Abstract
Molecular subgroup is now influencing risk stratification and disease management in medulloblastoma. Tissue metabolite profiles have shown promise in identifying the four consensus subgroups. A smaller number of metabolites can be measured non-invasively in patients using magnetic resonance spectroscopy (MRS). We investigated if a classifier constructed from tissue metabolites could be applied to in-vivo metabolite profiles to accurately predict subgroup non-invasively. Machine learning was used to construct a classifier with tissue concentrations from 10 metabolites reliably detected in-vivo. Retrospectively acquired diagnostic in-vivo MRS was available for 37 cases from four treatment centres. Although we identified WNT tumours by the presence of GABA with 100% accuracy in tissue, GABA was not reliability quantified in this in-vivo dataset. Therefore the classifier was developed using tissue profiles of known molecular subgroup (determined using DNA methylation array) from group 3(n=20), group 4(n=34) and SHH(n=23). The cross-validated accuracy of the tissue classifier was 86%. When applied to in-vivo metabolite profiles, subgroup was predicted non-invasively with an overall accuracy of 76%. Group 3 had the highest proportion of incorrectly classified cases (4/11), followed by SHH (2/10), and group 4 (3/16), largely due to differences in measuring lipids, glutamate, glutamine and hypotaurine in-vivo. We have established the feasibility of non-invasive metabolite profiling to identify medulloblastoma subgroups. With the ongoing optimization of MRS to target specific metabolites including GABA, we can further improve accuracy. Rapid, non-invasive preoperative diagnosis of subgroup will offer opportunities to stratify early therapeutic intervention especially surgery, avoiding long-term sequalae and improving quality of survival.