Machine Learning for Glioma Tumour Segmentation Based on the Methylation Subtypes

Illustration of a bird flying.
Project PNRR/2022/C9/MCID/I8, 1.07.2023 – 31.08.2026
Project code: 842027778

CF 68, PROJECT 760096

What We Do

Central nervous system tumors represent some of the most aggressive types of cancer. Gliomas are believed to develop from mutated progenitor or neuroglial stem cells and are notoriously difficult to treat. Their location within the brain, their often-quick growing nature, and their tendency to diffuse into healthy tissue with unclear delineations limits their accessibility and treatment via surgical resection. Treatment via chemotherapy is additionally challenging given the limitations of the blood-brain barrier and the difficulty generating targets specific enough to not damage surrounding brain tissue too greatly. Even with the newest lines of treatment, gliomas are incurable diseases. Despite our advancements in understanding the complexity of gliomas and our efforts to discover novel treatments, the survival of patients has not changed significantly. One limitation in the ability to treat these tumors is the limited knowledge of the glioma microenvironment and tumor biology. 

The long-term goal of this project is to provide a better understanding of the tumor microenvironment in different subtypes of gliomas as they progress and develop resistance over time to improve the disease outcome. 

The objective of this project is to develop a machine learning-based toolkit that enables the characterization of tumor microenvironment at a level unknown before. 

Our central hypothesis is that the use of Raman spectroscopy which creates a molecular fingerprint of the tumor at single-cell resolution, in conjunction with machine learning methods, can differentiate different types of cells existent in the tumor microenvironment and provide biological insights that guide further exploration into tumor biology. This will provide an understanding of tumor biology at a level unreachable before, by measuring single cells within their microenvironment and can lead to the discovery of novel targets and biomarkers with the potential to improve the disease outcome.

Who We Are

We are a group of computational scientists, modellers, and statisticians looking to apply artificial intelligence methods to provide a better understanding of the tumor microenvironment in different subtypes of gliomas as they progress and develop resistance over time to improve the disease outcome. Our methods are in machine learning, data science, and bioinformatics.

  • Florin Bilbie
  • Sandra Eremia
  • Daniela Florea
  • Eduard Milea
  • Victor Mitrana
  • Marian Necula
  • Bogdan Oancea
  • Ion Petre
  • Andrei Paun
  • Ana-Maria Seciu-Grama
  • Nicoleta Siminea
  • Laura Stefan
  • Iris Tusa
  • Alice Stoica
  • Eduard Szemeteanca
  • Ovidiu Vrinceanu
Objectives
Objective 1. Develop machine learning methods to classify the subtype of glioma using Raman spectroscopy of routine pathology slides.

We will develop machine learning methods for the classification of gliomas based upon the Raman fingerprint region. The focus will be on characterizing the tumor heterogeneity in terms of the methylation subtypes of the cells in various parts of the tumor, taking advantage of the single cell resolution of Raman spectroscopy. One major limitation in studying tumor heterogeneity is in the lack of methods capable of measuring a cell without removing it from the microenvironment. Raman spectroscopy has the potential to be the first technology to describe molecular changes in a single cell, thereby providing opportunities to study the dynamics of cellular transformation and the interconnections among different cells in the tumor by characterizing these cells in their microenvironment. The machine learning approach will be validated, and its prediction capability will be tested using tissue samples from glioma patients.

Objective 2. Develop machine learning methods to understand the brain tumor heterogeneity as a function of treatment and disease progression.

Tumor evolution over time represents one of the major limitations in the success of any therapeutic drug, thus leading to resistance. The Raman spectrum contains information regarding total proteins, lipids, and nucleic acids in single cells, which are commonly altered in tumor cells because of cell growth and proliferation. These advantages make Raman technology suitable for interrogating the dynamics of glioma cellular heterogeneity in tissues. Our collaborators have shown that the addition of drugs to cells changes the distribution of Raman signals drastically. However, in the highly concentrated biomolecular mixtures typically found in the tissues, quantitative analysis of such biomolecular classes is an extremely complicated task due to the overlap of their characteristic peaks in the Raman spectra. These challenges are addressed by our machine learning strategies.

List of publications
Software
Host Institute

National Institute of Research and Development in Biological Sciences, Romania.

Address: Splaiul Independenței 296, 060031 București, Romania

Website: https://www.incdsb.ro/

Phone: +40 21 220 77 80

Contact information

Prof. Dr. Andrei Păun, andrei.paun@incdsb.ro

Program page: https://mfe.gov.ro/pnrr/

Facebook page: https://www.facebook.com/PNRROficial

“PNRR. Funded by the European Union – NextGenerationEU” The content of this material does not necessarily represent the official position of the European Union or the Government of Romania.

RO: „PNRR. Finanțat de Uniunea Europeană – UrmătoareaGenerațieUE” Conținutul acestui material nu reprezintă în mod obligatoriu poziția oficială a Uniunii Europene sau a Guvernului României.