Advancing AI methods for intracellular localisation of transcripts within neurons

Neurons extend over long distances and must deliver genetic instructions (RNA molecules) to the right place at the right time. Instead of making all proteins in one central location, neurons transport RNAs to their branches and connections, where proteins can be produced locally. This process is essential for brain development, learning, and memory, and its disruption is linked to serious neurological conditions (e.g., Alzheimer’s disease). To understand how neurons function, and what fails in disease, we need to determine precisely where individual RNA molecules are located within neurons. 

Recent advances in spatial biology allow individual RNA molecules to be visualized directly within brain tissue, offering an unprecedented view of their organisation at subcellular resolution. However, these datasets are noisy and complex, making it difficult to assign RNAs to neuronal structures such as axons, dendrites, and synapses. This PhD project will address this challenge by developing novel AI approaches and tools. The student will use advanced deep learning methods, including Transformer-based models, to improve image quality and identify different neuronal compartments. By tackling image denoising and segmentation, plus integrating additional data modalities they will develop systems to enable accurate mapping of RNA molecules to specific parts of neurons. 

The project will deliver robust AI-powered image analysis tools for spatial biology and apply them to study how RNA regulation supports healthy brain function and how it is disrupted in disease. This interdisciplinary PhD combines training in AI with neuroscience, applying cutting-edge, AI-driven image analysis to a real-world scientific problem.

The overarching aim is to improve detectability and localization precision of true RNA transcripts (without inventing transcripts) and map these to different neuronal cell types and subcellular compartments within neurons. Specific objectives are: 

1. Develop strategies to denoise MERSCOPEspatial transcriptomic images of mouse brain to achieve single transcript localisation refinement. 

This will include leveraging self-supervised deep learning, physics-informed deconvolution, and decoding-aware modelling, to improve the signal quality and spatial precision, enabling improved three-dimensional localization and deblending. 

2. Use confocal microscopeimagesand as a complementary reference for multi-modal denoising of the spatial transcriptomic images. 

Serial confocal images will be used as a structural prior in multimodal AI models, including Transformer-based architectures trained for volumetric, cross-modal image alignment, to guide denoising of spatial transcriptomics data, improving signal fidelity and localization accuracy. 

3. Incorporateinto the AI model segmentation of different neural tissue cell types and subcellular compartments within neurons. 

Segmentation of neuronal cell types and subcellular compartments will be integrated with denoising through multi-task deep learning models with shared encoders, enabling biologically constrained noise removal that improves signal quality while preserving accurate cellular and compartmental RNA assignment. 

4. Apply the AI model developed toidentify RNA transcripts that have altered localisation in two paradigms of neuronal traffic dysfunction. 

Algorithms will be used to identify transcripts with altered distribution in neurons of mouse brains that lack the molecular chaperone protein sacsin and in a model where proteostasis is impaired through a tau dependent mechanism.