Linking scales through AI-derived mechanistic laws from spatial multimodal data

Code: BBSRC-DFA_2026_05

Primary Supervisor: Adrian Biddle
Email: a.biddle@qmul.ac.uk 
Institute: Blizard Institute

Secondary Supervisor: Sathiskumar Anusuya Ponnusami
Email: s.a.ponnusami@qmul.ac.uk 
Institute: School of Engineering and Materials Science

Abstract:

Hematoxylin and eosin (H&E) histopathology provides a ubiquitous, high-resolution view of tissue architecture, yet its mechanistic relationship to underlying molecular interactions remains poorly defined. Spatial transcriptomic and multiplexed proteomic technologies enable direct measurement of cell states and signalling within intact microenvironments, but their limited scale constrains population-level inference and clinical translation. We hypothesize that spatial molecular data contain latent causal structures that generate observable histomorphological patterns, and that these mechanistic relationships can be learned and transferred to large H&E-only cohorts. 

We propose a multimodal framework integrating causal inference and representation learning to derive interpretable mechanistic laws linking single-cell molecular interactions to tissue morphology in oral cancer. Using a discovery cohort with matched H&E, spatial transcriptomics, and CellDIVE proteomics, we will learn unified embeddings capturing cellular states, spatial dependencies, and interaction networks, and apply causal discovery to infer directional relationships among cell types and pathways. These mechanistic constructs will be mapped to aligned H&E image features through multimodal deep learning and interpretable histomorphometric analysis to identify morphological correlates of molecular processes. Validated models will then be deployed on large H&E cohorts to generate spatially resolved predictions of mechanistic interactions and evaluate associations with clinical outcomes. 

This study will establish a scalable strategy for translating information-rich spatial molecular datasets into broadly applicable histopathological predictors of biological mechanisms. By bridging molecular and tissue scales, the work advances mechanistic understanding of tumour microenvironment organization and enables population-level mapping of molecular interactions directly from routine histology. 

Lay Summary:

Routine tissue slides stained with hematoxylin and eosin (H&E) dye are the most widely used way to examine human biology under the microscope. These images show how cells are shaped, arranged, and interact within tissues, providing a rich visual record of biological organisation. Although these patterns arise from underlying molecular activity, such as gene expression and cell-to-cell signalling, we still do not fully understand how microscopic tissue structure relates to the processes occurring inside and between cells. New technologies can measure molecular activity directly within intact tissue, but they are costly and therefore available only for small numbers of samples, limiting their broader use. 

This project aims to uncover general biological principles linking molecular interactions to tissue architecture, and to use these principles to interpret routine microscopy images at scale. We will analyse a set of human tissue samples for which both detailed molecular measurements and matched H&E images are available. Advanced artificial intelligence methods will identify recurring molecular interaction patterns and determine how they correspond to specific microscopic tissue features. These relationships will be distilled into clear, interpretable biological rules connecting molecular behaviour to visible tissue structure. 

We will then apply these learned rules to large collections of standard H&E tissue images from many additional samples, enabling us to infer underlying molecular processes without specialised molecular tests. Ultimately, this work could allow routine microscopy images to reveal deeper information about how human tissues are organised and function, advancing fundamental understanding of tissue biology and supporting future biomedical research. 

Aims and Objectives:

The objective of the project is to develop a causal-mechanistic, scale-linking AI framework that: 
(i) Learns interpretable spatial biological mechanisms from multimodal transcriptomic and proteomic data; 
(ii) Translates these mechanisms into robust, scalable predictors operating on routine H&E histology images; 
(iii) Enables population-scale inference of molecular interaction states from large histology datasets.