Research
Multi-omics analysis: spatial, single-cell and regulatory genomics
Bulk measurements average over the cell types and anatomical structures that drive disease. I develop and apply analysis pipelines for spatial transcriptomics, working with both 10x Visium and NanoString GeoMx Digital Spatial Profiler data.
Using 10x Visium, we profiled the developing joint to show that microRNA-140 is more functional in joint development than in disease, resolving zonal gene expression changes across the growth plate. I developed a GeoMx analysis pipeline used to identify cell type and region-specific transcriptional changes in the endometrium of women with recurrent implantation failure, and applied the same approach to tissue gene expression profiles and cell communication networks in psoriasis and atopic dermatitis, informing candidate blood biomarker identification. These collaborations illustrate how methods developed for skeletal tissue transfer to other tissues and diseases.
My work also spans the assay types needed to link regulation to expression. This includes miRNA-seq, used in a comparative transcriptomic analysis of articular cartilage across post-traumatic osteoarthritis models, and ATAC-seq, used to characterise chromatin accessibility in mesenchymal stem cell chondrogenesis, adult hip cartilage chondrocytes and osteoarthritis. The computational problems these platforms raise, including cell type deconvolution, annotation across species, and integration with existing bulk data, remain an active interest.
Methods, software and data resources for interpreting omics data
A list of differentially abundant genes or proteins can be challenging to interpret. PhenomeExpress is a network-based algorithm that propagates expression evidence across biological networks and uses known disease and animal model phenotypes to identify the processes disturbed by an experimental perturbation. It is distributed as the PhenomeScape Cytoscape app, available from the Cytoscape App Store.
Public omics data can only be reused at scale if it is processed consistently and annotated properly. I build reproducible pipelines for processing transcriptomics data at scale, and the SkeletalVis R package allows search and visualisation of the resulting collection of uniformly processed skeletal transcriptomic datasets. It includes methods to identify datasets in which genes of interest are differentially expressed, and to find datasets with a gene expression pattern similar to a query dataset.
Alongside processed data, curated knowledge is required. OATargets integrates much of the molecular knowledge of osteoarthritis, collating evidence from human genetics, in vivo models and transcriptomics, including the genes experimentally modulated in animal models and their reported effect on joint damage.
Machine learning, knowledge graphs and large language models
Harmonised data supports prediction and recently we have applied machine learning to prioritise genes associated with osteoarthritis joint damage in animals, integrating curated experimental evidence with transcriptomic and network features to nominate candidates for functional validation.
Two current projects extend this work. The first applies large language models to metadata curation, since inconsistent and unstructured sample annotation in public repositories remains the principal barrier to large-scale data reuse, and to assisting the biological interpretation of omics results. The second builds biomedical knowledge graphs representing genes, phenotypes, pathways and experimental evidence in a structured form, providing a substrate for machine learning approaches such as link prediction and node classification for target discovery.