Improving Molecular Diagnosis Success Rates And Prognosis Prediction For Patients With Rare Diseases

Authors

• Tarun Karthik Kumar Mamidi, University of Alabama at Birmingham

Advisor(s)

Elizabeth Worthey

Committee Member(s)

Alecia Gross
Brittany Lasseigne
Ryan Melvin
Zsuzsa Bebok

School

Joint Health Sciences (Interdisciplinary)

Document Type

Dissertation

Department (new version)

Genetics

Date of Award

9-9-2024

Abstract

Rare diseases are medical conditions that affect a small number of people compared to the general population. There are more than 10,000 estimated rare diseases with 80% being genetic in origin, which although individually have a relatively low prevalence, collectively impact the lives of hundreds of millions of people worldwide. The diagnosis process is usually difficult and slow, but an accurate and early diagnosis can be crucial for disease treatment and management. This dissertation focuses on some of the important problems in this area by investigating the application of advanced machine learning methods to enhance the accuracy, precision, and efficiency of pathogenicity prediction for variants identified in patients with rare genetic disorders. The first part of the dissertation focuses on predicting the pathogenicity of any type of small variant. It leverages existing pathogenic and benign classifications from databases, such as ClinVar, and employs explainable, state-of-the-art machine-learning techniques to refine pathogenicity classification. The proposed approach, DITTO, can make predictions for any type of small variant, including SNVs, insertions, and deletions. It can also handle diverse variant consequence types, such as missense, splice site, and intergenic, across various transcripts and isoforms of a gene along with explanations of the predictions. This study shows that DITTO outperforms other state-of-the-art pathogenicity prediction methods. The second part of the dissertation focuses on classifying and prioritizing all small variants from a single gene for curation or experimental validation. The primary challenge with prioritizing variants in large genes, like NF1, is that many variants were previously unclassified, and different variants can cause different disease manifestations. The proposed approach, DITTO, accurately classified all previously classified variants based on agreement with existing functional and expert human-curated classifications. DITTO also aided in classifying all unclassified small variants and reclassifying uncertain class variants. DITTO performed well identifying some misclassified variants, prioritizing a subset of variants for follow-up curation or experimental validation. The third part of the dissertation addresses some critical challenges associated with diagnosing rare disease patients. The primary obstacle lies in manually curating and interpreting variants from the exome or genome sequencing data to identify potential disease-causing variants. This approach also demonstrates that applying a cutoff for the DITTO score, in conjunction with allele frequency (i.e., considering rare variants) and phenotype association with genes, significantly reduces the number of candidate variants requiring review. Furthermore, applying this approach to rare disease patients from Undiagnosed Diseases Network (UDN) reveals that reviewing only a limited number of top variant candidates, determined by the DITTO score, accurately identified molecular diagnosis for more than half the patients. In summary, this dissertation underscores the immense accuracy and precision of explainable, advanced machine-learning approaches for classifying and prioritizing disease-causing variants. These approaches also facilitate the reanalysis of ultra-rare, difficult-to-diagnose patients by reducing the number of variants for review. As a result, the diagnostic process can become automated, faster, and more cost-effective.

ProQuest ID

https://proquest.com/docview/3101846465