Latent variable models for dissecting cell-to-cell heterogeneity in single-cell profiling data

The cell is often viewed as the fundamental unit in biology. The cells within the organs of our bodies and other organisms differ from each other in their molecular state and can be characterised in terms of the set of genes they express. Conventional techniques for measuring this gene expression are based on large populations of cells; rather than quantifying the gene expression of single cells, they rely on measuring the average gene expression of a very large number of cells. Many important research questions, however, can only be addressed when analysing cells individually. For example, single-cell analyses are vital for identifying rare cell types (e.g. tumour cells in blood samples), or for providing insights into disease development (e.g. better characterisation of the differentiation of blood stem cells is crucial for understanding leukaemia). Many recent technological advances now allow us to perform such single-cell profiling and tackle some of these open questions. However, as with any new technology, there are challenges that must be overcome before this powerful approach can attain its full potential.

In this proposal I am planning to address the fundamental challenges which arise in the statistical analysis of this new type of data. In particular, these include problems in the reliable identification of rare cell types and the presence of confounding factors. When measuring the expression of all genes in an ensemble of individual cells, the observed RNA output (which is a measure of the gene expression strength) between cells will always vary. Some of the variation may arise from technical artefacts (the experimental procedure may just be more efficient in one cell than another) while other variation might be due to a multitude of hidden biological processes that cannot be observed directly, e.g. effects of the cell cycle, which is a confounding factor that in turn can mask other processes of interest, such as differentiation. Consequently, for any analysis of single-cell data, it is vital to understand and uncover the reasons behind the observed variability in gene expression.
In the course of this fellowship, I will use a statistical approach called latent variable models to unravel the observed variability between single cells. This in turn will enable the reliable identification of rare cell types within an ensemble of single cells and could be used to formally test the statistical significance of individual biological processes.
A related challenge I will address is the difficulty of identifying the order in which gene expression patterns change during differentiation. As gene expression can only be measured once for each cell, data-sets from cell differentiation studies consist of multiple 'snapshots', each containing a mixed population of cells that are in different stages of differentiation. I will therefore extend latent variable models to reveal the hidden temporal order of the cells and identify cells at various stages within the differentiation process.
I will work together with leading experimental groups and apply the newly developed methods to study differentiation processes in blood stem cells as well as in immune cells. This real-world application of my methods has the potential ultimately to provide new insights into the biology of infection, autoimmune diseases and leukaemia.
Finally, I will provide freely available and user-friendly open-source software, facilitating the broad applicability of my proposed methods in any lab performing single-cell experiments.

Conventional techniques for profiling cells quantify the average RNA abundance in large populations of cells. Recent technological advances now enable the quantification of transcriptional abundance at the single-cell level, by assaying the complete transcriptome of individual cells (e.g. using single-cell RNA sequencing technology). Datasets derived using this technology can be used to identify distinct molecular states (sub-populations) within heterogeneous populations of cells and yield new insights in core cellular processes. The aim of this proposal is to develop the necessary statistical methods to address pertinent statistical challenges when analysing these data. In particular, we aim to model the cell-to-cell heterogeneity that is caused by the complex interplay of multiple hidden biological factors. Importantly, both technical noise and biological processes underlie this variability. These hidden (latent) processes may include both processes of interest (e.g. differentiation) and confounding processes (e.g. cell cycle). I propose to develop latent variable models to dissect the observed variability. To this end, I will exploit prior knowledge on the sources of heterogeneity to establish a dictionary of latent variables that is able to fully capture the observed variability in gene expression data. In turn, these inferred latent variables will enable the efficient identification of sub-populations and can be used to formally test for the significance of individual latent processes. Furthermore, I will use Gaussian Process models to infer the transcriptional dynamics of differentiating cells from snapshot data of largely unsynchronised cells. Finally, I will develop user-friendly open-source software in order to facilitate broad application of the proposed methods. In collaboration with leading experimental groups, I will apply the newly developed methods to study differentiation processes in blood stem/progenitor cells as well as in immune cells.

The proposed research on developing statistical methods for analysis of single-cell profiling data has the potential to impact researchers from a wide range of disciplines-mainly in the biomedical sciences, but also in the field of statistical learning. In the short term, my work will benefit the increasing number of researchers from different disciplines conducting single-cell studies, as the challenges I am planning to address are likely to be issues in a considerable fraction of such experiments. While I will mainly develop methods for analysis of single-cell RNA-seq data, these methods are also highly relevant in other single cell "omics" applications, e.g. single-cell metabolomics, which I will explore in work package 3.
In order to maximise the circle of potential users of the proposed methods, I am planning to develop a user-friendly, open-source software package that includes detailed documentation. This will enable many researchers from different backgrounds to use the proposed methods. I am confident that my proposed methods will help some of these researchers gain novel, biologically relevant insights from their data.
Part of the proposed research involves the development and application of efficient inference methods in structured covariance models. Therefore, the community of researchers working in statistical learning will also benefit from the proposed work, potentially facilitating the development of novel methods and applications built upon its foundations.

In the long term, the proposed research has the potential to have an impact on central public health issues. I am planning to collaborate with Dr Teichmann and Prof Göttgens, whose research is focused on transcriptional regulation of the mouse immune system and blood stem cell differentiation, respectively. Ultimately, within these collaborations, the proposed research may contribute to new insights into the biology of infection, autoimmune diseases and leukaemia.

In addition, the proposed research may have an impact in the biotechnological sector. The EMBL-EBI Industry Programme provides a forum for knowledge exchange with industrial partners, which could pave the way for development of new commercial products.

I will also make my research accessible to the general public via press releases and non-technical summaries of scientific articles. By communicating my research to the general public I am hoping to raise awareness of the importance of biostatistics in the age of ever-increasing data and perhaps ignite the interest of the next generation of researchers.

On a related note, I am hoping to continue to engage with current students and researchers, as well as professionals from other disciplines, through various teaching activities. In this way, I am keen to contribute to the training of future computational biologists and biostatisticians.

In summary, the proposed research has the potential to be of great benefit for the knowledge economy of the UK-including the advancement of scientific knowledge-and the long-term improvement of public health.