicillin–Streptomycin (Wisent, China) were added to the completed DMEM medium. All cells were cultured in 5% CO₂ at 37 ℃.

An X-Ray generator (Faxitron, USA) for laboratory was used to treat cells. And the dose rate was 3 Gy/min. HCT 116 and HT-29 were seeded in six well culture dishes (100, 200, 1000, 2500 and 10,000 cells/well) and incubated eight hours (overnight) until adherent before irradiation. After adherent, cells were treated with various doses (0, 2, 4, 6, 8 Gy) of X-Ray respectively. All cells were fixed by paraformaldehyde of 4% fortnight later. And then, they were stained by crystal violet. This experiment was repeated three times independently. Any colony that contained fifty cells or more were counted and linear-quadratic model in Graphpad Prism 8 was employed to calculate the surviving fraction.

Total mRNA was extracted by using RNA Easy Fast Cell Kit (Tiangen, China). cDNA was reversely transcribed from mRNA by using the RT Kit. Q-PCR was conducted by using TB Green Kit. Relative mRNA level was calculated by the formula 2 − △△CT. The experiments were repeated three times. Both kits mentioned above were from Takara, Japan. Specific primers information is clarified in supplemental material. GAPDH, an internal reference gene, were used as a control for standardization.

All shRNA sequences were got from MERCK website. When the confluence of 293 T cells in the cell culture dish reached 80–90%, we mixed the target plasmid containing shRNA, tool plasmid psPAX2, pMD2.G and GFP fluorescent plasmid into the 400 μL Opti-MEM medium according to the protocol proportion, and then added the mixed system into 293 T cell culture medium. After 48 h, the virus solution from 293 T cells was obtained and added to HCT116 and HT29 cell culture medium with 10 μg/mL polybrene. After another 48 h, the virus group cells and control group cells were treated with 1.5 μg/mL purinomycin (the tool plasmid contained purinomycin-resistance gene to ensure that these cells successfully infected by the virus would not be killed). When the control cells were all killed by purinomycin, the surviving cells in virus group were thought to be successfully infected by the virus. Western Blot experiment was carried out to detect the expression level of genes.

Cell protein was obtained by using cracking liquid to crack the cells. The obtained protein denatured by boiling, and SDS-PAGE gel was used for electrophoresis to separate proteins with different molecular weights. After electrophoresis, SDS-PAGE gel and PVDF (polyvinylidene fluoride) paper were used to constructed gel-matrix sandwich to transfer. After that, the PVDF paper was incubated successively with the first and second antibodies, and exposed in developer to obtain the image of the target band. These experiments were repeated three times independently. The blank controls were wild type HCT116 and HT29 cells without any treatment. The negative control referred to two types of cells transferred into empty plasmids.

We conducted this research mainly based on R and R studio software. R packages employed include: “clusterProfiler”, “GEOquery”, “org.Hs.eg.db”, “HsAgilentDesign026652.db”, “R.utils”, “rjson”, “jsonlite”, “reshape2”, “ggfortify”, “limma”, “edgeR”, “GOplot”, “stringr”, “ggplot2”, “dplyr”, “pheatmap”, “glmnet”, “survival”, “survminer”, “rms”, “cowplot”, “WGCNA”, “ROCR”, “survivalROC”. The difference in progress-free survival in different groups of patients was analyzed using log-rank test in GraphPad Prism 8, and two tailed student’s t-test was used to compare mean values. * for P ≤ 0.05; ** for P ≤ 0.01; *** for P ≤ 0.001. Statistical significance was set at P ≤ 0.05 in most conditions, and specific principles are n.s for not significant.

