A novel DNA repair‐related nomogram predicts survival in low‐grade gliomas

Abstract Aims We aimed to create a tumor recurrent‐based prediction model to predict recurrence and survival in patients with low‐grade glioma. Methods This study enrolled 291 patients (188 in the training group and 103 in the validation group) with clinicopathological information and transcriptome sequencing data. LASSO‐COX algorithm was applied to shrink predictive factor size and build a predictive recurrent signature. GO, KEGG, and GSVA analyses were performed for function annotations of the recurrent signature. The calibration curves and C‐Index were assessed to evaluate the nomogram's performance. Results This study found that DNA repair functions of tumor cells were significantly enriched in recurrent low‐grade gliomas. A predictive recurrent signature, built by the LASSO‐COX algorithm, was significantly associated with overall survival and progression‐free survival in low‐grade gliomas. Moreover, function annotations analysis of the predictive recurrent signature exhibited that the signature was associated with DNA repair functions. The nomogram, combining the predictive recurrent signature and clinical prognostic predictors, showed powerful prognostic ability in the training and validation groups. Conclusion An individualized prediction model was created to predict 1‐, 2‐, 3‐, 5‐, and 10‐year survival and recurrent rate of patients with low‐grade glioma, which may serve as a potential tool to guide postoperative individualized care.


| INTRODUC TI ON
Diffuse low-grade gliomas (LGGs) are infiltrative, incurable lesions characterized by a continuous slow-growth and an almost unavoidable anaplastic transformation. [1][2][3] Median overall survival for patients with LGGs ranges from 5.6 to 13.3 years depending on tumor histopathologic feature, molecular phenotype, and growth rate. 2,4,5 Unlike their high-grade glioma counterparts, low-grade glioma with a more favorable prognosis pose unique challenges for both clinicians and patients for time-consuming monitoring of tumor recurrence. 6 An individualized plan of postoperative imaging assessment will facilitate the efficient use of medical resources and reduce medical costs.
Postoperative individualized care plan based on the highly accurate individualized recurrence prediction model. Nomograms, presenting the results of predictive models in a printed format, could be widely used in clinical practice. 7 Numerous nomograms for Overall

| Material and Methods
In total, 188 patients with primary or recurrent LGG were enrolled in the training group and 103 patients in the validation group.
Resected tumor samples were immediately placed in liquid nitrogen and only samples with more than 80% tumor cells, judged by HE staining of adjacent tissues, were selected for further sequencing. Transcriptome data of LGG samples were generated by the Agilent platform. Molecular testing was performed at the Molecular Pathology Testing Center of Beijing Neurosurgical Institute. All patients were followed up trimonthly by telephone or clinic for an average of 1813 days. 15 of 188 patients (7.98%) lost to follow-up in the training group and 5 of 103 patients (4.85%) lost to follow-up in the validation group. Clinical information of patients was summarized in Table S1.
The sequencing data, clinical, and follow-up information of primary and recurrent LGG patients were uploaded to the CGGA portal (http://cgga.org.cn/). All datasets used and/or analyzed in this study are available from the corresponding author on reasonable request.

| Biological functional enrichment scores
The biological functional enrichment score of each patient was generated by Gene Set Variation Analysis (GSVA) analysis based on tumor transcriptome sequencing data. GSVA analysis was performed using the default parameters by the gsva package in R as described in the previous study. 14 Gene list for each biological function was downloaded from AmiGO 2 Web portals (http://amigo.geneo ntolo gy.org) most recently.

| LASSO-COX dimension reduction analysis
LASSO-COX dimension reduction analysis was performed by glmnet and survival packages in R. The λ value corresponding to the minimum partial likelihood deviance was selected as the optimal λ in our study.
where expr gene was the expression level of the gene and λgene was the corresponding lambda value.  org/gsea/index.jsp) and gene ontology (GO) was performed in the DAVID portal website (https://david.ncifc rf.gov/summa ry.jsp). For all statistical methods, P < 0.05 was considered as significant difference.

| DNA repair functions significantly enriched in recurrent low-grade gliomas
A 5917 biological functional enrichment scores for 138 primary and 50 recurrent low-grade glioma patients were calculated by the GSVA algorithm. We found that 2596 biological functions were significantly increased in recurrent tumors, while 108 biological functions were significantly decreased ( Figure 1A). Classification of significantly elevated biological functions in recurrent tumors found that proliferation and cell cycle (24%), transcription and translation (15%), metabolic process (12%), and response to stimulus (11%) account for the highest proportion ( Figure 1B). The biological functions related to tumor progression-free survival (PFS) in low-grade glioma were screened out by multivariate COX analysis. The biological functions most related to PFS in each classification were shown in Figure 1C-F.
As expected, the results suggested that faster cell cycle, increased DNA repair and biosynthesis, and cellular response to radiation were significantly elevated in recurrent tumors.

| Development of a recurrent signature for lowgrade gliomas
The biological functions related to PFS were included in multivariate COX analysis to screen independent prognostic functions. Positive regulation of response to DNA damage stimulus was screened out as a function that was significantly elevated in recurrent gliomas and had the most independent prognostic value for PFS ( Figure 2A) Univariate and multivariate COX analysis revealed that the recurrent score was an independent prognostic factor in the training and validation database (Table S2-S5). The ROC curve was performed to verify the accuracy of the recurrent score in prognostic prediction ( Figure S1A-D). Also, the predictive role of the recurrent score was further verified in other LGG databases ( Figure S1E,F).

