Development, validation, and updating of prognostic models for m7G- associated genes in lower-grade gliomas

doi:10.21203/rs.3.rs-3785800/v1

Download PDF

Research Article

Development, validation, and updating of prognostic models for m7G- associated genes in lower-grade gliomas

https://doi.org/10.21203/rs.3.rs-3785800/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Studies are aiming at developing prognostic models using N7-methylguanosine (m7G)-related genes in gliomas, however, models with good predictive performance for lower-grade gliomas have yet to be developed.

Methods

Based on genes with m7G variants and clinical information, two prediction models have been derived to predict the probability of survival for patients with lower-grade gliomas in TCGA. The models were externally validated using independent datasets. Based on CGGA information, updated models that were created matched the features of the local population.

Results

Two models were derived, validated and updated. Model 1, which was derived on the basis of mRNA, only contains five genes: CD37, EIF3A, CALU, COLGALT1, and DDX3X. Model 2 included six variables: grade, age, gender, IDH mutation status, 1p/19q codeletion status and prognostic index of model 1. The C-statistic of revised model 1 was 0.764 (95%CI: 0.730–0.798) in the revised set and 0.700 (95%CI: 0.658–0.742) in the test set. Regarding internal validation, C-statistic for model 2 with 1000 bootstrap replications was 0.848, while in external validation, the C-statistic was 0.752 (95%CI: 0.714–0.788). Both models exhibited satisfactory calibration after updating in external validation. The models' web calculator is provided at https://lhj0520.shinyapps.io/M7G-LGG_model/.

Conclusion

we developed and validated two models and updated them, which makes the models better predictors for patients.

Prognostic model

lower-grade gliomas

m7G

model updating

tumor-associated macrophages/microglia

Diffuse gliomas, including glioblastomas and World Health Organization (WHO) grade II and III tumors (sometimes referred to as lower-grade gliomas, LGG), are the most common malignant brain tumors among adults (Ostrom, Patil et al. 2020, Mair, Geurts et al. 2021). Lower-grade gliomas account for 43.2% of gliomas in the central nervous system (Jiang, Nam et al. 2021). Secondary glioblastomas evolve from low-grade gliomas, most often type II or III lower-grade gliomas. Lower-grade gliomas have variable prognostic outcomes and a wide range of survival outcomes (Gittleman, Sloan et al. 2020), therefore, robust clinical prediction models might be helpful for physicians in treatment of patients.

Gliomas are complex tumors composed of tumor and non-tumor cells (Hambardzumyan, Gutmann et al. 2016). About 30–50% of cells in gliomas are microglia or macrophages, referred to as tumor-associated macrophages/microglia (TAMs) (Tan, Li et al. 2020). Pathologically, TAMs support glioma growth and progression via immunosuppression and enhancing proliferation, migration, invasion as well as angiogenesis (Jiang, Nam et al. 2021). Given cell heterogeneity, there is a need to investigate the prognostic impact of TAMs in lower-grade gliomas.

The N7-methylguanosine (m7G) is the seventh N to modify RNA guanine with a methyl group (Zhang, Song et al. 2021). Physiologically, M7G is a common modification on mRNAs, tRNAs, rRNAs, and miRNAs (Zhang, Song et al. 2021). The m7G methylation is closely associated with tumor development and is implicated in various tumor-related biological activities (Luo, Yao et al. 2022). M7G methyltransferases add m7G modifications to target RNAs, influencing RNA synthesis, structure, and maturation and, as a result, mediating various essential biological processes (Luo, Yao et al. 2022). The m7G-related gene variations in glioblastoma are associated with patient survival and are potential therapeutic targets in glioblastoma (Wu, Li et al. 2022). Various prognostic models for m7G-related genes in lower-grade gliomas have emerged, however, the models are not well developed (Wang, Zhong et al. 2022). Furthermore, m7G-related genes are directly and extensively correlated with tumor microenvironment and immune cells (Wu, Li et al. 2022, Zhou, Yang et al. 2022).

Genetic variants influence RNA modifications by specifically substituting the nucleotide at the changed position or by changing the sequence of nucleotides in the proximal flanking region (Ramaswami, Deng et al. 2015). Dysregulated RNA modification may play a role in cancer regulation (Woo and Chambers 2019, Li, Xia et al. 2020). However, the effects of variants on m7G methylation modifications in lower-grade gliomas have not been conclusively determined. Therefore, we aimed at identifying genes with m7G methylation variant sites and to investigate their impact on LGG prognosis.

We combined single-cell RNA sequencing and bulk RNA sequencing data to search for m7G variant genes associated with TAMs. Then, the identified genes were used to build a prognostic model for low-grade gliomas (Model 1). A second model was formed by extending Model 1 with addition of clinical variables. The two models were validated and appropriately updated against a CGGA dataset that matched the characteristics of the local population. With user convenience in mind, we also constructed an online web platform that includes both models.

2.1 Dataset collection and study population

The research design and workflow of this study is shown in Fig. 1.

Single-cell RNA sequencing (scRNA‑seq) data and the corresponding clinical information were downloaded from the Chinese Glioma Genome Atlas (CGGA) website (Yu, Hu et al. 2020). Four additional datasets with mRNA data and the corresponding clinical information were used in the model. The datasets were acquired from three databases: The Cancer Genome Atlas (TCGA), Gene Expression Omnibus (GEO), and CGGA.

The TCGA-LGG cohort was downloaded from the GDC Data Portal. A total of 496 patients met the inclusion criteria: primary tumor, WHO stage II and III, matched clinical and genetic information, and complete overall survival (OS) outcomes. The inclusion criteria were consistent for all datasets used in this study. The GDC Data Portal was also used to download somatic mutation data in the mutation annotation format (MAF) for LGG patients from TCGA for mutation analysis of genes.

The GSE16011 data set (the microarray data) was downloaded from GEO and collated with the required data information from 106 patients (Gravendeel, Kouwenhoven et al. 2009).

Two bulk-RNA sequencing and clinical data sets (mRNAseq_693 and mRNAseq_325) were acquired from the CGGA database (Yan, Zhou et al. 2021). After removing the batch effect by the “ComBat” function in “sva” R package (version 3.44.0), the two CGGA datasets were combined into one dataset (CGGA data set) (Johnson, Li et al. 2007). Information for 406 patients was available after collation.

2.2 Bioinformatics analysis

Four patients (S7, S8, S9 and S14) with lower-grade gliomas were only included in the CGGA single-cell sequencing dataset, all of which were retained to screen for marker genes for TAMs in LGG (Yu, Hu et al. 2020, Yan, Zhou et al. 2021). Single-cell sequencing data analysis was performed using the R package “Seurat” (version 4.3.0) (Yan, Zhou et al. 2021). Genes with expressions between 200–2500 and those with less than 5 expressed in mitochondria were retained for data quality assessment. The "vst" method was used to identify the top 2000 high-variance genes as variable characteristics. Principal component analysis (PCA) was performed to reduce data dimensions based on identified highly variable genes, from which the best principal components were used for t-SNE analyses. Based on source literature of the dataset, TAM type clusters were annotated (Yu, Hu et al. 2020) and the differentially expressed genes (DEGs) for TAMs and other clusters computed. The threshold for DEGs was adjusted by p value < 0.05 and |log₂FC|>1.5.

