Applying Neural Networks to Predict Antibiotic-Resistant Bacterial Infection from Shotgun Metagenomic Data

Saturday, September 16, 2017 - 9:00am - 9:30am
Keller 3-180
David Haslam (Cincinnati Childrens Hospital Medical Center)
There is an urgent need to develop new ways to combat the threat of rising antibiotic resistance rates, which has been nationally highlighted explicitly in recent action plans of the CDC and the White House. Our laboratory has found that hospitalized patients who develop infection almost always have the infecting bacterium on their body before they become ill, usually in the intestine but occasionally on the skin or in the mouth. We are using shotgun metagenomic sequencing to capture DNA sequences of all the bacteria on a patient’s body with the goal of identifying patients at risk of infection due to antibiotic resistant bacteria and intervening before the patient becomes ill. We call this approach Precision Metagenomics (PM).

A technical limitation of PM screening is the accuracy with which genetic information correlates with actual antibiotic resistance. For some organisms and many antibiotics there are clear and accurate correlates between genetic information and resistance testing results. However, prediction of antibiotic resistance from genomic data is more complicated for many organisms, particularly those in the Enterobacteriaceae family. Accurate prediction of antibiotic resistance is essential if metagenomic screening is to be used as a diagnostic method. To that end, we have developed machine learning algorithms that predict antibiotic resistance from raw metagenomic data. Using paired genetic and antibiotic resistance data from several hundred bacteria we trained machine learning algorithms to classify the bacteria as susceptible or resistant. We have found that neural networks have hightest accuracy and are able to accurately predict antibiotic resistance from metagenomic data, including samples that contain Enterobacteraceae. Our future work will optimize genetic analysis of resistant bacteria and neural network hyperparameters.