DPhil in Medical Science, University of Oxford
Doctoral research at the MRC WIMM Centre for Computational Biology, University of Oxford, in the lab of Prof. James Davies, academic co-founder of Nucleome Therapeutics.
Supervision and training
My co-supervisor is Prof. Christopher Yau, Professor of Artificial Intelligence at the Big Data Institute, who directs the Ellison Institute of Technology Centre for Doctoral Training in Fundamentals of AI and is a Turing AI Fellow.
Alongside the research I completed the AI modules of that CDT, whose taught component covers the foundations rather than applications: computational statistics including Bayesian statistics, machine learning and learning theory, and optimisation. The programme also runs on rotation projects with different supervisors before a DPhil topic is settled.
The pairing is the point. The genomics side supplies an assay that resolves regulatory contacts to the base pair; the AI side supplies the methods to learn from data at that resolution and to be honest about the uncertainty in what comes out.
The assay the company was built on
The lab’s Micro-Capture-C (MCC) assay maps the 3D genome at single base pair resolution, published in Nature in 2021 and invented in the founders’ Oxford laboratories. Nucleome holds an exclusive licence to it from the University, and the method is the technical foundation the company was spun out around.
Resolution is the whole point. Most chromosome conformation methods resolve interactions to kilobase-scale bins, which is enough to say that a region contacts a promoter but not enough to say which base does. At base pair resolution, an enhancer-promoter interaction can be tied to the specific variant sitting inside it, which is what turns a non-coding GWAS association into a testable target.
The project
I am developing AI architectures to scale that data from the handful of loci an assay can interrogate directly towards roughly 20,000 genes across the CD4+ T cell genome, learning gene regulation from sequence and chromatin accessibility.
This is the research side of the question rather than a production pipeline. My industry work was applied: models that had to run, and predictions a drug discovery team would act on. Here the scaling is the open problem, and generating training data at that scale is as much a data engineering question as a modelling one.
The DPhil follows directly from doing this work in industry first: I spent a year at Nucleome on exactly this problem, training and validating chromatin-accessibility models, before returning to the lab the technology came from.
Teaching and additional modules
Alongside the research I teach and take additional modules. Write-ups will appear under teaching on the Projects page.
Where this sits
This is epigenetics: not which genes are present, but which are reachable, and which regulatory element is talking to which promoter in three-dimensional space. It is the layer upstream of the transcriptome, deciding what gets transcribed before any of the downstream omics have anything to measure.
My first research project was on splicing and nonsense-mediated decay, which is gene regulation read off the RNA after the fact. This asks a similar question one layer earlier, of the chromatin that determines what is transcribed at all.