Machine Learning for Better Prognostic Stratification and Driver Gene Identification Using Somatic Copy Number Variations in Anaplastic Oligodendroglioma

Abstract

Background: 1p/19q-codeleted anaplastic gliomas have variable clinical behavior. We have recently shown that the common 9p21.3 allelic loss is an independent prognostic factor in this tumor type. The aim of this study is to identify less frequent genomic copy number variations (CNVs) with clinical importance that may shed light on molecular oncogenesis of this tumor type.

Materials and methods: A cohort of 197 patients with anaplastic oligodendroglioma was collected as part of the French POLA network. Clinical, pathological, and molecular information was recorded. CNV analysis was performed using single-nucleotide polymorphism arrays. Computational biology and feature selection based on the random forests method were used to identify CNV events associated with overall survival and other clinical-pathological variables.

Results: Recurrent chromosomal events were identified in chromosomes 4, 9, and 11. Forty-six focal amplification events and 22 focal deletion events were identified. Twenty-four focal CNV areas were associated with survival, and five of them were significantly associated with survival after multivariable analysis. Nine out of 24 CNV events were validated using an external cohort of The Cancer Genome Atlas. Five of the validated events contain a cancer-related gene or microRNA: CDKN2A deletion, SS18L1 amplification, RHOA/MIR191 copy-neutral loss of heterozygosity, FGFR3 amplification, and ARNT amplification. The CNV profile contributes to better survival prediction compared with clinical-based risk assessment.

Conclusion: Several recurrent CNV events, detected in anaplastic oligodendroglioma, enable better survival prediction. More importantly, they help in identifying potential genes for understanding oncogenesis and for personalized therapy.

Implications for practice: Genomic analysis of 197 anaplastic oligodendroglioma tumors reveals recurrent somatic copy number variation areas that may help in understanding oncogenesis and target identification for precision medicine. A machine learning multivariable model built using this genomic information enables better survival prediction.

Keywords: Genomics; Glioma; Machine learning; Oligodendroglioma; Survival.

Figure 1.

Copy number variation landscape. Landscape of recurrent deletions (A) and amplifications (B) identified by GISTIC.

Figure 2.

Chromosomal view of the copy number variation landscape.

Figure 3.

Multivariable random forest analysis. (A): Classification error as function of number of trees. (B): Relative importance measures of the variables associated with overall survival: VIMP (the reduction of prediction power by omitting a variable from the analysis) and average depth of the variables in the decision trees (smaller numbers represent variables closer to the tree root and thus greater influence on the classification). (C): Survival plots for nine of the variables with VIMP >0.005. Copy number variation coding is as follows. For amplifications: 0, normal; 1, gain; 2, high amplification. For deletions: 0, normal; 1, heterozygous deletion; 2, homozygous deletion. For copy‐neutral LOH: 0, normal; 1, LOH. Abbreviations: AMP, amplification; DEL, deletion; LOH, loss of heterozygosity; VIMP, variable importance.

Figure 4.

Nuclear atypia. (A): A case showing severe nuclear atypia with increased nuclear‐cytoplasmic ratio and enlarged and pleomorphic nuclei. (B): A case showing slightly irregular nuclei. (Hematoxylin and eosin, ×40.)