Dissertation
Predicting antibiotic resistance using machine learning
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2023
DOI: 10.25820/etd.007285
Abstract
Bacteria gain resistance mechanisms to antimicrobial agents over time. These resistance mechanisms lead to increasing Antimicrobial Resistance (AMR). Increasing AMR leads to higher morbidity and mortality over time. Because of this, AMR will become the leading cause of death in the world by 2050. The increase in morbidity and mortality is due to the time to determine Minimum Inhibitory Concentration (MIC) values. An MIC is a minimum concentration of an antimicrobial agent required to inhibit the visible growth of bacteria. MICs, along with clinical breakpoints or epidemiological breakpoints, determine if a bacterial isolate is susceptible or resistant to a given antimicrobial agent. Determining MICs is a time-consuming process that takes anywhere from 24 to 72 hours. During this time, a physician prescribes an empirical antimicrobial therapy to the patient to help stave off the bacterial infection. As AMR increases, these empirical therapies become less effective, leading to less time available to determine MICs.
Lowering the time to determine MICs is of key importance. One approach is to use Machine Learning (ML) to predict MICs. In this thesis, many avenues for ML are examined to find the most effective algorithm, input dataset, and processing of the dataset to create a generalizable ML model to predict MICs. A simulation can then use this generalizable model to determine future antimicrobial agents and future resistance mechanisms before they appear. This simulation utilizes two Genetic Algorithms to perform this task. With this simulation, physicians could know future AMR, and know how to counteract that AMR before a patient enters a hospital. Physicians being able to predict future AMR would lead to few MICs needing to be predicted for/determined and quicker response times rather than predicting MICs alone.
All the following algorithms are discussed including; Decision Trees, Random Forests, K-Nearest Neighbors, XGBoost, Dense Neural Networks, Convolutional Neural Networks, and Recurrent Neural Networks. Out of these algorithms, a Recurrent Neural Network is what was able to create a generalizable model using a bacterial isolate's assembled DNA sequence (in the form of contigs) and an antimicrobial agent's Simplified Molecular Input Line Entry System (SMILES) string. The combination of these two datasets was what gave the highest accuracy from all datasets tested, including; gene acquisition data, contigs alone, demographic data, K-Mer counts, and Single Nucleotide Polymorphisms (SNP).
Details
- Title: Subtitle
- Predicting antibiotic resistance using machine learning
- Creators
- Cory Kromer-Edwards
- Contributors
- Suely Oliveira (Advisor)Mariana Castanheira (Committee Member)David Stewart (Committee Member)Juan Pablo Hourcade (Committee Member)Colleen C Mitchell (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Computer Science
- Date degree season
- Summer 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007285
- Number of pages
- xvi, 148 pages
- Copyright
- Copyright 2023 Cory Kromer-Edwards
- Language
- English
- Date submitted
- 05/15/2023
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (pages 135-148).
- Public Abstract (ETD)
Antibiotic resistance is ever-increasing which leads to higher morbidity and mortality. Because of the increasing resistance, the time for a physician to prescribe a proper antibiotic to a patient is becoming shorter. To prescribe an antibiotic, Minimum Inhibitory Concentrations (MIC) must be determined. These MICs are then used to determine which antibiotics the bacteria is susceptible to or resistant to.
The complication with MICs is that they need 24 to 72 hours to determine. As antibiotic resistance increases, the 24 to 72 hours to determine an MIC remain the same. Physicians prescribe an initial antibiotic therapy while MICs are determined. This initial antibiotic becomes less effective as resistance increases. The combination of the time to determine MICs staying the same and the initial antibiotic becoming less effective leads to increased morbidity and mortality from bacterial infections. By 2050, bacterial infections will be the leading cause of death due to this rising morbidity and mortality.
Machine Learning can predict many MICs within minutes making it quick and efficient. This efficiency can lead to expedited treatment for patients. This thesis delves into many Machine Learning algorithms (including XGBoost, KNearest Neighbors, Random Forest, Dense Neural Networks, Decision Trees, and Recurrent Neural Networks) and datasets (including gene acquisition, K-Mer counts, patient demographics, and Single Nucleotide Polymorphisms). The goal was to determine which combinations of algorithms and datasets lead to the highest accuracy model. That model was a Recurrent Neural Network that takes an input of DNA from the bacteria and an antibiotic’s chemical structure to predict MICs. This model is able to predict new antibiotics that were not in the training dataset with high accuracy. Due to this model being able to generalize to new antibiotics, it was then used to simulate the creation of new resistance and new antibiotics to counter that resistance.
- Academic Unit
- Computer Science
- Record Identifier
- 9984454542102771
Metrics
5 File views/ downloads
47 Record Views