AI for Multimodal Spatial Analysis of Vascular Tissue Organisation

Code: BBSRC-DFA_2026_07

Primary Supervisor: Paul Evans
Email: paul.evans@qmul.ac.uk 
Institute: William Harvey Research Institute 

Secondary Supervisor: Ginestra Bianconi
Email: g.bianconi@qmul.ac.uk 
Institute: School of Mathematical Sciences 

Abstract:

Understanding how mechanical and molecular signals jointly regulate tissue organisation is a central challenge in spatial biology. Advances in high-resolution vascular imaging, deep learning–based structural reconstruction, and computational modelling of haemodynamic forces now enable detailed spatial characterisation of vascular systems. However, existing analytical approaches struggle to integrate structured flow fields with omics-scale biological variability, limiting our ability to learn generalisable principles of mechanosensitivity and physiological adaptation from complex spatial datasets. 

This PhD will develop advanced AI methods to integrate multimodal spatial imaging with transcriptomic data, generating physics-aware and network-based models of vascular mechanosensitivity. Emphasising methodological innovation rather than application of established pipelines, the project will design interpretable spatial AI frameworks capable of learning from structured spatial systems while respecting the fundamental biology underpinning vascular mechanosensitivity. By representing arterial networks as interconnected biological environments influenced by local mechanical cues, the research will uncover transferable computational strategies for analysing spatial biological data. 

The project will leverage a uniquely rich multimodal dataset generated through a British Heart Foundation–funded longitudinal programme. These resources combine photon-counting CT–derived vascular geometry, deep learning–enabled structural mapping, spatially resolved haemodynamic fields, and spatial molecular profiles capturing inter-individual variation in mechanosensitive pathways. Complete datasets will be available prior to the student’s start date, ensuring strong feasibility and immediate scope for AI development. 

Few datasets globally integrate vascular topology, simulated physical forces, and spatial omics measurements at this resolution. This provides an exceptional foundation for creating next-generation AI tools capable of extracting structured insight from high-dimensional spatial systems. By shifting the analytical paradigm toward discovering spatial signatures directly from data, the project places algorithmic innovation at its core while delivering broadly applicable methods for multimodal spatial biology. 

Aims and Objectives:

The overarching goal is to create transferable AI methodologies for learning from high-dimensional spatial datasets that encode both biological structure and mechanosensitivity. This will be achieved through a unified spatiotemporal AI framework that integrates three tightly coupled components: physics-aware modelling, heterophily-aware graph learning and structured multimodal fusion. The first objective is the development of physics-informed spatial AI models that approximate computational haemodynamic simulations while preserving key fluid-dynamic constraints (conservation laws and invariants). Conventional simulations are computationally intensive and difficult to scale; surrogate AI architectures with embedded physical structure in the learning process have the potential to dramatically accelerate analysis while maintaining biological realism. 

The second objective is to learn structured spatial representations of vascular mechanosensitivity systems. Rather than treating measurements as independent features, the project will model arterial networks as organised spatial graphs, explicitly designed for heterophilous biological networks, where adjacent regions may exhibit contrasting transcriptomic states. Instead of simple local aggregation as done in conventional graph networks, the project will develop toplogy-aware and non-local message-passing strategies to capture directional and long-range dependencies across vascular networks. Such network-based and geometric learning approaches provide a natural framework for capturing these dependencies and align closely with emerging AI strategies for complex biological systems. 

A third objective focuses on multimodal data fusion. The student will design models capable of integrating imaging-derived spatial features with spatial transcriptomic profiles, addressing a central frontier in spatial AI, reconciling heterogeneous data modalities within unified predictive representations. We hypothesise that vascular remodelling under disturbed blood flow patterns will be enhanced in patients that exhibit higher expression of mechanosensitive genes, identified by spatial transcriptomics, compared to patients with lower expression. By learning how mechanical environments influence molecular responses across structured spatial graphs, the framework will test whether disturbed flow interacts with mechanosensitive gene expression to drive vascular remodelling. The resulting framework is expected to generalise beyond vascular biology, including other spatially structured biological systems where physical coupling links functionally distinct regions, thus supporting broader applications in spatial-omics research. 

Interpretability will be embedded at the model architectural level, rather than applied post hoc, through topology-aware and physics-consistent attribution mechanisms to ensure that inferred relationships remain biologically plausible and mechanistically informative. Collectively, these advances position the project firmly within AI algorithm development and software creation rather than application of existing analytical workflows.