We screened for m7G variant-associated genes that were common in the four mRNA datasets (TCGA, CGGA, GSE16011 and RMVar). Spearman rank correlation was used to calculate correlation coefficients for the two gene sets, DEGs and m7G mutation-associated genes. Final gene numbers were obtained based on filtering criteria of an absolute rank correlation coefficient > 0.6 and adjusted p < 0.05.

The R package “maftools” (version 2.12.0) was used to analyze the TCGA somatic mutation data (Mayakonda and Koeffler 2016). ESTIMATE was used to predict the presence of stromal cells and immune cells in the tumor tissue (Yoshihara, Shahmoradgoli et al. 2013), while CIBERSORT was used to calculate the proportions of 22 immune cells derived from bulk tumors on the basis of gene expressions (Newman, Liu et al. 2015).

2.3 Statistical analysis

The TCGA and CGGA RNA-sequencing data were converted to log2(TPM + 1) format before data analysis. The microarray data (GSE16011) were processed via the Robust Multichip Average (RMA) method.

The TCGA-LGG cohort (n = 496) was used as the derivation set, while the GSE16011 data set (n = 106) and the CGGA data set (n = 406) were used as validation sets in model 1. The CGGA data set (n = 408) was used as the validation set in model 2. Before fitting and evaluating the model, the “ComBat” function was used to remove batch effects in a uniform manner across all datasets to reduce bias (Leek, Scharpf et al. 2010, Tang, Ji et al. 2021).

Patient characteristics are presented as median and interquartile ranges (IQR) for continuous variables while categorical variables are presented as counts and percentages. Overall survival outcomes are presented by Kaplan-Meier curves. The median follow-up time was calculated using the reverse Kaplan-Meier method (Mathoulin-Pelissier, Gourgou-Bourgade et al. 2008). Model predictions used OS as the event outcome. Event per variable (EPV) in the model was used to assess sample size adequacy (Moons, Altman et al. 2015). Prior to computing the correlation coefficient, one must first compute uniformly whether the data conforms to the bivariate normal distribution, otherwise, one uses the Spearman rank correlation method. p < 0.05 was the threshold for statistical significance, unless otherwise stated. The R software (version 4.2.1) was used for analyses.

2.4 Model development, validation

Two models have been developed: a genomics model that includes only mRNA (model 1) and model 2 obtained by an extension of model 1 using clinical variables. The genomic data were not missing, thus, imputation was not considered. The derivation and validation sets were missing some values for the clinical variables. To impute the missing values for clinical variables, the R package “mice” (version 3.14.0) was used for multiple imputation, based on the percentage of missing values in the derivation set (10 times) (Murad, Dankner et al. 2020). For the multiple imputation model, the variables used included 9 candidate predictors (prognostic index for model 1, age, sex, grade, radiation therapy, pharmaceutical therapy, IDH mutation status, 1p/19q codeletion status, and MGMT methylation status) and two outcome indicators (Nelson–Aalen estimator of the baseline cumulative hazard and outcome events) (Moons, Donders et al. 2006, White and Royston 2009).

Genes were filtered for the presence of variables using the least absolute shrinkage and selection operator (LASSO) in “glmnet” package of R (Goeman 2010). Ten-fold cross validation was used to select the optimal value of λ, and the corresponding variables with non-zero coefficients kept. For imputed data sets, variable selection was performed in the stacked data, which was formed by stacking multiply imputed data on top of each other. For each imputed set, optimal penalty (lambda) was determined to find the mean lambda, which was used as the final optimal penalty for LASSO for variable selection in the stacked data (Steyerberg 2009). Weighted Cox regression was fitted to the stacked data to estimate predictor coefficients (1/10 weight per patient) (Wood, White et al. 2008). Coefficients can be well estimated by stacking the data, however, this method can underestimate the standard errors (Steyerberg 2009). Therefore, performance of the model was assessed by parallel analysis of 10 imputed datasets, combined with results from each dataset and Rubin’s rules (Rubin 2004, Marshall, Altman et al. 2009, Wang, Yuan et al. 2021).

Coefficients corresponding to variables were obtained by fitting the Cox multivariable regression models. The prognostic index (PI), also referred to as the linear predictor, was extracted from the built Cox model. It is a weighted sum of the model's variables, with weights corresponding to coefficients of regression (Royston and Altman 2013).

Overall performance measures for the prognostic model were assessed by calibration and discrimination (Steyerberg, Vickers et al. 2010). Discrimination and calibration of the model were separately assessed in derivation and validation sets. Discrimination refers to the ability to discriminate between high- and low-risk patients and is measured by Harrell’s overall concordance (C) statistic and time dependent ROC (Harrell Jr, Lee et al. 1996, Steyerberg, Vickers et al. 2010, Wang, Yuan et al. 2021). Indicator values vary between 0.5 (no predictive discrimination) and 1.0 (perfect discrimination) (Wang, Yuan et al. 2021). The C-statistic of the model was calculated using the “validate” function in R package “rms” (version 4.3.0) (Harrell 2001). Time dependent ROC curves at 1, 3, and 5 years were created using the R package “survivalROC” (version 1.0.3) (Heagerty, Lumley et al. 2000). Time dependent C-statistic of the models were visualized using the “cindex” function in R package “pec” (version 2022.05.04) (Gerds, Kattan et al. 2013).

Calibration refers to the agreement between observed and predicted outcomes (Hilden, Habbema et al. 1978). Calibration plots can be used to assess calibration. In this study, patients were assigned into several groups according to sample size, and the predicted survival probabilities (X-axis) compared with observed survival probabilities (Y-axis) at 1, 3, and 5 years. We further performed a more sensitive calibration degree assessment (Houwelingen 2000). Specifically, patients were assigned into two risk groups based on the median of PI. Then, we compared the observed Kaplan-Meier survival curves and the average predicted probability curves for each risk group. Given the inconvenience of later updating the model in terms of applying the function, calibration plots were plotted using the R code, as described by Wang (Wang, Yuan et al. 2021). We also estimated the calibration slope, i.e., the regression coefficient of PI in the external validation set (Royston and Altman 2013, Stevens and Poppe 2020, Wang 2020). The calibration slope reflects whether the impact of predictors (regression coefficients) in the original model is over-or under-estimated (Janssen, Moons et al. 2008). A calibration slope of less than 1 indicates that regression coefficients are too large and may result in extreme predictions for new patients, such as weaker predictions for low risk and stronger predictions for high risk (Janssen, Moons et al. 2008).

