Interpretable Attention Map-Based Subcellular Localisation of Transcripts Using Compartment-Specific Protein Signatures
Code: BBSRC-DFA_2026_26
Primary Supervisor: Ziquan Liu
Email: ziquan.liu@qmul.ac.uk
Institute: School of Electronic Engineering and Computer Science
Secondary Supervisor: Mathieu Bourdenx
Email: m.bourdenx@ucl.ac.uk
Institute: UK Dementia Research Institute
Abstract:
Subcellular localisation of RNA transcripts is a fundamental mechanism of post-transcriptional gene regulation, enabling neurons to respond locally to activity-dependent stimuli in dendrites, axons, and synapses. Despite its importance for neuronal health, our understanding of which transcripts localise to which compartments remains incomplete, and robust computational tools to predict or assign transcript localisation from spatial transcriptomics data are lacking.
This project will develop a transformer-based deep learning framework that builds interpretable attention maps from compartment-specific protein markers and transcripts with established subcellular localisation, to perform genome-wide subcellular localisation prediction of transcripts profiled by spatial transcriptomics. Compartment-specific protein markers (e.g., MAP2 for dendrites, neurofilament for axons, PSD95 for post-synaptic densities, LAMP2 for lysosomes, RAB7 for late endosomes) will be used as spatial anchors, co-registered with single-molecule spatial transcriptomics data. A curated set of transcripts with validated subcellular localisation (from the literature and public databases) will serve as a supervised training signal. The model will learn cross-modal attention maps that capture spatial co-distributions between transcript signals and protein marker patterns, generating interpretable and reliable compartment-affinity scores for every transcript in the dataset.
The trained model will be applied to the full transcriptome profiled by MERSCOPE in mouse brain, producing subcellular localisation maps at scale. Validation will leverage held-out known-localisation transcripts, enabling rigorous benchmarking. The framework will then be applied to models of neurodegeneration to identify transcripts with disease-associated changes in localisation. This project will deliver open-source tools applicable to any spatial transcriptomics dataset with matched immunofluorescence.
Lay Summary:
Neurons are highly specialised cells that must deliver specific genetic instructions (RNA molecules) to precise locations within the cell — axons, dendrites, and synapses — far from the nucleus where these instructions are made. This spatial delivery of RNA molecules allows neurons to rapidly produce the proteins they need locally, supporting brain development, learning, and memory. When this targeting system breaks down, as it does in diseases such as Alzheimer's, the consequences for neuronal health can be severe.
Recent technologies allow scientists to measure the exact location of hundreds of different RNA molecules within intact brain tissue. However, assigning each RNA molecule to the correct part of the neuron — is it in the cell body, a dendrite, or a synapse? — remains a major challenge. This PhD project will tackle this problem using a cutting-edge artificial intelligence approach called attention mapping. The idea is to teach an AI system to recognise compartment-specific protein ‘landmarks’ within neurons (proteins that are exclusively found in dendrites, or exclusively in axons, for example) and to use these landmarks as signposts to determine where each RNA molecule is located. The AI will learn from RNA molecules whose compartment location is already known from the scientific literature and then apply this knowledge to predict the location of thousands of previously uncharacterised transcripts.
The project will deliver practical, open-source AI tools for the spatial biology community and will provide new insights into how RNA localisation is disrupted in neurodegenerative diseases.
Aims and Objectives:
The overarching aim is to develop and validate an attention map-based AI framework for predicting and interpreting subcellular RNA localisation in neurons at transcriptome scale. Specific objectives are:
- Curate a multimodal training dataset by co-registering compartment-specific immunofluorescence protein marker images with MERSCOPE spatial transcriptomics data from mouseand human brain tissue and compiling a reference set of transcripts with experimentally validated subcellular localisation.
- Develop a cross-modal transformer architecture with attention mechanisms that learns spatial associations between protein marker compartment maps and transcript distribution patterns, trained on known-localisation transcripts.
- Generate interpretable and robustattention maps that highlight the spatial features driving compartment assignment, providing mechanistic insights into transcript localisation rules.
- Apply the trained model to perform transcriptome-wide subcellular localisation prediction in aged and disease mouse brain models,identifying transcripts with altered compartment affinity in neurodegeneration.