Learning Early Cell Development Features in Spatial Biology Using Weak Supervision

Advances in spatial biology now allow detailed characterisation of tissues through high-resolution histology and spatially resolved molecular profiling. These technologies enable direct observation of cellular heterogeneity and differentiation within intact tissue environments. However, extracting meaningful biological insight from such data remains challenging, as most existing AI approaches rely on dense cell-level annotations that are costly, subjective, and often ill-defined for early stages of cell differentiation.

This project will develop advanced AI methods to discover and localise cellular and molecular features associated with early differentiation using weak supervision. Rather than requiring manual annotation of individual cells, the proposed approach will learn from coarse, tissue-level group labels that describe overall differentiation composition (e.g., relative abundance of mature cell types) but contain no information about individual cell identity or spatial location. The central methodological challenge is to infer which spatially localised patterns in multi-modal data explain these aggregate tissue-level properties. To address this, the project will investigate weakly supervised learning frameworks that can localise features from high-level labels, including multi-instance learning and reinforcement-learning-based optimisation, integrating histology with spatial transcriptomics and other omics data to enable objective feature discovery without dense annotations.

The methods will be evaluated across multiple cell types, including from the prostate, pancreas and thyroid, to assess robustness, transferability, and biological relevance. By reducing dependence on detailed annotation, the project aims to deliver scalable AI tools for analysing early cell differentiation and tissue organisation in complex spatial biology data.

Modern spatial biology technologies allow scientists to study tissues in unprecedented detail by combining high-resolution microscope images with information about gene activity and other molecular signals. These approaches make it possible to observe how different types of cells are organised within tissues and how cells progress from early stages of development to more mature forms. However, analysing these complex datasets remains challenging. Most current artificial intelligence (AI) methods require detailed manual annotation, such as marking individual cells and identifying their developmental stage. Producing these annotations is time-consuming, subjective, and often unreliable for early differentiating cells, which may not yet show clear defining features.

This project will develop new AI methods that can identify features of early cell development or change without relying on detailed cell-level labelling. Instead, the approach will use broader tissue-level information, such as the overall composition of mature cell types within a tissue region. From these coarse signals, the AI will learn which patterns in images and molecular data correspond to early stages of development or change and where they are located within the tissue.

By reducing the need for extensive manual annotation, this work will make analysis of spatial biology data more scalable and objective. In the longer term, it will provide new tools for understanding how cells develop and organise within healthy tissues and cancer.