Validation of the prediction model includes internal and external validation. Internal validation is aimed at testing the reproducibility of the construction process of the model in a population with similar sources of derivation sets (Steyerberg, Harrell Jr et al. 2001). To correct for optimism in model performance, internal validation was performed 1000 times via a bootstrap resampling procedure. External validation was used to verify whether the model has good predictive performance in data independent of the derivation set, as revealed by assessing differentiation and calibration of the model in the external validation set.

2.5 Model updating

When the model performs poorly in external validation, we tended to make some adjustments/updates to the model based on information from the validation set (Van Calster, McLernon et al. 2019). After updating, the model should be revalidated using an independent data set (Royston and Altman 2013). By estimating the calibration slope and the cumulative baseline hazard in the external validation set, this allows the determination on how to update the model: the calibration slope only, the baseline hazard only, or both (Van Calster, McLernon et al. 2019, Wang, Yuan et al. 2021). Two main approaches for updating the model have been proposed: recalibration and revision methods (Steyerberg, Borsboom et al. 2004, Janssen, Moons et al. 2008). In external validation, model recalibration improves model calibration, and model revision improves both model discrimination and calibration (Toll, Janssen et al. 2008).

Model revision was performed on model 1. Two validation sets were used in model revision: updating set (in which the model is updated based on its information) and the test set (in which the update model is validated). In the updating set, the Cox regression model was re-fitted to obtain new regression coefficients based on original predictors (Janssen, Moons et al. 2008). Taking into account the large sample size requirement for the proposed model revision method (Steyerberg, Borsboom et al. 2004), we combined the TCGA-LGG cohort (derivation set, n = 496) and the GSE16011 dataset (as updating set, n = 106) into a larger “revision set” (n = 602), after which the regression coefficients of the model were re-fitted in the revision set. Then, we observed and compared the baseline survival probability of the revision and test sets. If the difference was too large, we used the baseline survival probability (\({\text{S}}_{0}\left(\text{t}\right)\)) from the test set when calculating the predicted survival probability (\(\text{S}\left(\text{t}|\text{X}\right)\)). Predicted survival probability of the model was calculated by: \(\text{S}\left(\text{t}|\text{X}\right)={{\text{S}}_{0}\left(\text{t}\right)}^{\text{e}\text{x}\text{p}\left(\text{P}\text{I}\right)}\). Baseline survival probability was estimated as \({\text{S}}_{0}\left(\text{t}\right)=\text{e}\text{x}\text{p}[-{\text{H}}_{0}\left(\text{t}\right)]\), where \({\text{H}}_{0}\left(\text{t}\right)\) is the baseline cumulative hazard (Royston and Altman 2013). The baseline cumulative hazard can be computed using the “basehaz” function in “survival” R package (version 3.4.0). The CGGA dataset (n=408) was used as the test set, with at least 100 events and a sufficient amount of power to detect differences in prediction model performance between the two datasets (Vergouwe, Steyerberg et al. 2005, Janssen, Moons et al. 2008). The revised model was re-evaluated in both revision and test sets.

We re-calibrated model 2 based on baseline survival probability and calibration slope of the validation set (Janssen, Moons et al. 2008, Moons, Kengne et al. 2012). We used the \({\text{S}}_{0}\left(\text{t}\right)\) from the validation set rather than the derivation set to calculate the predicted survival probability in the validation set. To obtain the final regression coefficient, the regression coefficient for each predictor variable was multiplied by the calibration slope. After re-calibration, model calibration was re-evaluated in the validation set.

2.6 The Online Calculator

To make it easier for users to use the two models, they can be obtained from the online calculator website: https://lhj0520.shinyapps.io/M7G-LGG_model/. The site includes single person predictions for each model as well as batch predictions, and a tool for removing batch effects with the model derivation set (TCGA).

3.1 Data Summary

Patient characteristics are summarized in Table 1.

Table 1

patient characteristics^*
	Derivation Cohort	Validation Cohorts
Characteristics	TCGA(n = 496)	CGGA(n = 408)	GSE16011(n = 106)
Age (year)	41.0(33.0,53.0)	40.0(34.0, 46.0)	43.9(37.1, 54.8)
Gender (%)
Female	221 (44.6)	170 (41.7)	36 (34.0)
Male	275 (55.4)	238 (58.3)	70 (66.0)
Grade (%)
II	238 (48.0)	220 (53.9)	22 (20.8)
III	258 (52.0)	188 (46.1)	84 (79.2)
IDH mutation status (%)
Wild Type	92 (18.5)	100 (24.5)	38(35.9)
Mutant	402 (81.1)	277 (67.9)	47(44.3)
Unknown	2(0.4)	31(7.6)	21(19.8)
1p/19q co-deletion status (%)
Non-codeletion	333 (67.1)	243 (59.6)	38 (35.8)
Co-deletion	163 (32.9)	131 (32.1)	33 (31.2)
Unknown	0	34(8.3)	35(33.0)
MGMT promoter (%)
Unmethylated	86 (17.3)	150 (36.8)	-
Methylated	410 (82.7)	191 (46.8)	-
Unknown	0	67(16.4)	-
Radiation Therapy (%)
No	160 (34.3)	80 (19.6)	30 (28.3)
Yes	306 (65.7)	320 (78.4)	76 (71.7)
Unknown	3(0.6)	8(2.0)	0
Pharmaceutical Therapy (%)
No	178 (38.7)	158 (38.8)	63(59.4)
Yes	282 (61.3)	236 (57.8)	16(15.1)
Unknown	36(7.2)	14(3.4)	27(25.5)
Median follow-up (years)	2.3(2.1,2.6)	7.4(6.9,7.8)	NA**
Number of deaths (%)	123(24.8)	164(40.9)	91 (85.8)
^*Categorical variables are presented as n (%), and continuous variables are presented as median (interquartile range).
^** NA represents surviving subjects who did not exceed the median follow-up time

In the derivation cohort, which had a total of 496 patients, there were 123 (24.8%) mortalities, with a median follow-up time of 2.3 years (95%CI: 2.1–2.6). In the derivation set, the EPVs for model 1 and model 2 were 6.15 (123/20) and 13.67 (123/9), respectively. There were 408 patients in the CGGA data set, with 164 (40.9%) mortalities and a median follow-up time of 7.4 years (95%CI: 6.9–7.8). For the GSE16011 dataset, a total of 106 patients were available, with 91 (85.8%) mortalities and a follow-up time of 13.3 years for the last survivor.

Regarding the clinical data, there was a low percentage (7.2%) of missing values in the derivation set (Fig. S1A) and a low percentage (8.3%) of missing values of predictors in the CGGA data set (Fig. S1B). In the GSE16011 dataset, one of the key predictors (1p/19q co-deletion status) was missing up to 33% (Fig. S1C), which is the main reason why we did not select this dataset as a validation set for model 2.

