Developing a machine learning approach to analyse imaging mass cytometry data in bovine embryos

Poor fertility is a worldwide problem and probably the most economically important animal health issue. The UK dairy industry alone suffers annual losses of ~500M (Dairy Board–2019). Fertility problems in cattle are escalating. First service conception rates are currently less than 40% compared to over 50% in the 1980s. A direct effect of subfertility is that many more animals are required to generate replacements. This drives down welfare (fewer resources per animal) and increases emissions (~11-19% of green-house emissions are produced by farm animals - largely by cows). To mitigate these challenges, in-vitro embryo production (IVEP) is a critical tool. Through IVEP we also have a remarkable opportunity to accelerate adaptations to increasingly widespread environmental challenges (for example prioritising heat tolerant genetics), reducing stress on animals while improving overall farm resilience. However, despite it’s potential, IVEP pregnancy success rates do not outperform natural pregnancy, ~30% of embryos are simply inherently unable to develop to term for unknown reasons. This is exacerbated by reliance on subjective grading systems to select embryos for transfer. We fail to capture critical biomarkers of competency, presenting an urgent problem for sustainable farming. Our goal is to directly address this, producing a data-driven alternative -computational tools that enable better understanding of bovine embryonic development, and predict an embryo's potential to produce a healthy pregnancy. This minimises losses with multiple, positive, knock-on impacts. To achieve this, we propose to develop an analytical framework combining two cutting edge imagine technologies -Imaging Mass Cytometry (to report protein biochemistry in cells) and 3D-Correlative Light Electron Microscopy (to report the ultrastructure of cells), along with bespoke machine learning tools. This approach will integrate multidimensional data—protein composition, ultrastructure, spatial organisation, and temporal dynamics. Applying this framework to bovine embryos will allow us to model embryonic development and predict developmental potential with unprecedented accuracy. This solution perfectly aligns with embryo biopsy techniques, where cells can be removed with no deleterious effect to the embryo. Biopsied cells are highly likely to contain information able to predict competency before embryos need to be selected for transfer. We will capture biopsied cells by our imaging methods and relate findings to an embryo outcome that indicates competency versus non-competency. However, to prove feasibility, we must develop state-of-the-art approaches to analyse these complex imaging data sets. Data is multi-dimensional and multi-modal, and while individual analytical solutions exist for certain elements of the analysis pipeline, it is the integration of these data types that presents a substantial challenge. Therefore, our immediate aims are to develop proof-of-concept machine learning approaches capable of: Mapping changes in protein content and ultrastructure across development in healthy and unhealthy embryos. Using positive and negative embryo outcome related to prior protein composition to predict competency. In mouse, we have de-risked experiments necessary to provide the data feed to develop computational tools and we have substantial team experience developing algorithms for imaging data processing. The success of these proof-of-concept approaches will position us for longer follow-on grants, delivering on each algorithm, enabling their full application.