Neuro-oncology Programme Pump Prime Projects 2020-2021

The Neuro-oncology Programme are currently funding two pump prime projects: 

Cellular heterogeneity and stromal relationship differences in paediatric glioma subtypes Dr Manav Pathania, Milner Institute
Predicting peritumoural infiltration using an expectation-maximization regularized deep learning approach based on multiparametric MRI in glioblastoma Dr Chao Li, Dept of Clinical Neurosciences


Cellular heterogeneity and stromal relationship differences in paediatric glioma subtypes

Dr Manav Pathania, Department of Oncology, Milner Institute | £12,000 | March 2020 – March 2021

Collaborators: Prof Greg Hannon, CRUK Cambridge Institute

Background to the project

The majority of paediatric high grade gliomas (pHGGs) carry lysine-to-methionine (K27M) or glycine-to-arginine/valine (G34R/V) mutations in histone variants H3.1 or H3.3. These hotspot mutations in H3 variants are the defining genetic anomaly distinguishing paediatric from adult HGG.

Current therapies treat pHGGs similarly to adult brain tumours, even though these children’s tumours have completely different genetic drivers.

Project goals

The aim of this project is to use single cell RNA sequencing and imaging mass cytometry to identify pHGG subtype-specific reprogramming of the tumour microenvironment, by carrying out these analyses in models of two distinct pHGG subtypes.

The specific aims are:

  1. Develop models of pHGG driven by two different combinations of mutations and carry out comparative scRNA-seq using acutely dissociated tumour cells.
  2. Optimise an antibody panel for IMC and perform comparative, spatially resolved cellular heterogeneity analysis of pHGG subtype models.

What does this mean for patients?

Current therapies treat pHGGs similarly to adult brain tumours, even though these children’s tumours have completely different genetic drivers. These treatments are broad-acting and invasive, involving chemoradiotherapy and neurosurgery, and do not substantially improve survival. In fact, often there are unintended effects on the developing brain and adverse effects on quality of life.

The cross-species validation described in this project increases the probability of identifying the most therapeutically tractable mechanisms of stromal co-option present in these tumours.

Background to brain tumours in children

Brain tumours cause the most cancer-related death in children. Paediatric high-grade gliomas (pHGGS) are the most lethal and aggressive type of children’s brain tumour, affecting approximately 70-100 children in the UK every year.

H3 mutant pHGGs most frequently occur in infratentorial locations such as the brainstem and midline (where they are referred to as Diffuse Intrinsic Pontine Gliomas or Diffuse Midline Gliomas; DIPGs or DMGs). Currently there are no effective treatments for this universally fatal paediatric disease.

 

Predicting peritumoural infiltration using an expectation-maximization regularized deep learning approach based on multiparametric MRI in glioblastoma

Dr Chao Li, Department of Clinical Neurosciences | £12,000 | March 2020 – March 2021

Collaborators: Mr Stephen Price, Cambridge Brain Tumour Imaging Laboratory, Professor Carola-Bibiane Schönlieb, Centre for Mathematical Imaging in Healthcare

Background to the project

A radical surgical resection of glioblastoma is shown to benefit patient survival but may casue neurological deficits.  Therefore, there is a crucial need to define tumour invasive margins more accurately using pre-treatment imaging in order to maximise the benefits of surgery.

Multiparametric MRI may facilitate a better tumour characterisation by reflecting the multifaceted tumour physiology. However, effective models to integrate multiparametric MRI to inform clinical decision remains a challenge, where a supervised deep learning paradigm may not be applicable, due to the difficulty in accurately delineating tumour infiltration at the pixel level and the large amounts of training data needed.

Project goals

This project will integrate statistical learning (specifically the expectation maximization (EM) approach) into a deep learning (DL) model. The hypothesis is that by combining the strength of DL and statistical inference, a pixel-wise prediction could be achieved.

To train the model, a cohort of 129 glioblastomas has been recruited in a previous clinical trial. To validate the model, imaging data from 90 patients in an ongoing clinical trial (PRaM-GBM) will be included.

What does this mean for patients?

Improving the delineation of tumour invasion areas could have significant clinical impact on more targeted surgical treatment for glioblastoma patients.

Background to glioblastoma

Glioblastoma is the most common primary malignant tumour in adults, characterised by poor outcomes. Currently, surgery remains the mainstay of patient care and although contrast enhancing T1 weighted imaging is considered the ‘gold standard’ in treatment planning, glioblastoma infiltration beyond the enhancement margin is well known to occur.