Cumulative event rates for the two models are shown in Fig. S2. In model 1, patients in the revision set had a higher cumulative mortality rate than patients in the testing set (Fig. S2A). Model 2 also had a higher cumulative mortality rate for patients in the derivation set than in the validation set (Fig. S2B).

3.2 Identification of TAM cell marker gene expression profiles

The retrieved data were subjected to quality assessment, and 142 samples from four patients (3 WHO grade II, 1 WHO grade III) with 19667 features from CGGA single-cell sequencing data were obtained (Yu, Hu et al. 2020). We used PCA to reduce the dimensionality of the top 2000 variable genes and retained the top four principal components based on the heatmap (Fig. 2A). Four cell clusters were clustered by t-SNE (Fig. 2B), and marker genes cited in the original literature revealed that cluster 0 was classified as TAMs (Fig. 2C) (Yu, Hu et al. 2020).

3.3 Identification of TAMs-associated m7G-related genes

There were 120 DEGs between cluster 0 and other clusters of clustered cells (Fig. 3A). A total of 118 genes that intersected in the TCGA gene dataset were retained. There were 66 common m7G-related genes from the four databases (RMVar, CGGA, TCGA and GSE16011) (Fig. 3B). Correlation coefficients for 118 TAM marker genes and 66 m7G related genes were calculated, and 20 genes that satisfied the threshold conditions (|Rs|>0.6 and adjusted p < 0.05) retained.

3.4 Derivation, validation and revision of the gene prediction model (Model 1)

The 20 genes were screened by LASSO (Fig. 3C-D) and five predictors were retained to fit the model (CD37, EIF3A, CALU, COLGALT1, and DDX3X). Table S1 shows the Ensemble ID and Symbol ID corresponding to the five predictors of model 1.

The original model was first fitted in the derivation set and externally validated in each of the two validation sets. In internal validation, the apparent C-statistic for the original model was 0.786 (95%CI: 0.742–0.832), while the optimism-corrected C-statistic with 1000 bootstrap replications was 0.779. Evaluation performance of the original model on derivation and validation sets was as shown in Fig. 4. The original model performed well in both discrimination (Fig. 4A) and calibration (Fig. 4B-C) in the derivation set, but exhibited a significant decrease in performance in the validation sets (Fig. 4D-I) especially with regards to calibration. In Fig. 4E-F, survival probability in the CGGA dataset is underestimated. For the GSE16011 dataset, patient survival is markedly overestimated (Fig. 4H-I). In the CGGA dataset, the C-statistic of the original model was 0.691(95%CI:0.648–0.734) while that for the GSE16011 dataset was 0.717(95%CI: 0.663–0.770). The calibration slope for the original model was 0.831 in the CGGA dataset and 0.667 (far from 1) in GSE16011 dataset, which is a reflection of the poor predictive ability of the original model in the validation sets.

Given poor performance of the original model on the validation set, we revised it to achieve better predictive power in the validation set. Regression coefficients for the five prognostic genes were re-estimated using the revision set (TCGA + GSE16011, n = 602) and the final PI obtained as:

The updated model 1 required re-evaluation of model performance in the revision and test set (CGGA set, n = 408). The C-statistic for the revised model was 0.764 (95%CI: 0.730–0.798) in the revised set and 0.700 (95%CI: 0.658–0.742) in the test set. It had a calibration slope of 1 in the revision set and 0.998 (close to 1) in the test set, indicating that the revised model had significantly improved in terms of prediction discrimination.

The AUCs for the model in the revision set at 1, 3, and 5 years were 0.854, 0.819, and 0.762, respectively (Fig. 5A). In the revision set, its calibration curve lies close to the 45° dotted line (Fig. 5B). For the two risk groups based on median of PI as the cut-off, the predicted mean survival curve for each group (dashed line) closely approximated the observed K-M survival curve (solid line) (Fig. 5C). Both points indicate that the model was performing well in calibration of the revision set. The AUCs in the test set at 1, 3, and 5 years were 0.725, 0.799, and 0.724, respectively (Fig. 5D). In Fig. 5E-F, the calibration ability of the model has a certain bias.

Since the cumulative event rate for the revision set differed from that of the test set (Fig. S2B), we considered adjusting the baseline hazard of the revised model to the baseline hazard of the test set. Note that adjusting for model baseline risk does not affect the model discrimination, but improves calibration.

The revised model's temporal dependence C-statistics from 1 to 10 years in both datasets are shown in Fig. 5G, both decreasing over time but still close to 0.7. Figure 5H-I shows the revised model's calibration performance after adjustment. The tuned revised model achieved satisfactory prediction performance on the test set.

3.5 Derivation, validation and recalibration of Model 2

We performed LASSO on each imputed dataset and obtained the optimal lambda values in each cross-validation (10 in total) (Fig. S3A). The mean for the ten lambda values was 0.01685655, and LASSO was used in the stacked data set to filter variables with the mean of lambda values as the optimal penalty (Fig. S3B). Finally, six predictors were included in the model: grade, age, gender, IDH mutation status, 1p/19q codeletion status, PI_model1.

In internal validation, the apparent C-statistic for model 2 was 0.849 (95%CI: 0.817–0.881), while the optimism-corrected C-statistic with 1000 bootstrap replications was 0.848. Evaluation performance for model 2 on derivation and validation sets was as shown in Fig. 6. The AUCs for the model in the derivation set at 1, 3, and 5 years were 0.888, 0.895, and 0.861, respectively (Fig. 6A). Calibration of the model in the derivation set was satisfactory (Fig. 6B-C). In external validation, C-statistic for the model was 0.752 (95%CI:0.714–0.788) while that of the calibration slope was 0.801. The AUCs for the model in the validation set at 1, 3, and 5 years were 0.725, 0.801, and 0.740, respectively (Fig. 6D). Even though the calibration for model 2 was good at 1, 3, and 5 years (Fig. 6E), over time, there was a significant decrease in calibration in both risk groups (Fig. 6F). The time-dependent C statistic for model 2 revealed good performance in both datasets, maintaining above 0.7, even in the tenth year (Fig. 6G).

Due to large calibration deviation of the model in the validation set, the complex revision method reduced the discrimination for model 2. Therefore, model recalibration was considered based on calibration slope and baseline hazard in the validation set. Final coefficients for prognostic factors in the model are equal to the original Cox regression coefficient multiplied by the calibration slope. The final PI was obtained as:

After model recalibration, the calibration for model 2 in the validation set was reassessed. In Fig. 6H, the calibration for model 2 was improved and was more pronounced in the risk group survival curves (Fig. 6I), indicating that this model has a satisfactory predictive power on the validation set and that it can be considered for extrapolation to similar study populations. Fig. S3C-D shows the distributions of prognostic indices for the two models within the data sets.

