Multimodal 3D Tissue Modelling (M3TM) for Metabolic Status

Insulin is a key hormone regulation blood glucose levels and its primary target is skeletal muscle. We have studied how insulin impacts on >500 humans, with muscle biopsy samples from people with different levels of physical activity and metabolic fitness. We have generated a unique set of  Merscope 960-gene profiles generating 100 regions along with standard pathology “H&E” serial section images, and high-resolution images of proteins within each samples. We have for each person measures of the blood insulin levels, aerobic fitness and fasting glucose. Our aim is to use these data to build novel models that help us understand how muscle tissue contributes to healthy insulin action. This PhD project will exploit these uniquely large and deeply phenotyped human muscle dataset to develop advanced AI-based models for spatial and molecular tissue analysis. Using advanced AI methods we will focus on studying the blood vessels in muscle, and how they vary with insulin action. Long term our work may help identify the pathways that control muscle metabolism and help screen for potential drugs to treat diabetes. Further, the nature of AI research components may have broad impact on image-based data modelling beyond single cell spatial biology. 

Evaluate existing 3D segmentation methods​11,13,17​ and develop data integration methods that enable stitching of both Z-stack (up to 7) and 10um serial sections to enable distance calculations between cell types and cell molecular sub-phenotypes to be constructed and related to clinical status. This will involve learning-based image registration and stitching methods to align Z-stack volumes and serial tissue sections, coupled with 3D reconstruction models designed to handle missing or sparsely sampled slices​15​ . The reconstructed volumes can be represented using graph-based spatial models​14​, enabling robust calculation of distances and spatial relationships between cell types and molecular sub-phenotypes, and facilitating their association with clinical status. 

Use graph neural network (GNN) frameworks to construct hierarchical spatial graphs that capture cell–cell interactions and relate them to variation in insulin sensitivity and glucose tolerance. Spatio-temporal representation learning​14​ will be employed to derive embeddings that encode differences in tissue organization between the high and low insulin sensitivity groups. Contrastive learning strategies​18​ will be exploited to rank representations between conditions, as well as between loaded and unloaded conditions, enabling systematic evaluation of the added value of 3D spatial modelling over conventional 2D approaches. 

Finally, use paired Spatial images/gene counts, H&E images and bulk transcriptomics data to build a model that attempts to infer spatial molecular profiles from only H&E data. First, multi-instance learning (MIL) methods​19​ will be applied to learn whole-slide and patch-level embeddings that link H&E tissue morphology to underlying gene expression. Then, generative models, such as conditional variational autoencoders and diffusion-based approaches​20​, will be developed to synthesise spatial gene expression maps directly from H&E images, trained using the paired H&E, spatial transcriptomic, and paired bulk transcriptomic data. To improve robustness and generalisability, self-supervised pretraining and transfer learning strategies will be explored​21​  - applied to n=150 H&E samples, with accompanying bulk transcriptomics, and measures of insulin sensitivity.