Machine Learning Approaches to Multi-Modal Data Integration and Translation in Single-Cell Biology

Abstract

Building a complete picture of cell state requires measuring different properties of the cells, such as their gene expression, morphology, etc., and understanding 1) how these properties relate to each other, 2) how they change over time, 3) how they are affected by different perturbations. It is often difficult to collect this information through experimentation alone. High-throughput single-cell assays such as single-cell RNA-sequencing are destructive to cells, making it difficult to make other observations of the same cells at other time points or using different measurement tools.

In this thesis, I develop new machine learning methodology to integrate and translate between single-cell data. In the first half, I develop methods based on generative modeling, representation learning and optimal transport to learn mappings between cells collected at different time points. In the second half, I develop methods based on generative modeling and representation learning to map between different data modalities, including both observational measurements and interventions. Overall, this body of work progresses towards the larger goal of complete cell models that predict cell state under different measurements, time points, and perturbations.