Both models can be obtained at a website (Li 2022) (https://lhj0520.shinyapps.io/M7G-LGG_model/), and the website operation instructions are given in the supplementary materials (p8-10). Therefore, it is easier for other researchers to further validate or use our models. Baseline survival rates for the two models from 1 to 10 years are shown in Table S2.

3.6 Somatic mutation and immune analyses

The TCGA-LGG population was assigned into two groups based on the median for PI_{model 1}, and waterfall plots for the top 15 genes with the highest number of mutations in both risk groups plotted (Fig. 7A-B). Nine hypermutated duplicate genes were identified in both risk groups (IDH1, PIK3CA, TTN, NOTCH1, FUBP1, CIC, ATRX, TP53). We calculated the tumor mutational burden (TMB) for the population (Fig. 7C), which revealed that it was positively correlated with PI_{model 1} (Fig. 7D) (R = 0.339, p < 0.001). Regarding percent expressions for 22 immune cells in each sample, monocytes and macrophages were highly prevalent (Fig. S4A). Correlations between immune cells and prognostic genes were calculated again, and statistically significant positive correlations were found for three of the genes (CD37, CALU, COLGALT1) as well as PI_{model 1} with macrophages M2 (Fig. S4B). With respect to the tumor microenvironment, PI_{model 1} was positively correlated with all 3 scores: Stromal Score (R = 0.470, p < 0.001), Immune Score (R = 0.428, p < 0.001), ESTIMATE Score (R = 0.459, p < 0.001) (Fig. S4C).

This study aimed at developing and validating two clinical prediction models, based on m7G-variant-related genes associated with TAMs and clinical variables. The models were updated with new information that better fits the characteristics of the local population, allowing the final models to obtain better predictive power. This study is reported according to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Statement (TRIPOD) (Moons, Altman et al. 2015). The complete checklist is shown in Table S3.

Five genes exhibited a good prognostic capacity. Two variants were associated with the m7G modifier site in each gene, and mutations resulted in the loss of modifier sites in all genes (Table S4) (Posternak, Ung et al. 2017, Zhang, Liu et al. 2019, Luo, Li et al. 2021). Confidence levels for all m7G sites within five genes were moderate and were obtained based on previously published MeRIP-seq data (Luo, Li et al. 2021). Based on disease-related data collected from GWAS and ClinVar databases, variants in the five genes were pathogenic mutations, resulting in dysregulation of m7G modifications in LGG (Luo, Li et al. 2021). The RBPs corresponding to m7G-associated SNPs in the RBP binding region in five genes were also obtained from RMVar database (Fig. S5A).

Studies have assessed the significance of five prognostic factors in glioma. A pan-cancer study found that CD37 is aberrantly expressed in various cancers, and that it plays an important immunosuppressive role in diffuse gliomas (Yan, Zhou et al. 2021). The CD37 promotes polarization of M2 macrophages and promotes the malignant progression of glioma (Yan, Zhou et al. 2021). Expressions of the EIF3A gene have been associated with cell cycle progression, migration, invasion and colony formation in glioma cancer cells (Abdul Rahman, Mokhtar et al. 2019). CALU acts as a potential serum-specific marker in glioblastoma, binds calcium ions in the endoplasmic reticulum, and participates in protein folding and sorting (Sreekanthreddy, Srinivasan et al. 2010). However, its regulatory role in LGG is unclear. DDX3X is a valuable biomarker in human glioma (Hueng, Tsai et al. 2015). The significance of COLGALT1 in glioma has yet to be determined, however, its mutation leads to defects in small blood vessels in the brain and brain perforation defects.

The models developed in this study differ from the published 3 prognostic LGG models, which were developed using m7G-related genes (Table S5) (Chen, Zhang et al. 2022, Shao, Li et al. 2022, Wang, Zhong et al. 2022). First, the selected candidate gene sets differed. The m7G regulators were selected by previous authors, whereas our gene set was selected for genes predicted to affect m7G modification upon mutations. Therefore, we computed the correlation coefficients between the latest regulators and our prognostic genes, which revealed significantly high correlation coefficients (Fig. S5B). Second, we focused more on the normative processes of model derivation, validation and updating, accompanied by a detailed evaluation report (TRIPOD). Third, the previously published studies did not detail how the missing data were handled. To address the missing data, we reviewed a large body of literature and implemented a comprehensive set of procedures, including multiple imputation, modeling, and evaluation of imputed data. Fourth, some of the previous studies succinctly possessed 3 prognostic genes but had poor discrimination in external validation (maximum was 0.705). One study reported a good external validation discrimination (0.788–0.873) but had up to 15 predictor variables (Shao, Li et al. 2022). They all lacked the assessment of external validation calibration. One of our two models had 5 predictors while one model had 10 predictors. After the update, the externally validated ROCs were maintained at 0.7–0.8 and were well calibrated.

A notable aspect of this study is that the latest models utilized both prior knowledge obtained from the derivation set and newly acquired data from the validation set. Once the model is updated, the model has good predictive performance on the validation set. For the validation set (CGGA), we included data from the Chinese population, implying that the model constructed in this study was more applicable to the local patients in China. To facilitate the preprocessing of the genetic data set, we created an additional web page on our website, enabling users to upload their own genetic data for uniform batch effect removal (the core is the "ComBat" function in R software), to reduce the bias caused by batch effect. We also provided interactive pages for individual prediction and a batch of patients’ prediction for the two uploaded models, which were user friendly. Given that the treatments were not included in model 2, we further performed sensitivity analysis which revealed that the model 2 still exhibited good predictive performance in different treatment populations (Fig. S6-Fig. S7). The model also had satisfactorily discrimination and calibration in the population which received radiotherapy (Fig. S6A-C), but model performance was somewhat erratic in the population that did not receive radiotherapy (Fig. S6D-F), possibly due to the small sample size (n = 80). Its performance in both populations which received pharmaceutical therapy (Fig. S7A-C) and those not administered with pharmaceutical therapy was good (Fig. S7D-F).

This study has some limitations. First, model 2 contained prognostic factors, including 5 clinical variables and 5 genes. Therefore, to reduce the complexity of using the model, we developed an online web calculator. Second, the generalizability of the models still needs to be improved. Although the models were constructed and validated using large publicly available databases, the sample size is not particularly large compared to that of widely available models. For this purpose, we provided some model parameters and formulae in the analysis which may need to be updated in subsequent investigations. Third, as observed, none of our original models exhibited accurate predictions on the validation set. This can be attributed to two possible factors: inadequate sample sizes and lack of representativeness in both the derivation and validation sets. There are difference among the datasets, obtained from the US (TCGA), China (CGGA) and Netherlands (GSE16011), in terms of population characteristics. For this reason, the dataset with national population characteristics was applied in the model revision and recalibration. Fourth, few data are available for the mutation loci of the five prognostic genes. Thus, we studied the protein-coding genes that affect m7G modification after mutation, but few mutant loci were found for these five genes from the TCGA database. Follow up experiments should further examine the effects of mutations in these genes. Sixth, the sample size enrolled in the single-cell dataset was too small and the marker genes for TAMs may not be representative. Before the study, different datasets from multiple databases, but the results were not better. This problem needs to be resolved in subsequent studies.

