Probabilistic Algorithms for Integrated Analysis of Single-Cell Multi-Omic Data

Demetci, Pinar

Full Metadata

Overview

Year:: 2023
Contributor:: Demetci, Pinar (creator); Singh, Ritambhara (Advisor); Istrail, Sorin (Advisor); Crawford, Lorin (Reader); Larschan, Erica (Reader); Brown University. Center for Computational Molecular Biology (sponsor)
Genre:: theses
Subject:: Machine Learning; Genomics; Computational biology; multi-omic data; Multi-omics data; Data integration (Computer science); optimal transport; manifold alignment; single-cell sequencing
Extent:: , None p.

Files

Description

Abstract:: Advances in sequencing technologies in the last decade have enabled us to profile various genomic features at the single-cell resolution, such as gene expression and chemical modifications on the DNA. Studying how these features co-vary across cells can reveal how they interact to regulate cellular processes across cell types and states. However, with some exceptions, it is not possible to simultaneously take multiple types of genomic measurements on the same cells due to the destructive nature of sequencing technologies. Multi-modal studies of single-cell genomes thus require computational methods that integrate data from different sequencing experiments. This dissertation presents three probabilistic algorithms designed to address certain real-world challenges that existing algorithms fail to address when integrating various single-cell measurements. The first one, Single-Cell alignment with Optimal Transport (SCOT), is an unsupervised algorithm that compares dataset geometries to yield probabilistic cell alignments between two datasets. When there is validation data available for hyperparameter tuning, SCOT gives results on par with the state-of-the-art alignment algorithms. Unlike these algorithms, however, SCOT heuristically self-tunes its hyperparameters and still yields high-quality alignments when users do not have sufficient validation data. This is a realistic scenario as different features are profiled in different cells in single-cell experiments. The second algorithm, SCOT version 2 (SCOTv2), extends SCOT to align more than two datasets at a time. It also handles datasets with disproportionate cell-type representation, which we show is a common phenomenon in real-world experiments that most alignment algorithms fail to account for. The third algorithm, Single-cell fused Gromov CO-Optimal TRansport (SCOOTR), uses a novel optimal transport formulation to jointly align both cells and features through an alternating optimization scheme. This joint formulation not only improves cell alignments but also generates hypotheses about the relationships between genomic features. SCOOTR additionally allows for users to provide weak supervision on either the feature or the cell alignments in order to improve the quality of both.
Notes:: Thesis (Ph. D.)--Brown University, 2023

Content

Citation

Demetci, Pinar, "Probabilistic Algorithms for Integrated Analysis of Single-Cell Multi-Omic Data" (2023). Center for Computational Molecular Biology Theses and Dissertations. Brown Digital Repository. Brown University Library. https://repository.library.brown.edu/studio/item/bdr:er5vz2ca/

Relations

Collection:

Center for Computational Molecular Biology Theses and Dissertations

Theses and Dissertations for the Center for Computational Molecular Biology department.
...