Centres for Doctoral Training

Representing and interacting with virus-infected cell data via natural language

Supervisors: 

David L Robertson, MVLS, School of Infection and Immunity 

Craig MacDonald, CoSE, Computing Science 

Project Summary:

Recent advances in deep learning applied to modelling biological complexity are enabling new approaches to understanding cells and viral infection. Building on advances in protein and genome language models, transformer-based approaches can learn contextual representations of biological systems from large-scale data. Because viruses depend on host cells, models of infection must capture the infected cell state, including interactions with the host immune response. Conventional approaches, however, rely on graph-based biomolecular networks that are difficult to interrogate with natural language. Cell foundation models adapt transformer architectures to gene expression by representing genes as tokens and encoding their expression levels to generate rich embeddings of cellular states.

This project will investigate the integration of foundation models trained on single-cell transcriptomic data with large language models to represent, organise and query information about the biology of virus infection using natural language. This integration offers opportunities to apply methods from natural language processing, including information retrieval, data representation, question answering and recommender systems, to the analysis of infected cells. Key questions include whether cell foundation models can generalise to virus infection, predict cellular permissiveness to infection, prioritise host factors that could be targeted for antiviral intervention and identify shared host responses across different viruses.