In this work, we identified 20 differentially expressed genes in TAMs containing m7G in their RNA. Five genes were used to fit the model 1/genetic, which was initially studied in a derivation cohort (TCGA) and then validated in two validation cohorts (TCGA and GSE16011). Model 1 optimization was performed using a “revision set” which comprised patients from the TCGA and GSE16011 datasets and was validated in a “test set” using data from the CGGA dataset. By integrating clinical variables into the initial prognostic index, the authors created a second model that was trained and validated using separate cohorts: TCGA for derivation and another for validation (CGGA). After appropriate updates to models, we believe that the latest models have the potential to be applied for predicting the Chinese cohort data.

LGG

Low grade glioma

TAMs

Tumor-associated macrophages/microglia

AUC

Area under the ROC curve

Confidence interval

ROC

Receiver operating characteristic

overall survival

Contribution to the field statement

Accurate prediction of survival in LGG patients is important for clinical decision making, and the development of a model with high predictive performance and ease of use in the clinical setting is of great practical importance. We have followed the TRIPOD guidelines to developed two prediction models with good discrimination and satisfactory accuracy that can be used for genomics or for the combination of genetic and clinical information. Further, models were updated on the basis of the validation set in order to provide better predictive power and generalizability to new patients. In particular, it should be noted that we also provide an easy-to-use online calculator containing two models and preprocessing of the genetic data. LGG prediction models using genes with mutations affecting RNA modifications as predictors have not been available, our study is groundbreaking and these genes with predictive value are worthy of further exploration by the biological community.

Funding:

This work was supported in part by the National Natural Science Foundation of China (81773541), funded from the Priority Academic Program Development of Jiangsu Higher Education Institutions at Soochow University, the State Key Laboratory of Radiation Medicine and Protection (GZK1201919) to ZXT, A Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD) to ZXT, The study was supported by National Natural Science Foundation of China (CN) (82172441) (Ke Lu) and (82273472) (Lei Shi), Kunshan High-level Medical Talent Post Training Project (2018), and Suzhou health personnel training project (GSWS2020112) (Lei Shi), Suzhou Medical Center(Szlcyxzx202103).

Author Contribution

HJL, YFD, LS, KL for literature search; HJL, JJS, ZXT for figures ; HJL, ZXT, JJS, HS for study design ; HJL, JJS, HS for data collection ; HJL, RRG for data analysis ; HJL, RRG, JJS, ZXT for data interpretation ; HJL, RRG, JJS for writing. All authors read and approved the final manuscript.

Acknowledgments

The authors thank the patients and investigators who participated in TCGA, CGGA, GEO and RMVar for providing the data.

Data availability statement

All data used in this article are from publicly available databases.

TCGA Database: https://xenabrowser.net/

CGGA Database: http://www.cgga.org.cn/

GEO Database: https://www.ncbi.nlm.nih.gov/gds/

RMVar Database: https://rmvar.renlab.org/

CellMiner Database: https://discover.nci.nih.gov/cellminer/home.do

Conflict of Interest

All authors declare that they have no conflicts of interest.

