Bottom-up phenotyping of tissues across cell-types (Prof. Christoffer Nellaker)

PROJECT TITLE

Bottom-up phenotyping of tissues across cell-types.

SUPERVISORS

• Christoffer Nellaker

• Craig Glastonbury

DESCRIPTION OF PROJECT

We have developed a bottom-up histology analysis pipeline (HAPPY) which we trained in a supervised fashion for placenta tissue phenotyping. This is a powerful approach of analysing tissue structures using graph convolutional network analysis of cell communities. However, this is not sustainable to expand to other tissues due to the difficulty of manual annotation of cell types.

A project with the group would focus on adapting HAPPY for additional tissue contexts, enabiling compressed representations of large regions of tissues, and building additional tissue feature models.

For instance to make HAPPY work on other tissues without extensive manual labelling it would be better to build self- and weakly-supervised modelling approached to extract low dimensional representations of cell morphologies. There are existing models that perform this task (e.g. cellViT) which will be used as benchmarks. The project will focus on improved self-supervised learning models, and utilizing published spatial transcriptomic and histological datasets as a source of weak labels for training.

We will utilize cell representations to build self-supervised graph convolutional networks to capture tissue structures. Better cell representations should lead to better tissue structure discrimination and therefore be a means to measure the performance of models.

Finally there are aspects of tissue structures which could be quantified more directly through HAPPY models – such as vascularization, diffusion distance, and exchange surface predictions.

We would then look to apply these to reproductive tissue questions, which is a main focus of the group.

The broader ambition is to build a research tool set to enable tissue analysis and quantification across all tissue types.

TRAINING OPPORTUNITIES

• Research training in machine learning, statistics, quantitative genetics and beyond
• Membership to journal clubs and research group to understand current research landscape, cutting edge methodology and interdisciplinary applications
• Skills training in presentation, scientific writing, funding applications and scientific leadership
• Training plans will be developed holistically to benefit the scientific needs and interests, as well as the career goals of the individual.

Funding Information

Unfunded, however, the Deep Medicine group will assist the DPhil candidate in securing funding. 

HOW TO APPLY

To apply for this research degree, please click here.