Statistical learning approaches for obtaining interpretable reduced representations of multimodal sequencing datasets

Leistico, Jacob R

Permalink

https://hdl.handle.net/2142/115458 Copy

Description

Title

Statistical learning approaches for obtaining interpretable reduced representations of multimodal sequencing datasets

Author(s)

Leistico, Jacob R

Issue Date

2022-04-19

Director of Research (if dissertation) or Advisor (if thesis)

Song, Jun S

Doctoral Committee Chair(s)

Gruebele, Martin

Committee Member(s)

• Kahn, Yonatan
• Kuehn, Seppe

Department of Study

Physics

Discipline

Physics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Epigenetic
• multimodal
• tensor decomposition
• single-cell

Abstract

The highly specialized functions of cells in a multicellular organism are effected by the proteins contained within and produced by the cell. Establishing the distinct protein composition of specialized cell types is accomplished largely through the regulation of gene transcription. Several epigenetic mechanisms, including chemical modifications of histone proteins, collectively function to establish and maintain the transcriptional state necessary for the specialized functions of the cell. The collection of RNA transcript abundance, protein content, and epigenetic features create a multimodal description of cell identity. Advances in high throughput sequencing technologies are increasingly enabling the profiling of multiple cellular modalities in bulk tissues and single cells. Computational algorithms that can integrate information from the multiple profiled modalities are needed to fully utilize the resulting information in classifying tissue or cell types and identifying the salient molecular features. This thesis presents rigorous computational methods for integrating multimodal genomic data to extract biologically relevant features that can help identify tissue or cell types. These methods share the common approach of obtaining reduced representations of the multimodal sequencing datasets that can be used for downstream analysis. The first method applies a higher order singular value decomposition (HOSVD) to decompose an epigenetic data tensor obtained from profiling multiple histone modifications in human tissue samples. The reduced representations obtained with this method are shown to capture features differentiating disease conditions, mutational subtypes, and tissue types. Additionally these representations are connected to covarying profiles in genomic location space that are shown to identify recurrent epigenetic differences between disease conditions. The second method applies the spectral clustering on multilayer graphs (SCML) algorithm to cluster cells from a multimodal single-cell sequencing dataset that simultaneously profiles mRNA transcripts and selected surface proteins. Clusters obtained with the SCML algorithm are shown to compromise between the community structure obtained from the mRNA transcripts and the protein surface markers. The results presented in this thesis demonstrate the benefits of integrating multiple molecular modalities in identifying cell and tissue types.

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Jacob Leistico

Owning Collections