Abdul Rahman, A., N. M. Mokhtar, R. Harun, R. Jamal and W. Z. Wan Ngah (2019). "Transcriptome analysis reveals the molecular mechanisms of combined gamma-tocotrienol and hydroxychavicol in preventing the proliferation of 1321N1, SW1783, and LN18 glioma cancer cells." J Physiol Biochem 75(4): 499-517.
Chen, Z., Z. Zhang, W. Ding, J. H. Zhang, Z. L. Tan, Y. R. Mei, W. He and X. J. Wang (2022). "Expression and Potential Biomarkers of Regulators for M7G RNA Modification in Gliomas." Front Neurol 13: 886246.
Gerds, T. A., M. W. Kattan, M. Schumacher and C. Yu (2013). "Estimating a time‐dependent concordance index for survival prediction models with covariate dependent censoring." Statistics in medicine 32(13): 2173-2184.
Gittleman, H., A. E. Sloan and J. S. Barnholtz-Sloan (2020). "An independently validated survival nomogram for lower-grade glioma." Neuro Oncol 22(5): 665-674.
Goeman, J. J. (2010). "L1 penalized estimation in the Cox proportional hazards model." Biom J 52(1): 70-84.
Gravendeel, L. A., M. C. Kouwenhoven, O. Gevaert, J. J. de Rooi, A. P. Stubbs, J. E. Duijm, A. Daemen, F. E. Bleeker, L. B. Bralten, N. K. Kloosterhof, B. De Moor, P. H. Eilers, P. J. van der Spek, J. M. Kros, P. A. Sillevis Smitt, M. J. van den Bent and P. J. French (2009). "Intrinsic gene expression profiles of gliomas are a better predictor of survival than histology." Cancer Res 69(23): 9065-9072.
Hambardzumyan, D., D. H. Gutmann and H. Kettenmann (2016). "The role of microglia and macrophages in glioma maintenance and progression." Nat Neurosci 19(1): 20-27.
Harrell, F. E. (2001). Regression modeling strategies: with applications to linear models, logistic regression, and survival analysis, Springer.
Harrell Jr, F. E., K. L. Lee and D. B. Mark (1996). "Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors." Statistics in medicine 15(4): 361-387.
Heagerty, P. J., T. Lumley and M. S. Pepe (2000). "Time‐dependent ROC curves for censored survival data and a diagnostic marker." Biometrics 56(2): 337-344.
Hilden, J., J. D. F. Habbema and B. J. M. o. i. i. m. Bjerregaard (1978). "The measurement of performance in probabilistic diagnosis." Methods of Information in Medicine 17(04): 227-237.
Houwelingen, H. C. v. (2000). "Validation, calibration, revision and combination of prognostic survival models." Statistics in medicine 19(24): 3401-3415.
Hueng, D. Y., W. C. Tsai, H. Y. Chiou, S. W. Feng, C. Lin, Y. F. Li, L. C. Huang and M. H. Lin (2015). "DDX3X Biomarker Correlates with Poor Survival in Human Gliomas." Int J Mol Sci 16(7): 15578-15591.
Janssen, K. J., K. G. Moons, C. J. Kalkman, D. E. Grobbee and Y. Vergouwe (2008). "Updating methods improved the performance of a clinical prediction model in new patients." J Clin Epidemiol 61(1): 76-86.
Jiang, T., D. H. Nam, Z. Ram, W. S. Poon, J. Wang, D. Boldbaatar, Y. Mao, W. Ma, Q. Mao, Y. You, C. Jiang, X. Yang, C. Kang, X. Qiu, W. Li, S. Li, L. Chen, X. Li, Z. Liu, W. Wang, H. Bai, Y. Yao, S. Li, A. Wu, K. Sai, G. Li, K. Yao, X. Wei, X. Liu, Z. Zhang, Y. Dai, S. Lv, L. Wang, Z. Lin, J. Dong, G. Xu, X. Ma, W. Zhang, C. Zhang, B. Chen, G. You, Y. Wang, Y. Wang, Z. Bao, P. Yang, X. Fan, X. Liu, Z. Zhao, Z. Wang, Y. Li, Z. Wang, G. Li, S. Fang, L. Li, Y. Liu, S. Liu, X. Shan, Y. Liu, R. Chai, H. Hu, J. Chen, W. Yan, J. Cai, H. Wang, L. Chen, Y. Yang, Y. Wang, L. Han, Q. Wang, G. Chinese Glioma Cooperative, C. Society for Neuro-Oncology of, A. Chinese Brain Cancer, A. Chinese Glioma Genome and n. Asian Glioma Genome Atlas (2021). "Clinical practice guidelines for the management of adult diffuse gliomas." Cancer Lett 499: 60-72.
Johnson, W. E., C. Li and A. Rabinovic (2007). "Adjusting batch effects in microarray expression data using empirical Bayes methods." Biostatistics 8(1): 118-127.
Leek, J. T., R. B. Scharpf, H. C. Bravo, D. Simcha, B. Langmead, W. E. Johnson, D. Geman, K. Baggerly and R. A. Irizarry (2010). "Tackling the widespread and critical impact of batch effects in high-throughput data." Nat Rev Genet 11(10): 733-739.
Li, H. (2022, 2022). "M7G-LGG Prognostic Model." from https://lhj0520.shinyapps.io/M7G-LGG_model/.
Li, Y., L. Xia, K. Tan, X. Ye, Z. Zuo, M. Li, R. Xiao, Z. Wang, X. Liu, M. Deng, J. Cui, M. Yang, Q. Luo, S. Liu, X. Cao, H. Zhu, T. Liu, J. Hu, J. Shi, S. Xiao and L. Xia (2020). "N(6)-Methyladenosine co-transcriptionally directs the demethylation of histone H3K9me2." Nat Genet 52(9): 870-877.
Luo, X., H. Li, J. Liang, Q. Zhao, Y. Xie, J. Ren and Z. Zuo (2021). "RMVar: an updated database of functional variants involved in RNA modifications." Nucleic Acids Res 49(D1): D1405-D1412.
Luo, Y., Y. Yao, P. Wu, X. Zi, N. Sun and J. He (2022). "The potential role of N(7)-methylguanosine (m7G) in cancer." J Hematol Oncol 15(1): 63.
Mair, M. J., M. Geurts, M. J. van den Bent and A. S. Berghoff (2021). "A basic review on systemic treatment options in WHO grade II-III gliomas." Cancer Treat Rev 92: 102124.
Marshall, A., D. G. Altman, R. L. Holder and P. Royston (2009). "Combining estimates of interest in prognostic modelling studies after multiple imputation: current practice and guidelines." BMC Med Res Methodol 9: 57.
Mathoulin-Pelissier, S., S. Gourgou-Bourgade, F. Bonnetain and A. Kramar (2008). "Survival end point reporting in randomized cancer clinical trials: a review of major journals." J Clin Oncol 26(22): 3721-3726.
Mayakonda, A. and H. P. Koeffler (2016). "Maftools: Efficient analysis, visualization and summarization of MAF files from large-scale cohort based cancer studies." BioRxiv: 052662.
Moons, K. G., D. G. Altman, J. B. Reitsma, J. P. Ioannidis, P. Macaskill, E. W. Steyerberg, A. J. Vickers, D. F. Ransohoff and G. S. Collins (2015). "Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration." Ann Intern Med 162(1): W1-73.
Moons, K. G., R. A. Donders, T. Stijnen and F. E. Harrell Jr (2006). "Using the outcome for imputation of missing predictor values was preferred." Journal of clinical epidemiology 59(10): 1092-1101.
Moons, K. G., A. P. Kengne, D. E. Grobbee, P. Royston, Y. Vergouwe, D. G. Altman and M. Woodward (2012). "Risk prediction models: II. External validation, model updating, and impact assessment." Heart 98(9): 691-698.
Murad, H., R. Dankner, A. Berlin, L. Olmer and L. S. Freedman (2020). "Imputing missing time-dependent covariate values for the discrete time Cox model." Statistical methods in medical research 29(8): 2074-2086.
Newman, A. M., C. L. Liu, M. R. Green, A. J. Gentles, W. Feng, Y. Xu, C. D. Hoang, M. Diehn and A. A. Alizadeh (2015). "Robust enumeration of cell subsets from tissue expression profiles." Nat Methods 12(5): 453-457.
Ostrom, Q. T., N. Patil, G. Cioffi, K. Waite, C. Kruchko and J. S. Barnholtz-Sloan (2020). "CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2013–2017." Neuro-Oncology 22(Supplement_1): iv1-iv96.
Posternak, V., M. H. Ung, C. Cheng and M. D. Cole (2017). "MYC Mediates mRNA Cap Methylation of Canonical Wnt/beta-Catenin Signaling Transcripts By Recruiting CDK7 and RNA Methyltransferase." Mol Cancer Res 15(2): 213-224.