In order to study the genes that played a key role in radiotherapy resistance, we selected the dataset GSE97543 as the research object [15]. In this database, the radiation-resistant human rectal adenocarcinoma cell line SW1463 was established through repeated radiation exposure, named SW1463RES. Compared with control cells after 4 Gy radiation, SW1463RES presented 710 upregulated genes and 1954 down-regulated genes, when cut-off values were set as |Log2 fold change|≥ 1 and P. value ≤ 0.05 in DEGs analysis (Fig. 1a, Additional file 6: Table S1). In order to clarify the function of the obtained upregulated genes, we used the Metascape tool website to perform Gene Ontology (GO) analysis and pathway analysis on this group of genes, which primarily enriched in: (1) cell cycle related pathways (Cell cycle, Mitotic cell cycle process, Mitotic G2-G2/M phase, E2F pathway, FOXM1 pathway, DNA IR-damage and cellular response via ATR); (2) DNA replication and repair related pathways (DNA replication, DNA biosynthetic process, DNA repair, Fanconi pathway and Homologous recombination repair); (3) stress response pathways (response to xenobiotic stimulus) (Fig. 1b). According to the results of enrichment analysis, radiation resistant SW146RES cells exhibited stronger DNA damage repair capabilities than control cells, which were in accordance with previous views [23, 24]. Because these genes that represented radio-resistance may promote tumor to recover from the stress state caused by X-Ray beam more quickly, they were inclined to lead tumor repopulation after radiation. Therefore, we further conducted a DEGs analysis, tumor vs. normal tissue, in the TCGA-CO/READ databases, and acquired 3669 upregulated genes while 5477 down-regulated genes were spotted (Fig. 1c, Additional file 7: Table S2). By using the Venn plot to intersect the upregulated genes of two groups, we finally found 252 co-upregulated genes (Fig. 1d), and numerous similar pathways were enriched in cell cycle, DNA replication and repair (Additional file 1: Figure S1). The specific items we selected in the TCGA and GEO databases were listed in Fig. 1e, and the detailed process of how we used these genes expression data to build our practice was described through a visual flow chart (Fig. 1f). These 252 co-upregulated genes were considered as candidates for constructing the radiation-resistance gene panel.

To investigate the key genes that had the greatest impact on the prognosis of colorectal cancer patients who undergoing radiotherapy, we selected 35 colorectal cancer patients who received radiotherapy in TCGA database to develop univariate and multivariate Cox proportional hazard regression models (baseline information of patients in Additional file 8: Table S3). Among the 252 co-expressed upregulated genes mentioned above, a total of five genes (LGR5, KCNN4, PTRH1, CENPH, TNS4) were considered to be associated with the patients’ PFS in the result of univariate Cox regression model. And among the five genes, LGR5, KCNN4, CENPH, and TNS4 were risk factors. Therefore, we further included those four genes in multivariate regression analysis to construct a gene panel related to radiation resistance. Then, we used coefficients in the multivariate Cox regression and the mRNA expression of key genes to establish a formula to calculate the risk score: Risk score = (CENPH*0.2551 + LGR5*0.05586 + KCNN4*0.02614 + TNS4*0.01099). Hazard ratios superior to 1 indicates that patients with high expression of these four genes are more likely to develop tumor progression after receiving radiotherapy (Fig. 2a). The heatmap clarified the expression level of four genes in each single patient, and all 35 patients were divided into two groups based on the median risk score, with 17 patients in high-risk group and the rest in low-risk group (Fig. 2b, c). As depicted in Fig. 2c, patients with higher risk scores after radiotherapy were more likely to occur tumor progression, and their median progression-free survival were relatively short. Specifically, when we constructed a progression-free survival curve on account of the expression of a single key gene, only one gene could obtain a statistically significant result (KCNN4, P = 0.0042, Additional file 2: Figure S2); but when we grouped patients according to risk score, high-risk group had a significantly higher progression rate in 3 years, 5 years and even longer periods than the low-risk group (P = 0.0001, HR = 6.402) (Fig. 2d), proving the superior performance of this prognostic model. Based on the radiation-resistance gene panel, ROC curve was introduced to reflect this model’s ability in predicting prognosis, with the AUC of 1, 3 and 5- year curves reaching to 0.87, 0.94 and 0.9 respectively, which further indicated a reliable predictive ability (Fig. 2e).

In order to observe the relationship between remaining clinical characteristics and risk score, we established a multivariate regression model (C index = 0.858), including age, gender, stage, T, N, M levels and chemotherapy, which indicated that risk score had a significant association with PFS (P = 0.001) (Additional file 3: Figure S3A). The nomogram was utilized to sketch above mentioned results, combining the risk score and other clinical records, and offered a quantitative tool to predict the probability of progression for each patient (Additional file 3: Figure S3B). What mentioned above suggested the independent prognostic function of risk score (Additional file 9).

To further verify the reliability of the radiation-resistance gene panel, we established a multivariate logistic regression analysis to evaluate its efficacy to predict tumor progression after radiotherapy. This inspection method can reflect the relationship between gene panel and tumor progression more intuitively. By building a series of regression models with a single gene or combinations with multiple genes, we compared the AUC values in various situations (Fig. 3a). Generally speaking, the AUC values of combinations with two or more key genes were higher than that of a single gene. The AUC value of a single candidate gene ranged from 0.513 for TNS4 to 0.755 for LGR5. Among the various combinations, the one that included all four key biomarkers had the highest AUC value, 0.827, representing the best predictive ability (Fig. 3b). And Fig. 3c–e showed three combinations with AUC values slightly lower than the highest one, reaching 0.823, 0.817 and 0.801 respectively. The data obtained through the above regression models supported our hypothesis that four key genes were participated in radiation resistance of colorectal cancer patients and associated with poor PFS. They may also involve in past-radiotherapy tumor repopulation given the intimate connection between gene panel and PFS.

