Dissertation
Computational modeling of proteins implicated in hearing loss
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2023
DOI: 10.25820/etd.007180
Abstract
Hearing loss is associated with over 128,500 missense variants spanning 219 genes, and approximately 6,300 of those missense variants have been classified as disease causing. However, seventy percent of these variants are classified as “variants of uncertain significance” (VUS) due to the inability to ascertain phenotypic effect from genotype. Experimental assessment (i.e., protein assays, model organisms, etc.) of over 90,000 missense variants is intractable, leaving variant characterization as the primary bottleneck to providing a genetic diagnosis from an identified missense variant.
Computational methods for variant classification can be used to highlight variants that merit costly experimental evaluation. However, computational tools for identifying deleterious variants often require the existence of an accurate three-dimensional protein model, a priori. To address the need for three-dimensional protein models, we developed Likelihood Recurrent Geometric Network (Likelihood-RGN), a machine learning algorithm for ab initio prediction of protein models. Likelihood-RGN builds on the RGN algorithm by enhancing the least squares loss function with a maximum likelihood loss function that includes experimental uncertainty data in the form of atomic displacement parameters. We then used machine learning (namely, the emergent AlphaFold2 algorithm) to develop protein models for the entire deafness proteome.
Although machine learning-based protein models provide coordinates where experimental data are not available, such models often contain discrepancies from biophysical expectations such as steric clashes and high-energy side-chain conformations. To make these protein predictions consistent with biophysical expectations, we accelerated the many-body side-chain optimization algorithm with novel approximations and used the algorithm to repack all machine learning models for the deafness proteome. Finally, we used our models to predict folding free energy differences (∆∆G), compute solvent accessible surface areas, identify protein model confidence, and collect CADD scores and minor allele frequencies for all 128,500 missense variants associated with deafness. We exploited these features to identify ∆∆G and CADD cutoffs that are consistent with deleteriousness, and we identify 3,456 destabilizing VUSs that are pathogenic at a probability of 99.0%. The work presented in this dissertation establishes a protein modeling protocol that identifies VUSs that are most likely disease-causing and uses this protocol to prioritize deafness-implicated VUSs.
Details
- Title: Subtitle
- Computational modeling of proteins implicated in hearing loss
- Creators
- Mallory RaNae Tollefson
- Contributors
- Michael J. Schnieders (Advisor)Richard J. H. Smith (Advisor)Terry A. Braun (Committee Member)Thomas L. Casavant (Committee Member)Michael A. Mackey (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007180
- Number of pages
- xxii, 135 pages
- Copyright
- Copyright 2023 Mallory RaNae Tollefson
- Comment
This thesis has been optimized for improved web viewing. If you require the original version, contact the University Archives at the University of Iowa: https://www.lib.uiowa.edu/sc/contact/.
- Language
- English
- Date submitted
- 03/17/2023
- Date approved
- 04/23/2023
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 102-113).
- Public Abstract (ETD)
The genetic causes of hearing loss vary greatly among patients. Hearing loss can be caused by any of approximately 6,300 different mutations in over 200 genes, and many mutations that cause hearing loss continue to be discovered by genetic sequencing. To date, over 90,000 protein-changing mutations are present in deafness-associated genes with too little experimental information to determine if those variants are benign or pathogenic. Such mutations are termed “variants of uncertain significance” (VUSs). VUSs are the primary bottleneck to providing a diagnosis for patients based on their genetic sequencing.
Though designing and performing experiments for over 90,000 VUSs is currently infeasible, computational methods can be used to identify the VUSs that are most likely to cause disease. In this dissertation, we develop a computational protein modeling protocol that results in identification of the VUSs most likely to cause disease. This protocol is comprised of three aims: 1) improve an existing machine learning algorithm and use it to predict protein models for all genes implicated in deafness, 2) implement a protein optimization algorithm to improve inaccuracies in the machine learning protein predictions, and 3) analyze VUSs using these protein models and establish a method for prioritizing VUSs that are most likely to cause deafness. Using this protein modeling protocol, we ultimately identify 3,399 VUSs that are pathogenic at a probability of 99.1%.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984424792202771
Metrics
3 File views/ downloads
20 Record Views