APOLLO: Raman-based AI platform to accelerate the diagnosis and management of gliomas via rapid methylation profiling

Adrian Lita¹, Joel Sjöberg², David Păcioianu³, Nicoleta Siminea^3,4, Orieta Celiku¹, Andrei Păun^3,4, Mark R. Gilbert¹, Houtan Noushmehr⁵, Ion Petre^2,4, , and Mioara Larion^1,

¹National Cancer Institute, National Institutes of Health, Neuro-Oncology Branch, Bethesda, 20892, USA.
²University of Turku, Department of Mathematics and Statistics, Turku, 20500, Finland.
³University of Bucharest, Faculty of Mathematics and Computer Science, Bucharest, 010014, Romania.
⁴National Institute for Research and Development in Biological Sciences, Department of Bioinformatics, Bucharest,060031, Romania.
⁵Henry Ford Health System, Department of Neuro-Oncology, Detroit, 48202, USA.
^*mioara.larion@nih.gov, ion.petre@utu.fi

---

Methylation profiling of glioma has become an integral part of the diagnosis and treatment of this disease. However, determining the methylation subtype of glioma in a fast manner that would make it suitable for intraoperative decision-making remains a challenge. Herein we developed a novel workflow which we call APOLLO, based on spontaneous Raman spectroscopy and machine learning, to predict the methylation subtype of gliomas using FFPE slides. We demonstrate that our workflow discriminates tumors from non-tumor areas with a high accuracy of 0.98. We also show that APOLLO can discriminate IDH1mut versus IDH1WT tumors with accuracy 0.78. Moreover, we achieved high discriminatory power between G-CIMP-high and G-CIMP-low molecular phenotypes with accuracy 0.75, both of which are lower-grade IDH1mut gliomas but display vastly different clinical outcomes. Furthermore, we identified novel Raman shifts that are important for discriminating between different methylation subtypes and validated them using stimulated Raman spectroscopy and immunohistochemistry. We uncovered novel metabolic signatures of IDH1mut glioma that set them apart from IDH1WT due to their unique lipid biology. The development of APOLLO allows fast, reliable, and accurate prediction of methylation subtypes of glioma which can be implemented in the Operating Room.

Data

Link to the dataset: https://seafile.utu.fi/f/6b787f5046334c808fc8/?dl=1

Library versions

The models were trained and evauated using python 3.9.15, scikitlearn 1.1.3, pandas 1.4.4, numpy 1.21.5, matplotlib 3.5.3, seaborne 0.12.1.