After clarifying the reliability of the radiation resistance gene panel in predicting the tumor progression of colorectal cancer patients after radiotherapy, we brought into the weighted gene co-expression network analysis (WGCNA). In this way, we could further explain the correlation between gene panel, PFS and risk score, as well as clustering genes with similar expression patterns to key genes. And the top 10,000 DEGs with the largest variance employed in WGCNA were listed in Additional file 10: Table S5.

First, we set the scale free topology model fit R^2 cutoff value to 0.9, and pick beta = 9 as the soft threshold power, building a scale-free network analysis on this basis (Fig. 4a). By merging modules with similar expression profiles in the dynamic tree cut, we eventually got a merged one composed of 13 modules (Fig. 4b). The quantitative connection between each module and clinical trials, PFS and risk score, was shown in Fig. 4c. Generally, PFS and risk score shown a negative correlation trend in post-radiotherapy colorectal cancer patients, providing evidence for the effectiveness of gene panel in forecasting radiation resistance and tumor repopulation. Four key genes in the gene panel were clustered into four different modules respectively (Fig. 4d). Detailed groupings and correlation index (Cor.) were as follows. KCNN4 was clustered in salmon module with 0.119 risk score Cor. and − 0.05 PFS Cor., while TNS4 in brown module with 0.206 risk score Cor. and 0.036 PFS Cor. For these two genes, high risk score corresponded to a relatively negative outcome in PFS. Although LGR5 in green-yellow module (0.032 risk score Cor., 0.115 PFS Cor.) and CENPH in grey module (− 0.099 risk score Cor., 0.089 PFS Cor.) were classified into the low risk score and good PFS outcome modules, the two should be risk factors according to hazard ratios in Cox regression models (Fig. 2a). Moreover, GO enrichment analysis was carried out on the genes in these four modules. And the results were mainly presented as cell cycle control related pathways and activation of immune related pathways caused by the release of intracellular substances in stress status after radiotherapy (Fig. 4e–h).

In addition to establishing and testing gene panels in colorectal cancer patients, we selected patients in other TCGA projects and used a series of in vitro experiments to further illustrate the sensitivity and specificity of gene panels in predicting radiotherapy resistance and tumor repopulation. Among other types of tumors in the TCGA database, we chose BRCA (n = 38), LUAD (n = 36) and CESC (n = 20) patients who received radiotherapy as the only adjuvant treatment to further proof our hypothesis. Multivariate logistic regression analysis was established in patients of the three kinds of tumor as well, and gene combinations with top 5 AUC values of each cancer were listed in Fig. 5a, c and e. As for breast cancer patients who received radiotherapy, the predictive model containing all four genes reached to the highest AUC value, 0.974 (Fig. 5b). Prognostic model without LGR5 worked best in LUAD patients, with an AUC value of 0.861 (Fig. 5d). In CESC patients, the highest AUC value of 0.745 was a model constructed by KCNN4 and TNS4 (Fig. 5f). In HNSC and ESCA patients, the gene panel also showed acceptable predictive ability (Additional file 4: Figure S4).

Besides, two colorectal cancer cell lines, HT29 and HCT116, were selected for clonogenic formation assay. The results showed that the colony formation ability of HT29 was significantly higher than that of HCT116 after being exposed to various doses of radiation, which meant HT29 cells showed more resistant to radiation (Fig. 5g, Additional file 4: Figure S4). In parallel, the basic expression of radiation-resistance genes in two cells was generally consistent with our expectation, except that KCNN4 had a higher baseline in HCT116 (Fig. 5h).

In order to further confirm the role of the four key genes we found in radio-sensitivity, we constructed four gene knockdown cell lines respectively by using shRNA, and HCT116 cell lines with better knockdown effect were selected through Western Blot experiments (Fig. 6a). Then, we carried out clone formation assay with these cells (Fig. 6b). Compared with the control cells transfected with empty plasmid (NC shRNA), a decreased clonogenic ability of CENPH and LGR5 knockdown cells could be observed in the untreated group, while the ability of KCNN4 and TNS4 knockdown cells was not significantly affected (Fig. 6c). However, when treated after 8 Gy irradiation, CENPH knockdown cells could hardly form clone, and the other three gene knockdown cells could only form few clones (Fig. 6b). And the clonogenic ability of all knockdown cells decreased unexpectedly after radiation treatment compared with the control groups (Fig. 6c). The same gene knockdown and clonogenic assay for further validation were also carried out in Ht29 cells (Additional file 5: Figure S5).