Using AI-Based Evolutionary Algorithms to Elucidate Adult Brain Tumor (Glioma) Etiology Associated with IDH1 for Therapeutic Target Identification

Abstract

Adult brain tumors (glioma) represent a cancer of unmet need where standard-of-care is non-curative; thus, new therapies are urgently needed. It is unclear whether isocitrate dehydrogenases (IDH1/2) when not mutated have any role in gliomagenesis or tumor growth. Nevertheless, IDH1 is overexpressed in glioblastoma (GBM), which could impact upon cellular metabolism and epigenetic reprogramming. This study characterizes IDH1 expression and associated genes and pathways. A novel biomarker discovery pipeline using artificial intelligence (evolutionary algorithms) was employed to analyze IDH-wildtype adult gliomas from the TCGA LGG-GBM cohort. Ninety genes whose expression correlated with IDH1 expression were identified from: (1) All gliomas, (2) primary GBM, and (3) recurrent GBM tumors. Genes were overrepresented in ubiquitin-mediated proteolysis, focal adhesion, mTOR signaling, and pyruvate metabolism pathways. Other non-enriched pathways included O-glycan biosynthesis, notch signaling, and signaling regulating stem cell pluripotency (PCGF3). Potential prognostic (TSPYL2, JAKMIP1, CIT, TMTC1) and two diagnostic (MINK1, PLEKHM3) biomarkers were downregulated in GBM. Their gene expression and methylation were negatively and positively correlated with IDH1 expression, respectively. Two diagnostic biomarkers (BZW1, RCF2) showed the opposite trend. Prognostic genes were not impacted by high frequencies of molecular alterations and only one (TMTC1) could be validated in another cohort. Genes with mechanistic links to IDH1 were involved in brain neuronal development, cell proliferation, cytokinesis, and O-mannosylation as well as tumor suppression and anaplerosis. Results highlight metabolic vulnerabilities and therapeutic targets for use in future clinical trials.

Keywords: TCGA; artificial intelligence; biomarker; brain cancer; evolutionary algorithm; glioblastoma; glioma; isocitrate dehydrogenase 1.

Figure 1

All pairwise comparisons of mRNA expression between GBM and non-tumor samples and also the other glioma subtypes were significantly different (p < 0.001; t-tests) for each of the potential prognostic genes: testis-specific protein Y-encoded 2; (a,b) (TSPYL2), Janus kinase and microtubule-interacting protein 1; (c,d) (JAKMIP1), citron rho-interacting serine/threonine kinase; (e,f) (CIT), and transmembrane O-mannosyltransferase targeting cadherins 1; (g,h) (TMTC1).

Figure 1

All pairwise comparisons of mRNA expression between GBM and non-tumor samples and also the other glioma subtypes were significantly different (p < 0.001; t-tests) for each of the potential prognostic genes: testis-specific protein Y-encoded 2; (a,b) (TSPYL2), Janus kinase and microtubule-interacting protein 1; (c,d) (JAKMIP1), citron rho-interacting serine/threonine kinase; (e,f) (CIT), and transmembrane O-mannosyltransferase targeting cadherins 1; (g,h) (TMTC1).

Figure 2

Results of the survival analysis with risk tables and Kaplan–Meier curves comparing overall survival of patients with high vs. low mRNA expression (median split). Each of the genes (a) TSPYL2; (b) JAKMIP1; (c) CIT; and (d) TMTC1 were prognostic for GBM (IDH-wildtype; p < 0.001; Log-rank test). TMTC1 was also prognostic for recurrent GBM (IDH-wildtype; p < 0.05; Log-rank test; not shown).

Figure 3

Correlations of IDH1 gene expression with expression and methylation data for TSPYL2 (a) and TSPYL1 * (b); JAKMIP1 (c,d); CIT (e,f); and TMTC1 (g,h). The linear regression line is provided as well as the R-squared “Score”. Parentheses indicate that transformed data (natural logarithm (l), arcsine (a), or square root (s)) provided a higher correlation in ACE. * Methylation data were not available for TSPYL2 in TCGA-GBM dataset, so TSPYL1 is presented instead.

Figure 3

Correlations of IDH1 gene expression with expression and methylation data for TSPYL2 (a) and TSPYL1 * (b); JAKMIP1 (c,d); CIT (e,f); and TMTC1 (g,h). The linear regression line is provided as well as the R-squared “Score”. Parentheses indicate that transformed data (natural logarithm (l), arcsine (a), or square root (s)) provided a higher correlation in ACE. * Methylation data were not available for TSPYL2 in TCGA-GBM dataset, so TSPYL1 is presented instead.