Skip to main content
The William Harvey Research Institute - Faculty of Medicine and Dentistry

Breadcrumb

  1. WHRI
  2. Study
  3. Advanced AI for Multi-modal Spatial Biology - PhD Programme
  4. Projects and Partners
  5. BBSRC-DFA_2026_02

AI-Driven Multimodal Modelling of Spatial Immune Recognition and Evasion in Gastrointestinal Cancer

Code: BBSRC-DFA_2026_02

Primary Supervisor: Dr Vivek Singh
Email: vivek.singh@qmul.ac.uk 
Institute: Barts Cancer Institute

Secondary Supervisor: Prof Marco Gerlinger
Email: m.gerlinger@qmul.ac.uk 
Institute: Barts Cancer Institute

Abstract:

Effective immune recognition is critical for the control of early-stage tumours and for the success of immunotherapies. Yet many cancers evade immune surveillance through poorly understood spatial and molecular mechanisms. Recent advances in spatial transcriptomics now enable whole-transcriptome, single-cell gene expression profiling directly in clinical FFPE samples, generating unprecedented multimodal datasets that integrate molecular, spatial, and pathological information. 

This project will develop advanced AI methods to analyse uniquely deep spatial datasets comprising whole-transcriptome in situ RNA profiling of over 2.5 million cells across >110 biopsies from highly immunogenic or immunotherapy-treated gastrointestinal cancers, complemented by multiplex immunofluorescence, multi-region genetics, and expert pathological annotation. A core research output will be the development of novel spatially aware multimodal learning objectives and model architectures, rather than the application of existing spatial analysis pipelines. The central hypothesis is that multimodal AI models that explicitly encode spatial context and prior biological knowledge can identify the molecular and cellular mechanisms that govern CD8 T-cell tumour infiltration and functional states, which ultimately determine whether the cancer is controlled or progresses. 

The project will develop a multimodal spatial foundation model using graph-based deep learning to learn robust representations of cells, niches, and tumour regions. These representations will be used to predict spatial immune phenotypes and to identify ligand–receptor interactions, signalling pathways, and niche-specific programs associated with immune infiltration or exclusion. An agent-based AI framework, coupled to a biologically grounded language model, will enable reproducible, expert-guided interrogation of the data and integration of curated prior biological knowledge. 

Lay Summary:

The immune system can recognise and destroy cancer cells, but many tumours develop ways to avoid immune attack. Understanding how immune cells move into tumours, where they become blocked, and why they often stop working is essential for improving cancer immunotherapies. Studying these processes in detail has been difficult because most techniques either lack spatial information or measure only a small number of molecules. 

New spatial transcriptomics technologies allow measuring the activity of thousands of genes in each individual cell of a cancer sample while simultaneously recording their precise location in the tumour. We will apply these to gastrointestinal cancers, combined with protein measurements, genetic information, and expert pathology assessment, to study how immune cells interact with cancer cells in space. 

The development of a new AI method that integrates and interprets all of these data types at once is central to the project. By modelling how various cell types are arranged in tumours and how each communicates with its neighbours, the AI system will identify molecular signals and cell–cell interactions that promote or prevent immune cell infiltration. These insights will help explain why some tumours are controlled by the immune system while others continue to grow”. 

The project will furthermore create an interactive AI tool that allows researchers to ask biological questions of the data, such as which genes are active near immune cells in successful versus unsuccessful immunotherapy responses. Ultimately, this will improve our understanding of cancer immune evasion and support the development of better immunotherapies. 

Aims and Objectives:

  • Methods: Develop a multimodal spatial foundation model that integrates COSMx whole-transcriptome (WTx) spatial transcriptomics, multiplex IF, pathological annotations, and genetic context to learn hierarchical representations of individual cell states, local microenvironmental niches, and higher-order tumour regions, disentangling intrinsic cell programmes from spatial organisation to enable robust immune phenotype prediction and mechanistic analysis. 
  • Prediction: Predict spatial immune phenotypes (e.g., T-cell infiltration depth, inactivation states, and immunotherapy response) and benchmark against non-spatial and single-modality models under strict patient-level validation, including evaluation of zero-shot prediction using structured tumour–immune priors in settings with limited labelled data. 
  • Mechanisms: Identify novel pathways and cell–cell communication programs associated with T-cell infiltration versus exclusion and activation versus anergy or exhaustion using spatially conditioned interaction modelling and counterfactual sensitivity analysis. 
  • Translation/usability: Implement a reproducible, agent-based analysis framework with a biologically grounded query interface that links user questions to audited analyses and curated prior knowledge. 
Back to top