Generative 3D Reconstruction of Cellular Architecture from Multimodal Spatial Omics Data
Code: BBSRC-DFA_2026_20
Primary Supervisor: Dr. Shanxin Yuan
Email: shanxin.yuan@qmul.ac.uk
Institute: School of Electronic Engineering and Computer Science
Secondary Supervisor: Prof. Dr. Simon Haas
Email: s.haas@qmul.ac.uk
Institute: Precision Healthcare University Research Institute
Abstract:
Spatial transcriptomics integrates gene expression with high-resolution imaging, yet current analyses
remain fundamentally two-dimensional, limiting interpretation of inherently three-dimensional tissue
architecture. This project will develop a generalisable AI framework to reconstruct anatomically
interpretable three-dimensional cellular models directly from two-dimensional multimodal spatial data.
We hypothesise that single-cell spatial omics datasets contain sufficient structural information to learn
probabilistic priors of cellular anatomy, enabling volumetric inference from single tissue sections.
Building on advances in single-view 3D computer vision and generative modelling, cells will be
represented using 3D Gaussian Splatting and refined into compartment-resolved volumetric models
of membrane, cytoplasm and nucleus. Differentiable rendering and latent generative models will
enable uncertainty-aware reconstruction and consensus three-dimensional archetypes for major cell
types.
The inferred geometries will be integrated into spatial transcriptomic workflows to enable geometry aware transcript assignment and correction of segmentation-induced artefacts. Aggregated
volumetric representations will define canonical three-dimensional cellular models and quantify
structural divergence across differentiation and disease states.
Human bone marrow provides an ideal biological context for the development of such approaches.
As the primary site of haematopoiesis, the bone marrow harbours haematopoietic stem and
progenitor cells that continuously generate all blood and immune lineages through intermediate
cellular states characterised by distinct molecular programmes and progressive changes in cellular
and nuclear morphology. Whilst bone marrow will serve as an ideal test bed, the methodological
advances will be broadly applicable to any spatial omics dataset and tissue.
Lay Summary:
Cells in our body are not just defined by the genes they express, but also by their shape, internal structure, and position within tissues. Modern technologies can now measure gene activity directly within intact tissue sections, providing detailed molecular maps. However, these methods capture only flat, two-dimensional images, even though cells are three-dimensional structures. This makes it difficult to fully understand how gene activity relates to cell shape and organisation.
This project will develop new artificial intelligence (AI) methods that can reconstruct realistic three-dimensional models of cells from standard two-dimensional microscope images. By learning from thousands of cells, the system will build typical or “consensus” 3D models for different cell types, capturing their characteristic shapes and internal organisation. These models will allow us to link gene activity patterns directly to cellular structure in a quantitative way.
The approach will also improve how gene signals are assigned to individual cells in dense tissues, where overlapping cells often cause errors. Beyond technical improvements, the project will create a framework to compare 3D cell structures across development and disease.
We will develop and test this method using human bone marrow, a complex tissue where blood cells form and where many blood cancers arise. However, the tools will be designed to work across different tissues. Ultimately, this research will provide new ways to understand how changes in gene activity are reflected in cell structure in health and disease.
Aims and Objectives:
The overarching aim of this project is to develop a computational framework that integrates multimodal spatial transcriptomic and imaging data to reconstruct 3D cellular models from 2D images to link transcriptional identity to cellular morphology and spatial context. It will then be applied to address key biological questions in human bone marrow.
We will pursue three specific aims:
Aim 1 – Learn and Validate 3D Generative Cell Anatomy.
Develop a generative modelling framework to reconstruct three-dimensional cellular representations from two-dimensional multimodal spatial data. Cells will be represented as collections of 3D Gaussian “splat” ellipsoids for each anatomical compartment (membrane, cytoplasm, nucleus). Validation of reconstruction fidelity will be performed with synthetic and experimental data to ensure that inferred structures are coherent, robust and reproducible.
Objective 1: Demonstrate accurate and uncertainty-aware volumetric reconstructions of held-out 2D cell images, with quantitative validation.
Aim 2 – Integrate 3D Cellular Geometry with Spatial Transcriptomics
Incorporate inferred 3D cellular representations into a spatial transcriptomics analysis pipeline to enable geometry-aware transcript assignment. By embedding volumetric constraints into transcript-to-cell mapping, this aim seeks to improve molecular attribution and reduce segmentation-induced inaccuracies. The geometry-informed approach will be compared against standard segmentation tools.
Objective 2: Demonstrate measurable improvement in spatial transcriptomic analysis through three-dimensional geometry-aware modelling.
Aim 3 – Define Consensus 3D Cellular Models and Quantify Structural Divergence
Define consensus three-dimensional cellular models for all major bone marrow cell types by aggregating learned volumetric representations and linking them to transcriptional identity and spatial context. This framework will then be applied to detect and quantify structural divergence across stem cell differentiation trajectories and between healthy and malignant states.
Objective 3: Establish canonical three-dimensional cellular archetypes and quantify architectural variation associated with lineage commitment and disease.