Structured Deep Generative Models for Exploring the Single Cell Landscape

Abstract

The rise of single-cell sequencing technologies has enabled characterization of cellular states at unprecedented scale and resolution. Despite the promise of single-cell profiling, interpretation of the data is not straightforward due to substantial technical artifacts from the sequencing process unrelated to any meaningful biological phenomena. For example, single-cell RNA sequencing measurements may be confounded by transcriptional noise, variable capture efficiency, and batch effects across experiments among other issues. Towards obtaining robust biological insights from single-cell data, recent works have proposed hierarchical Bayesian models that explicitly account for known technical source of variation in the data generation process. By doing so, these models may disentangle meaningful biological variations of interest from irrelevant nuisance factors. Under this paradigm, the choice of how to represent ``meaningful'' variations largely determines the efficacy of a given model. Yet, far from being static, this designation may vary wildly between different analyses and is intimately linked with the specific analysis being pursued. Thus, to draw meaningful insights from our data, we cannot simply reuse models with relatively loose assumptions (e.g. data points being independently and identically distributed), but must instead carefully design our model's structure in tandem with a given line of inquiry. Concretely, the work presented in this thesis revolves around the following claim: No single model is suitable for all lines of inquiry. Distinct scientific questions require distinct model structures to obtain meaningful insights from single-cell data. To validate this claim, this thesis presents a suite of novel generative models tailored for the investigation of specific classes of hypotheses in single-cell data science. Beyond just single-cell analyses, we have found that the core ideas behind these models may be of use in other machine learning domains more generally. The remainder of this thesis is organized as follows: In Part I we provide an overview of necessary biological and machine learning background and summarize the specific contributions of this thesis. We proceed in Part II to describe our proposed models and present accompanying experimental results demonstrating their efficacy. Part III concludes and discusses potential directions for future work.