Interpretable Deep Neural Networks for More Accurate Predictive Genomics and Genome-Wide Association Studies - A Dissertation

Aug 2, 2026

Upload by Patrick

The dissertation presents interpretable deep neural network approaches to improve predictive genomics, with a focus on genome-wide association studies and polygenic risk score estimation. It develops deep learning and multi-task learning models for breast cancer and expands interpretability beyond first-order feature attribution to structured, second-order explanations. An additional linearizing neural network architecture, LINA, is introduced to enable more accurate first- and second-order interpretations, supported by benchmarking on real-world and synthetic datasets. Comparative analyses evaluate estimation quality against statistical and alternative machine learning baselines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages149

Reading Time6 h 15 min

AccessFree to download

Acknowledgments

List of Tables

List of Figures

Abstract

Introduction

Machine Learning (ML)

Interpretable Machine Learning

Supervised Machine Learning Algorithms

Genomics

Genome-Wide Association Studies

Deep neural network improves the estimation of polygenic risk scores for breast cancer

Methods

Results and Discussion

LINA: A Linearizing Neural Network Architecture for Accurate First-Order and Second-Order Interpretations

Methods

Data and Experimental Setup

Results and Discussion

Explainable multi-task learning improves the parallel estimation of polygenic risk scores for many diseases through shared genetic basis

Methods

Results

UNIVERSITY OF OKLAHOMA GRADUATE COLLEGE INTERPRETABLE DEEP NEURAL NETWORKS FOR MORE ACCURATE PREDICTIVE GENOMICS AND GENOME-WIDE ASSOCIATION STUDIES. A DISSERTATION SUBMITTED TO THE GRADUATE FACULTY in partial fulfillment of the requirements for the Degree of DOCTOR OF PHILOSOPHY By ADRIEN BADR´E Norman, Oklahoma 2023 INTERPRETABLE DEEP NEURAL NETWORKS FOR MORE ACCURATE PREDICTIVE GENOMICS AND GENOME-WIDE ASSOCIATION STUDIES. A DISSERTATION APPROVED FOR THE SCHOOL OF COMPUTER SCIENCE BY THE COMMITTEE CONSISTING OF Dr. Chongle Pan (Chair) Dr. Dean Hougen Dr. Qi Cheng Dr. Henry Neeman Dr. Krithivasan Sankaranarayanan © Copyright by ADRIEN BADR´E 2023 All Rights Reserved. Acknowledgments I want to thank Dr. Chongle Pan for being my committee chair, advisor, coauthor, mentor, and friend. His support, specifically during the pandemic, and tireless persistence helped me realize this work. None of my achievements would have been possible without him. I would also like to thank my wife and the love of my life, Charissa, for her endless love and support through this journey. Without her, I would have collapsed many times. I also want to thank my parents, J´erˆome and Sophie, my sisters, Aurore and Flore, and my family for their continuous support. Next, I would like to thank Virginie Perez-Woods. She was an amazingly helpful academic coordinator. More than being a great friend, she was a french presence and launched my dissertation journey in the best possible way. I also want to address my gratitude to all the current staff at the CS department that helped me and guided me during this fantastic experience. Then, my gratitude goes to all my labmates, and I wish them the best for their Ph.D. journey. Finally, I would like to thank the professors (Dr. Cheng, Dr. Hougen, Dr. Neemanand Dr. Sankaranarayanan) on my committee for their time and advice. Table of Contents chapterAcknowledgmentsivsection*.1 List Of Tables viii List Of Figures ix Abstract xiii 1 Introduction 1 1.1 Machine Learning (ML) .......................... 1 1.1.1 Types of Machine Learning .................... 1 1.1.1.1 Supervised Learning ................... 1 1.1.1.2 Unsupervised Learning .................. 3 1.1.2 Regression vs Classification .................... 3 1.1.3 Interpretable Machine Learning .................. 6 1.2 Supervised Machine Learning Algorithms ................. 8 1.2.1 Linear Regression .......................... 8 1.2.2 Logistic Regression ......................... 9 1.2.3 Deep Learning ............................ 10 1.2.3.1 Artificial Neural Network ................ 10 1.2.3.2 Multi-task learning .................... 18 1.3 Genomics .................................. 18 1.4 Genome-Wide Association Studies ..................... 22 1.5 Conclusion .................................. 23 2 Deep neural network improves the estimation of polygenic risk scores for breast cancer 25 2.1 Introduction ................................. 26 2.2 Methods ................................... 28 2.2.1 Breast cancer GWAS data ..................... 28 2.2.2 Development of deep neural network models for PRS estimation 30 2.2.3 Development of alternative machine learning models for PRS estimation ............................... 31 2.2.4 Development of statistical models for PRS estimation ...... 32 2.2.5 DNN model interpretation protocol ................ 33 2.3 Results and Discussion ........................... 33 2.3.1 Development of a machine learning model for breast cancer PRS estimation .............................. 34 2.3.2 Comparison of the DNN model with statistical models for breast cancer PRS estimation ....................... 40 2.3.3 Interpretation of the DNN model ................. 47 3 LINA: A Linearizing Neural Network Architecture for Accurate FirstOrder and Second-Order Interpretations 54 3.1 Introduction .................................

FAQs

What problem does the dissertation address in predictive genomics?

It targets more accurate estimation of polygenic risk scores using genome-wide association studies, while making deep models interpretable. The work aims to connect predictive performance with explanation mechanisms.

How does the dissertation improve interpretability beyond standard feature attribution?

It introduces a linearizing neural network architecture (LINA) designed for accurate first-order and second-order interpretations. The protocol supports both instance-wise and model-wise interpretations.

What models are evaluated for polygenic risk score estimation?

Deep neural network models, alternative machine learning and statistical baselines, and explainable multi-task learning models are compared. Breast cancer-focused experiments and many-disease parallel estimation are both discussed.

Aug 2, 2026

Upload by Patrick