Ramaswami, G., P. Deng, R. Zhang, M. Anna Carbone, T. F. C. Mackay and J. Billy Li (2015). "Genetic mapping uncovers cis-regulatory landscape of RNA editing." Nat Commun 6: 8194.
Royston, P. and D. G. Altman (2013). "External validation of a Cox prognostic model: principles and methods." BMC medical research methodology 13(1): 1-15.
Royston, P. and D. G. Altman (2013). "External validation of a Cox prognostic model: principles and methods." Bmc Medical Research Methodology 13: 15.
Rubin, D. B. (2004). Multiple imputation for nonresponse in surveys, John Wiley & Sons.
Shao, D., Y. Li, J. Wu, B. Zhang, S. Xie, X. Zheng and Z. Jiang (2022). "An m6A/m5C/m1A/m7G-Related Long Non-coding RNA Signature to Predict Prognosis and Immune Features of Glioma." Front Genet 13: 903117.
Sreekanthreddy, P., H. Srinivasan, D. M. Kumar, M. B. Nijaguna, S. Sridevi, M. Vrinda, A. Arivazhagan, A. Balasubramaniam, A. S. Hegde, B. A. Chandramouli, V. Santosh, M. R. Rao, P. Kondaiah and K. Somasundaram (2010). "Identification of potential serum biomarkers of glioblastoma: serum osteopontin levels correlate with poor prognosis." Cancer Epidemiol Biomarkers Prev 19(6): 1409-1422.
Stevens, R. J. and K. K. Poppe (2020). "Validation of clinical prediction models: what does the "calibration slope" really measure?" J Clin Epidemiol 118: 93-99.
Steyerberg, E. W. (2009). Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating, Springer.
Steyerberg, E. W., G. J. Borsboom, H. C. van Houwelingen, M. J. Eijkemans and J. D. F. Habbema (2004). "Validation and updating of predictive logistic regression models: a study on sample size and shrinkage." Statistics in medicine 23(16): 2567-2586.
Steyerberg, E. W., F. E. Harrell Jr, G. J. Borsboom, M. Eijkemans, Y. Vergouwe and J. D. F. Habbema (2001). "Internal validation of predictive models: efficiency of some procedures for logistic regression analysis." Journal of clinical epidemiology 54(8): 774-781.
Steyerberg, E. W., A. J. Vickers, N. R. Cook, T. Gerds, M. Gonen, N. Obuchowski, M. J. Pencina and M. W. Kattan (2010). "Assessing the performance of prediction models: a framework for traditional and novel measures." Epidemiology 21(1): 128-138.
Tan, Y. Q., Y. T. Li, T. F. Yan, Y. Xu, B. H. Liu, J. A. Yang, X. Yang, Q. X. Chen and H. B. Zhang (2020). "Six Immune Associated Genes Construct Prognostic Model Evaluate Low-Grade Glioma." Front Immunol 11: 606164.
Tang, K., X. Ji, M. Zhou, Z. Deng, Y. Huang, G. Zheng and Z. Cao (2021). "Rank-in: enabling integrative analysis across microarray and RNA-seq for cancer." Nucleic Acids Res 49(17): e99.
Toll, D. B., K. J. Janssen, Y. Vergouwe and K. G. Moons (2008). "Validation, updating and impact of clinical prediction rules: a review." J Clin Epidemiol 61(11): 1085-1094.
Van Calster, B., D. J. McLernon, M. van Smeden, L. Wynants, E. W. Steyerberg, t. Topic Group 'Evaluating diagnostic and S. i. prediction models' of the (2019). "Calibration: the Achilles heel of predictive analytics." BMC Med 17(1): 230.
Vergouwe, Y., E. W. Steyerberg, M. J. Eijkemans and J. D. F. Habbema (2005). "Substantial effective sample sizes were required for external validation studies of predictive logistic regression models." Journal of clinical epidemiology 58(5): 475-483.
Wang, J. (2020). "Calibration slope versus discrimination slope: shoes on the wrong feet." J Clin Epidemiol 125: 161-162.
Wang, J., T. Yuan, H. Ding, J. Xu, W. R. Keusters, X. Ling, L. Fu, Q. Zhu, Q. Li, X. Tang, W. Cai, H. Shang, L. Li and H. Zou (2021). "Development and external validation of a prognostic model for survival of people living with HIV/AIDS initiating antiretroviral therapy." Lancet Reg Health West Pac 16: 100269.
Wang, Z., Z. Zhong, Z. Jiang, Z. Chen, Y. Chen and Y. Xu (2022). "A novel prognostic 7-methylguanosine signature reflects immune microenvironment and alternative splicing in glioma based on multi-omics analysis." Front Cell Dev Biol 10: 902394.
White, I. R. and P. Royston (2009). "Imputing missing covariate values for the Cox model." Statistics in medicine 28(15): 1982-1998.
Woo, H. H. and S. K. Chambers (2019). "Human ALKBH3-induced m(1)A demethylation increases the CSF-1 mRNA stability in breast and ovarian cancer cells." Biochim Biophys Acta Gene Regul Mech 1862(1): 35-46.
Wood, A. M., I. R. White and P. Royston (2008). "How should variable selection be performed with multiply imputed data?" Stat Med 27(17): 3227-3246.
Wu, X., C. Li, Z. Wang, Y. Zhang, S. Liu, S. Chen, S. Chen, W. Liu and X. Liu (2022). "A bioinformatic analysis study of m(7)G regulator-mediated methylation modification patterns and tumor microenvironment infiltration in glioblastoma." BMC Cancer 22(1): 729.
Yan, X., Q. Zhou, H. Zhu, W. Liu, H. Xu, W. Yin, M. Zhao, X. Jiang and C. Ren (2021). "The clinical features, prognostic significance, and immune heterogeneity of CD37 in diffuse gliomas." iScience 24(11).
Yoshihara, K., M. Shahmoradgoli, E. Martinez, R. Vegesna, H. Kim, W. Torres-Garcia, V. Trevino, H. Shen, P. W. Laird, D. A. Levine, S. L. Carter, G. Getz, K. Stemke-Hale, G. B. Mills and R. G. Verhaak (2013). "Inferring tumour purity and stromal and immune cell admixture from expression data." Nat Commun 4: 2612.
Yu, K., Y. Hu, F. Wu, Q. Guo, Z. Qian, W. Hu, J. Chen, K. Wang, X. Fan, X. Wu, J. E. Rasko, X. Fan, A. Iavarone, T. Jiang, F. Tang and X. D. Su (2020). "Surveying brain tumor heterogeneity by single-cell RNA-sequencing of multi-sector biopsies." Natl Sci Rev 7(8): 1306-1318.
Zhang, L. S., C. Liu, H. Ma, Q. Dai, H. L. Sun, G. Luo, Z. Zhang, L. Zhang, L. Hu, X. Dong and C. He (2019). "Transcriptome-wide Mapping of Internal N(7)-Methylguanosine Methylome in Mammalian mRNA." Mol Cell 74(6): 1304-1316 e1308.
Zhang, M., J. Song, W. Yuan, W. Zhang and Z. Sun (2021). "Roles of RNA Methylation on Tumor Immunity and Clinical Implications." Front Immunol 12: 641507.
Zhou, K., J. Yang, X. Li, W. Xiong, P. Zhang and X. Zhang (2022). "N7-Methylguanosine Regulatory Genes Profoundly Affect the Prognosis, Progression, and Antitumor Immune Response of Hepatocellular Carcinoma." Front Surg 9: 893977.

No competing interests reported.

supplementarymaterial.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Development, validation, and updating of prognostic models for m7G- associated genes in lower-grade gliomas

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

1 Introduction

2 Materials and Methods

2.1 Dataset collection and study population

2.2 Bioinformatics analysis

2.3 Statistical analysis

2.4 Model development, validation

2.5 Model updating

2.6 The Online Calculator

3 Results

3.1 Data Summary

3.2 Identification of TAM cell marker gene expression profiles

3.3 Identification of TAMs-associated m7G-related genes

3.4 Derivation, validation and revision of the gene prediction model (Model 1)

3.5 Derivation, validation and recalibration of Model 2

3.6 Somatic mutation and immune analyses

4 Discussion

5 Conclusions

Abbreviations

Declarations

Contribution to the field statement

Funding:

Author Contribution

Acknowledgments

References

Additional Declarations

Supplementary Files

Status:

Version 1