Decoding the Topological Grammar of Ectopic Lymphoid Structures using Multimodal Graph Neural Networks
Code: BBSRC-DFA_2026_12
Primary Supervisor: Prof Greg Slabaugh
Email: g.slabaugh@qmul.ac.uk
Institute: Digital Environment Research Institute
Secondary Supervisor: Elena Pontarini
Email: e.pontarini@qmul.ac.uk
Institute: William Harvey Research Institute
Abstract:
Ectopic lymphoid structures (ELS) are highly organised immune microenvironments that arise within non-lymphoid tissues during sustained immune activation. Unlike diffuse immune infiltration, ELS exhibit structured architecture, including segregated B-cell and T-cell zones, stromal networks, and hierarchical organisation. Although typically assessed using semi-quantitative histological scoring, there are currently no rigorous computational frameworks for modelling or quantifying their spatial organisation across scales.
This project will use model organ tissue known to be heterogeneous for the presence of ELS to develop an end-to-end artificial intelligence framework for analysing organised immune microenvironments using multiplex imaging and spatial transcriptomics data (including Xenium and Visium HD platforms).Efficient self-attention–based segmentation models, building on additive-attention UNet architectures, will be developed to accurately identify and characterise individual cells. Segmented cells will then be represented as multimodal spatial graphs incorporating spatial proximity, molecular features, and cell identity.
Persistent homology and related topological descriptors will be applied to quantify higher-order architectural features such as clustering, zonation, and hierarchical organisation. These descriptors will be integrated into topology-aware graph neural networks to learn continuous embeddings that capture graded variation in immune microenvironment organisation.
The resulting framework will provide a generalisable approach for modelling organised immune architectures across tissues and biological contexts. By integrating efficient segmentation, graph learning, and topological analysis within multimodal spatial datasets, this project advances AI methodology for spatial biology and directly aligns with the objectives of the Doctoral Focal Award.
Lay Summary:
In some situations, immune cells organise themselves into highly structured clusters within tissues. These clusters, known as ectopic lymphoid structures, resemble small lymph nodes and contain organised regions of different immune cell types. Their structure is important because it reflects how immune responses are coordinated locally within tissue. However, current assessment of these structures relies mainly on visual inspection under a microscope and simple scoring systems, which do not fully capture their complexity.
This PhD project will develop new artificial intelligence (AI) methods to analyse high-resolution tissue images and spatial molecular data. The first step will be to build AI models that accurately identify and map individual cells in complex tissue images. These cells will then be represented as networks, allowing the model to study how cells are arranged and interact in space.
Using mathematical tools that describe spatial structure, the project will measure patterns such as clustering, organisation into zones, and hierarchical layering. Rather than simply detecting whether a structure is present or absent, the approach will quantify different degrees of organisation across samples.
The outcome will be a general framework for analysing organised immune structures in tissue using advanced AI techniques. Although the project focuses on immune microenvironments as a case study, the methods developed will be broadly applicable to many forms of spatial imaging data in biology.
Aims and Objectives:
Aim 1: Develop efficient attention-based segmentation models for spatial immune imaging
This aim focuses on identifying the cellular building blocks of ELS in heterogeneous tissue spatial images. We will establish robust cell-level segmentation across multiplex immunofluorescence and spatial transcriptomic imaging platforms. Building on efficient additive self-attention UNet architectures [2] developed within the Slabaugh lab, models will be adapted to handle densely packed immune cells and heterogeneous morphology.
Aim 2: Construct multimodal spatial graph representations of immune microenvironments
Following segmentation, tissue will be represented as cell-level spatial graphs incorporating spatial proximity, cell identity, and molecular features. Graph neural network (GNN) models will be developed to learn embeddings that capture neighbourhood organisation and zonation patterns beyond simple clustering or density-based approaches. This will help understand how B-cells, T-cells, and stromal networks coordinate during the early stages of ELS formation.
Aim 3: Map the Biological Ontology of ELS through Topological Descriptors
This aim will introduce formal topological characterisation of immune architecture. Persistent homology and related topological data analysis techniques will be applied to spatial graphs to derive descriptors capturing clustering, segregation, and hierarchical organisation across multiple spatial scales. Quantitative metrics of architectural organisation will be defined to move beyond semi-quantitative grading.
Aim 4: Develop Cross-Modality Generative Models for "Reductional" Spatial Inference
We will develop AI models to map advanced spatial transcriptomic signals (ground truth) onto basic H&E slides. This reductional approach will test if the AI can predict the latent molecular architecture of an ELS from morphology alone, delivering a user-friendly software tool.