| Relationship between recurrent scores of the recurrent signature and clinicalpathologic characteristics
The relationship between recurrent scores and clinical-pathologic factors was further tested. The recurrent score significantly increased in recurrent tumors and moderately slightly increased in patients with postoperative radiotherapy in the training database ( Figure 4A). The recurrent score significantly increased in 1p/19q non-codeletion tumors in the validation database ( Figure 4B).
However, the recurrent score showed no correlation with histology, gender, age, postoperative chemotherapy, and IDH mutation status in both training and validation databases.

| The recurrent score is closely related to cell division and DNA metabolism
To explore the biological functions and pathways associated with the recurrent score, the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analysis and Gene Set Enrichment Analysis (GSEA) were performed. After screening the genes most related to the recurrent score, GO and KEGG analyses were performed based on these genes. GO analysis showed that the recurrent score was closely related to G2/M transition of mitotic cell cycle and DNA repair in training and validation databases ( Figure 5A,C). KEGG analysis showed that the recurrent score was closely related to the p53 signaling pathway and mismatch repair in both databases ( Figure 5B,D). The close relationship between the recurrent score and DNA repair-related functions was further verified by GSEA analysis in training and validation databases ( Figure 5E,F).

| The recurrent score is closely related to the DNA repair function
The relationship between the recurrent score and DNA repair function was further explored. Functional enrichment scores of DNA repair functions of each patient were calculated. Correlation analysis found that the recurrent score was significantly positively correlated with most DNA repair functions. The recurrent score was significantly positively correlated with 16 DNA repair functions in the training database. In the validation database, the recurrent score was significantly positively correlated with 14 kinds of DNA repair functions ( Figure 5G).

| The individualized prediction model showed robust predictive accuracy
To facilitate the clinical application of the prognostic prediction model, an individualized prediction model was constructed. The  Figure 6B). The C-index of this nomogram model was 0.78, which is higher than any other prediction model ( Figure 6C). To further expand the application range of our prediction model, the individualized prediction model for OS prediction based on predictive factors, including recurrent score, primary/ recurrent status, histology, age, postoperative radiotherapy, and IDH mutation status was also established for (1-,2-, 3-,5-and 10- year) survival probability prediction of low-grade glioma patients ( Figure S2A). The OS prediction model also showed robust predictive accuracy ( Figure S2B,C). LGG. 13,20,21 We further explored the relationship between recurrent score and tumor location. As shown in Figure S3, there is no significant correlation between the recurrent score and tumor location in both training and validation databases. Furthermore, the nomogram provided better predictive accuracy than the clinical factor-based model or recurrent signature alone, demonstrating the incremental value of the nomogram to the current early diagnosis of recurrent LGG. Moreover, our nomogram is easy to use, and it could serve as a quick and efficient tool for individualized prediction of prognosis and for guiding treatment in recurrent LGG patients.

| D ISCUSS I ON
As a standard adjuvant treatment for low-grade gliomas, postoperative radiotherapy and chemotherapy kill tumor cells by inducing DNA damage. 4,6 The unrepaired DNA damage is also a major source of potentially mutagenic lesions that promotes malignant progression of tumor. 22,23 Therefore, response to DNA damage is a key factor in tumor progression and recurrence of LGG. Our study found that response to DNA damage is a key factor in the recurrence of LGGs, and that DNA damage response-based signature is independent of the clinicopathological state of patients, such as histology, gender, WHO risk status, and postoperative treatment. 5,24 Further analysis confirmed that the recurrent score, as an accurate reflector of the DNA repair function, was constructed as a robust predictor of recurrence of LGG. This study suggested that DNA repair function targeted therapy may prevent the progression and recurrence of LGG.
The recurrent score can also be used as a predictor of the sensitivity of targeted therapy.
As a clinical application tool, our nomogram included only routine clinical examination items for glioma and did not use factors that may require statistical software or trained analysts such as tumor volume, the extent of resection, and epilepsy seizure types. 25,26 Although not perfect, this represents an encouraging level of predictive accuracy. Calibration shows how closely the predicted probabilities agree numerically with the actual outcomes. Of note, easily acquired factors and user-friendly operation methods make this prediction model more widely applicable. An online individualized prediction model is being developed. Clinicians without special training, or even patients themselves, will be able to predict tumor survival and recurrence through online operations in the near future.
The present study contains several limitations. A limited sample size may affect the model training. With the widespread application of the individualized prediction model, the parameters and predictive factors of the model may need to be updated to achieve higher prediction accuracy. The limited sample size also leads to the deviation of results in the relationship between the recurrence score and clinicopathological characteristics. Besides, the calculation of recurrent score requires a test kit, which increases the workload of pathologists as well as the cost of patients. Therefore, the test kits need to be more convenient and cheaper or replaced by other methods, such as radiomics. But it is worth noting that convenience and cost are contradicted, and we need to constantly explore the best balance in clinical applications.
In conclusion, the current research not only provides a tool for the objective assessment of the recurrence probability and survival rate of postoperative LGG, but also provides a theoretical basis for the targeted therapy of recurrent LGG. The individualized prediction model is simple and accurate enough to be widely applied to a broad clinical setting.