From single-cells to patient diagnosis: statistical machine learning approaches to biomedical discovery

Recent technological advances have enabled routine collection of large quantities of biomedical information, ranging from molecular data such as gene expression in single-cells to population level datasets such as critical care databases. In this talk I will outline three current projects that apply and adapt cutting edge statistical approaches to effectively leverage this information to provide biomedical insights. Firstly, I'll introduce methodology to integrate single-cell RNA and DNA sequencing data to assign gene expression states to mutational cancer clones, discussing how we implemented our model using stochastic gradient Variational Bayes in Tensorflow. Secondly, I will introduce cellassign, a probabilistic model that automates the annotation of single-cell RNA-sequencing data to known types, allowing effective quantification of the tumour microenvironment in human cancers. Finally, I will introduce a probabilistic framework for optimizing the order of diagnostic test acquisition in a critical care context, and demonstrate how this can be applied to black box prediction functions (such as deep neural nets) in the context of differential diagnosis.