Screening of glioma susceptibility SNPs and construction of risk models based on machine learning algorithms.

BACKGROUND

Glioma is a common primary malignant brain tumor. This study aimed to develop a predictive model for glioma risk by these screened key SNPs in the Chinese Han population.

METHODS

These 614 participants were randomly assigned to two datasets: a training dataset (217 cases and 213 controls) and a validation dataset (93 cases and 91 controls). Genotyping for 59 SNPs in 35 genes was conducted using Agena MassARRAY platform. Key SNPs associated with glioma susceptibility were identified through LASSO, SVM-RFE algorithm, and likelihood ratio. A nomogram was constructed to predict glioma risk, and its predictive accuracy was evaluated using calibration and ROC curves.

RESULTS

Twelve overlapping SNPs were identified by LASSO (λ = 0.022, 23 SNPs) and SVM-RFE algorithm (Accuracy = 0.6845, 15 SNPs). Moreover, likelihood ratios displayed 9 SNPs associated with glioma risk (p < 0.05). Combining these methods, ultimately 15 SNPs from 59 SNPs were identified as hub SNPs. Nomogram and ROC curves displayed that the model had good prediction performance in the training cohort (AUC = 0.7950) and the validation cohort (AUC = 0.7433), with rs3950296 and rs1317082 emerging as important risk factors.

CONCLUSIONS

We screened fifteen SNPs, especially rs3950296 and rs1317082 that were associated with glioma risk using machine learning and likelihood ratio tests. The predictive nomogram demonstrated good discrimination ability and potential utility for glioma risk prediction.