That metabolomics work became a Python package using multivariate machine learning to model pathway features together. At Nucleome Therapeutics I moved to deep learning, fine-tuning genomic foundation models on chromatin accessibility, with the data engineering on AWS S3 and SageMaker.
My DPhil works a layer further up, in the epigenome, and uses all of it: machine learning, deep learning and multi-omics applied to gene regulation. The industry work was applied; this is the research side of the same question. Scaled across roughly 20,000 genes in the CD4+ T cell genome, an assay that resolves regulatory contacts to the base pair would give these models much better training data.
Grouped by stage below. Personal and laboratory projects are on the Projects page, and venture capital work on the Venture Capital page.
University of Oxford. Doctoral research in the lab of Prof. James Davies, academic co-founder of Nucleome Therapeutics, co-supervised by Prof. Christopher Yau, who directs the Ellison Institute of Technology CDT in Fundamentals of AI. Scaling base-pair-resolution 3D genomics (Micro-Capture-C) across the CD4+ T cell genome with machine learning, to read gene regulation out of chromatin structure.
Replicate five external genomics data sources inside the in-house target-information portal, exactly as presented on their own web pages, in a format that loads and plots efficiently.
Fine-tune genomic foundation models, including Enformer and Evo2, on ATAC-seq and DNase-seq data to try to improve on the in-house DeepHaem architecture.
Establish allelic skewing as a functional validation method for chromatin-accessibility predictions, and use it to compare model versions and direct training protocols.
Nucleome Therapeutics. Joined as a Computational Biologist building the data pipelines behind an in-house target-information portal, then moved into the ML team as a Junior ML Engineer, fine-tuning genomic foundation models on chromatin accessibility data and establishing allelic skewing as a functional validation method. Please click the title for more detailed information on results and outcomes.
Build computer vision prototypes for intra-operative AI, working on video from live surgical procedures.
Pre-train and test models supporting the company's remote and tele-mentored surgery ambitions.
Contribute to data infrastructure improvements and surgical AI validation protocols.
Touch Surgery, acquired by Medtronic for $300M five months later. Pre-training and testing computer vision models for intra-operative video, in support of the company’s remote and tele-mentored surgery work.
Use LLM-based semantic searching to map more metabolite names to IDs, widening the scope of pathway analysis (Figure 1A)
Apply novel pathway databases augmented with metabolites using deep learning, facilitating more comprehensive pathway mapping and attaining more robust biological predictions (Figure 1B)
Develop an extension to PathIntegrate which will provide an unsupervised utility for pathway-based multivariate analysis that can be benchmarked with synthetic data simulations (Figure 1C)
Figure 1: General Workflow of the project. Please click the title for more detailed information on results and outcomes.
Meta-analyse various single-cell studies of neuroblastoma, building a single-cell atlas that spans tumour cell diversity (Figure 2.1)
Identify cell populations and tissue heterogeneity in single cell spatial transcriptomics (SCST) data using transfer learning from the single-cell reference.
Validate, test and scrutinise available tools for exploratory analysis of novel SCST data, including fine tuning and adpatation of a novel, multi-modal deep learning clustering approach (SiGra) (Figure 2.2)
Figure 2.1: single-cell RNA-seq workflow. Figure 2.2: single-cell spatial RNA-seq workflow. Please click the title for more detailed information on results and outcomes.
Process 4 single cell/single nuclear datasets from the Allen Brain Institute in R
Generate high throughput splicing data using specialised tools in Unix (e.g Whippet).
Identify evidence to support the mechanism and presence of NMD containing transcripts in the nuclei of PV interneurons
Map different isoform profiles of GABAergic and glutamatergic layers
Validate the utility of Single-Cell data for splicing analysis compared to Bulk RNA-seq
Figure 4: Poster displaying the outcome of this project investigating alternative splicing during Neurogenesis. Please click the title for more detailed information on results and outcomes.
