Immune Microenvironment Reconstruction and Interaction Modelling via Generative Graphical Vision Transformers
Code: BBSRC-DFA_2026_30
Primary Supervisor: Muhammad Salman Haleem
Email: m.haleem@qmul.ac.uk
Institute: School of Electronic Engineering and Computer Science
Secondary Supervisor: Stuart McDonald
Email: s.a.mcdonald@qmul.ac.uk
Institute: Barts Cancer Institute
Abstract:
This PhD proposal aims to develop novel computational methodologies for scalable, multi-scale representation of cells in three-dimensional tissue images, enabling robust analysis of spatial organisation and cell-to-cell interactions. Recent advances in spatial biology generate highly detailed maps of cellular position, morphology, and molecular state, but extracting reliable quantitative information from these data remains limited by current approaches to cell representation and segmentation. In particular, variability in cell size, shape, and density poses major challenges for accurately defining cell boundaries and neighbourhood relationships at scale.
The project will use established spatial imaging datasets, as well as publicly available multiplexed imaging data, as exemplar systems to develop and validate new methods. The core objective is to design scalable computational models that can represent cells at multiple levels of detail, ranging from simple spatial coordinates to detailed three-dimensional cellular geometries. To achieve this, the project will integrate generative diffusion models for multi-scale cellular reconstruction with graph-based vision transformer architectures for segmentation and neighbourhood feature extraction. These approaches will enable efficient handling of large tissue datasets while accounting for cellular heterogeneity and limited manual annotation.
In addition, the project will explore multimodal data fusion strategies to link cellular structure and spatial organisation with functional molecular information. The resulting framework will be broadly applicable across tissue types and experimental systems, supporting species-independent analysis of cellular organisation. Overall, this work will deliver enabling computational tools that advance the quantitative study of tissue architecture and collective cellular behaviour.
Lay Summary:
Understanding how cells behave requires studying how all cells are organised and interact within tissues, rather than focusing on individual cell types in isolation. Much like an ecosystem, tissue function emerges from the collective behaviour of many different cells arranged in complex spatial patterns.
Advances in spatial biology now allow researchers to label and locate every cell in a tissue section, generating detailed maps of cellular organisation. However, extracting meaningful biological insight from these images remains a major challenge. To measure how cells interact, computers must accurately identify the boundary of each individual cell (cell segmentation). This is difficult because cells vary widely in size, shape, and appearance, even within the same cell type, and tissue images can contain millions of cells. Small inaccuracies in cell boundaries can lead to large errors in measuring cell position, shape, and interactions.
This proposal aims to develop new computational methods that improve how cells are represented and reconstructed from complex biological images. The project will create scalable models that describe cells at multiple levels of detail, ranging from simple spatial points to detailed cell shapes, allowing analysis to be matched to biological question being asked. Advanced computer vision approaches will also be developed to segment cells accurately using limited manual annotation.
By improving how cells are segmented and quantified, this work will enable more reliable analysis of tissue organisation and cell-to-cell interactions. The methods will be tested using available spatial biology datasets and will be broadly applicable across many areas of biological research.
Aims and Objectives:
The overarching aim of this PhD project is to develop enabling computational frameworks for the quantitative analysis of tissue organisation through multiscale cellular representation.
The specific objectives are:
1) To design scalable generative diffusion models capable of reconstructing cellular representations at multiple spatial scales, ranging from point-based abstractions to detailed three-dimensional cell geometries.
2) To develop weakly supervised, multimodal vision transformer architectures for accurate extraction of composite cellular characteristics, explicitly accounting for spatial context and heterogeneity.
3) To apply graph-based modelling approaches to quantify cell–cell interactions and neighbourhood structure as functions of multiscale cellular features.
Together, these objectives address key methodological bottlenecks in spatial bioscience and support the development of generalisable analytical tools aligned with BBSRC priorities in quantitative, data-driven biology.