Using deep learning, a form of artificial intelligence, researchers at MIT have discovered a group of compounds that can eliminate methicillin-resistant Staphylococcus aureus (MRSA), a drug-resistant bacterium that causes over 10,000 deaths in the US annually. The compounds, which display low toxicity against human cells, are considered good potential drug candidates.
In a paper published in Nature, researchers described how the compounds effectively kill MRSA in a lab dish and in two MRSA-infected mouse models. An innovative aspect of the study is understanding the type of data the deep learning model utilizes to predict antibiotic potency. Insights from this could help in designing even better drugs. The research aligns with the Antibiotics-AI Project at MIT, helmed by James Collins. The project aims to unearth new classes of antibiotics for seven deadly bacterial types within seven years.
Deep learning has been used by the researchers to identify new antibiotics. This work has led to potential drugs against Acinetobacter baumannii, a germ often seen in hospitals, as well as other drug-resistant bacteria. Current methods, while fruitful, are limited as the models act as “black boxes” with no understanding of the prediction features. By shedding light on these models, scientists could design better drugs more efficiently.
In the study, the team trained a deep learning model using considerably increased data sets. Approximately 39,000 compounds were tested for antibiotic activity against MRSA. This data, with additional information on the compounds’ chemical structures, was plugged into the model, which generates predictions on a compound’s antimicrobial activity.
To enhance understanding of model predictions, the researchers used an algorithm called the Monte Carlo tree search. This assists the model in developing an estimate for each molecule’s antimicrobial activity. They trained three more models to predict compounds’ toxicity to different types of human cells. Blending these predictions diplayed that some compounds can effectively kill microbes with minimal human body side effects.
On scanning about 12 million compounds, the model identified five different classes with potential activity against MRSA. Two of these, from the same class, were found to be promising candidates after they were tested against MRSA in a lab. The studies showed that the compounds disturb the bacteria’s ability to maintain an electrochemical gradient across their cell membranes, an essential function for cell activities. Collins’ lab identified another antibiotic candidate, halicin, which works similarly but is specific to Gram-negative bacteria. MRSA is a Gram-positive bacterium.
The team has shared their data with Phare Bio, a nonprofit launched by Collins. This entity now plans to conduct a more comprehensive analysis of the chemical properties and possible clinical applications of these compounds. Simultaneously, Collins’ lab is aiming to design additional drug candidates based on the new study, and utilizing the models to discover compounds that can terminate other bacteria types. Other contributors to the research paper include Harvard, the Broad Institute, Integrated Biosciences, Inc., the Wyss Institute for Biologically Inspired Engineering, and Germany’s Leibniz Institute of Polymer Research.
