Using AI, Genes and Game Theory on Antimicrobial Resistance

Source: Geralt/Pixabay

Antimicrobial resistance (AMR) is the ability of microorganisms like bacteria, viruses, fungi and certain parasites to resist drugs such as antibiotics, antifungals, and antivirals from destroying it. AMR is a worldwide public health threat that is projected to rise. Globally, by 2050, over 10 million deaths per year will be due to antimicrobial resistance according to projections from a report by Wellcome Trust and the UK government. For antibiotic resistance alone, each year over two million people in the U.S. are affected, and 23,000 die, according to figures from the U.S. Centers for Disease Control and Prevention (CDC). Researchers at Washington State University have combined game theory with artificial intelligence (AI) to create a tool that can identify genes that are antibiotic-resistant in bacteria, and published their study in Scientific Reports on October 9, 2019.

Having a way to identify antimicrobial resistance in bacteria is critical in deciding which antibacterial drugs to use for infections. One method is to culture the bacteria in lab dishes and observe the impact of different drugs in various doses on bacterial growth. This is a slow process that may take many days. In many cases, bacteria cannot be cultured in a lab dish. For these, metagenomic analysis may be used. Metagenomics is the analysis of the DNA of microbes in environmental samples without the need for in vitro culturing.

In recent years, the rise of whole genomic sequencing has enabled researchers to discover genes that are resistant to antibiotics by performing database searches for similar known antimicrobial resistance sequences. The gap in this method is in researching resistance genes with sequences that are not similar to known antimicrobial resistance sequences, which may result in false negatives.

To address this gap, scientists have applied AI machine learning to identify potential genes that are believed to be antimicrobial resistant. In February of 2018, Virginia Tech researchers announced DeepARG, a deep learning model with an artificial neural network classifier that can predict antibiotic-resistant genes from metagenomic data with high precision and recall.

Later in October 2018, UC San Diego researchers published a study in Phys.org that demonstrated a machine learning algorithm that could accurately predict 33 known antibiotic-resistant genes and 24 new genetic signatures of antimicrobial resistance using the genomic sequences and phenotypes of over 1500 strains of tuberculosis-causing bacteria.

A month later in November 2018, a team of French researchers and their colleagues published in Nature Microbiology another machine learning approach to identifying antimicrobial-resistant genes by using three-dimensional protein structure information called pairwise comparative modeling (PCM). In PCM, two structure models are constructed for each candidate sequence for AMR and non-AMR sequences, then machine learning is used to determine the best structural model for the sequence. From a catalogue of 3.9 million proteins, the French researchers' PCM approach predicted over 6,000 antibiotic resistance determinants.

According to the Washington State University researchers, what sets their recently published study apart is that they applied a feature reduction strategy with their machine learning. “Identifying important features from a set of features to attain high classification accuracy is a challenging problem in machine learning because irrelevant or redundant features can compromise accuracy,” wrote the researchers in the study.

In their model, they took into account the many possible candidate features for protein sequences based on the characteristics, including physicochemical, composition, structural, and evolutionary traits. Taking it one step further, the researchers incorporated the relevance and interdependency of features in predicting antimicrobial resistance.

To achieve this, the Washington State University researchers applied game theory to their tool. Game theory is a branch of mathematics used to analyze situations where decisions made by the players are interdependent. Introduced in 1944 by Princeton University mathematician John von Neumann and economist Oskar Morgenstern, game theory is used in psychology, behavioral science, economics, business, social science, political science, computer science, and more fields.

Artificial Intelligence Essential Reads

The Emergence of Private LLMs

4 Surprising Ways AI Poses a Threat to Humanity

Applying game theory and machine learning toward researching complex biological systems is a clever way to bring a greater level of specificity because it takes into account the interdependencies of features. Unlike other models, just because a single variable is poorly predictive does not necessarily automatically eliminate it completely as a candidate. In this new type of model, the interactions of the features as a unit are analyzed by deploying game theory for machine learning.

“We consider features from the protein sequences—both AMR and non-AMR—of the Gram-negative bacterial genera Acinetobacter, Klebsiella, Campylobacter, Salmonella, and Escherichia for acetyltransferase (aac), β-lactamase (bla), and dihydrofolate reductase (dfr),” reported the Washington State University researchers. “Next we apply our game theory approach to select a small subset of features from the bacterial protein sequences, and finally we utilize this small feature subset to predict AMR genes using a support vector machine (SVM).”

“By using game theory to choose which protein characteristics to use in our machine learning model, we can predict AMR protein sequences for Gram-negative bacteria with an accuracy ranging from 93% to 99%,” reported the researchers. “With growth in both antimicrobial resistance and the number of sequenced genomes available, implementation of machine learning models for accurate prediction of AMR genes represents a significant development toward new and more accurate tools in the field of predictive antimicrobial resistance.”

As next steps, the researchers plan to create a user-friendly version of their model that will be available for public use for predicting antimicrobial resistance in bacteria in the future.

Copyright © 2019 Cami Rosso All rights reserved.