Virtual spatial omics profiling from histopathology by foundation AI to reveal cross-species tissue and cellular dynamics in skin ageing

Code: BBSRC-DFA_2026_04

Primary Supervisor: Prof Jun Wang
Email: j.a.wang@qmul.ac.uk 
Institute: Barts Cancer Institute

Secondary Supervisor: Dr Ines Sequeira
Email: i.sequeira@qmul.ac.uk
Institute: Institute of Dentistry

Abstract:

Spatial omics techniques profile tissue sections at single cell resolution, enabling spatially resolved investigation of cellular heterogeneity and microenvironment. This has revolutionised our understanding of cross-cell and cross-tissue dynamics in physiology and health. However, such spatial techniques are often very expensive, requiring specialised high-end instruments and trained personnel, limiting their large-scale cross-species applications. Histopathology with H&E staining is widely used for tissue and cell examination, and is cheap and easily accessible. A large body of deep learning models have been developed to examine H&E images and identify morphological features associated with events such as genome doubling, mutations and gene expression, and clinical outcomes. However, all these models lack the high-resolution capability to identify spatially resolved features, such as cellular neighbourhood and interactions.

We hypothesise that using matched H&E and spatial omics profiling data, by implementing state-of-the-art foundation AI we can perform virtual spatial omics profiling and predict cellular phenotypes from H&E images. This streamlined approach can spatially profile hundreds of thousands of publicly available H&E samples at single-cell resolution without increasing costs. Next, we will develop a transformer-based foundation AI that learns multimodal spatial and cell representations for the prediction of spatial labels and compositions, maintaining cross-modality and cross-species representations. We will apply these foundation models to analyse our in-house and publicly available spatial omics data of human and mouse skin samples, along with thousands of virtual spatial omics samples, to systematically uncover cross-species cellular dynamics and gene regulatory networks that govern skin ageing. 

Lay Summary:

During skin ageing, the tissue and cells undergo re-organisation of structure and interactions to cope with internal and external reactions. To understand this complex process, it requires high-resolution cell level investigation across the whole tissue space. Recent advances in spatial biology technologies have made it feasible by scanning and quantifying proteins and genes for every cell within a selected tissue section. However, these technologies are very expensive, and rely on high-end instruments and trained personnel, limiting their translation to large numbers of end-users.  

Histology slides have been widely used for tissue and cell examination for over 100 years. This technique is standardised, cheap and easily accessible. A large body of AI models have been developed in the last few years to analyse histology images to predict changes in the genome and clinical outcomes. However, their applications in cell organisation and interactions are still limited. 

In this project, we will apply the latest foundation AI techniques to develop AI tools to scan histology images and perform computational measurement of proteins and genes for every cell within the images, using the information learnt from matched samples measured by spatial instruments. Our tools can then automatically identify cell types and how cells are organized and interacted with each other. We will use our AI tools to analyse thousands of histology images of human and mouse skin samples, and in-house skin data generated from spatial instruments. We aim to identify cross-species changes in cell organisation and interactions, and responsible genes associated with skin ageing.

Aims and Objectives:

Aim 1: Train a foundation AI model ‘CellPredict’ to identify cellular phenotypes from H&E images using matched histology and spatial omics data in both human and mouse. 

Aim 2: Develop a transformer-based foundation model for multimodal learning with cross-modality and cross-species representations and perform downstream spatial label and composition prediction. 

Aim 3: Uncover cross-species changes in cellular niches, dynamics and gene regulatory networks that govern skin ageing.