Date of Award
Doctor of Philosophy (PhD)
Biomedical data is being generated in extremely high throughput and high dimension by technologies in areas ranging from single-cell genomics, proteomics, and transcriptomics (cytometry, single-cell RNA and ATAC sequencing) to neuroscience and cognition (fMRI and PET) to pharmaceuticals (drug perturbations and interactions). These new and emerging technologies and the datasets they create give an unprecedented view into the workings of their respective biological entities. However, there is a large gap between the information contained in these datasets and the insights that current machine learning methods can extract from them. This is especially the case when multiple technologies can measure the same underlying biological entity or system. By separately analyzing the same system but from different views gathered by different data modalities, patterns are left unobserved if they only emerge from the multi-dimensional joint representation of all of the modalities together. Through an interdisciplinary approach that emphasizes active collaboration with data domain experts, my research has developed models for data integration, extracting important insights through the joint analysis of varied data sources. In this thesis, I discuss models that address this task of multi-modal data integration, especially generative adversarial networks (GANs) and autoencoders (AEs). My research has been focused on using both of these models in a generative way for concrete problems in cutting-edge scientific applications rather than the exclusive focus on the generation of high-resolution natural images. The research in this thesis is united around ideas of building models that can extract new knowledge from scientific data inaccessible to currently existing methods.
Amodio, Matthew, "Deep Learning for Embedding and Integrating Multimodal Biomedical Data" (2022). Yale Graduate School of Arts and Sciences Dissertations. 550.