Utilizing deep learning, researchers from the Massachusetts Institute of Technology (MIT) have identified a new class of compounds that effectively kill methicillin-resistant Staphylococcus aureus (MRSA), a drug-resistant bacterium responsible for over 10,000 deaths each year in the United States. The compounds have been found to have low toxicity against human cells, a key characteristic for potential drug candidates.
The success of this research is grounded in the team’s ability to decipher the deep-learning model’s decision-making process, improving the model’s antibiotic potency predictions. This advance means scientists could design additional drugs that may prove superior to those identified by the model.
The team behind this development is part of the Antibiotics-AI Project at MIT, led by James Collins. The Project’s goal is to discover new types of antibiotics to fight seven dangerous bacteria over seven years.
MRSA alone infects over 80,000 people in the US annually. It can cause a variety of illnesses including skin infections, pneumonia, and sepsis. In their search for new antibiotics, the MIT researchers used deep learning to identify chemical structures linked to antimicrobial activity. However, the way these models make predictions has typically been considered a “black box”, with no clear insight into the underlying criteria.
In this study, researchers began ‘opening the black box’ by training their deep learning model with a dataset significantly larger than before. The dataset contained the results of 39,000 compounds tested for MRSA antibiotic activity, with additional information about the compounds’ chemical structures. The researchers implemented an algorithm known as Monte Carlo tree search to determine how the model was reaching its predictions, further improving the predictions’ accuracy.
Building upon this work, the researchers also trained three additional deep learning models to predict a compound’s toxicity to three types of human cells. Using these models, the team screened roughly 12 million compounds, identifying five classes that appeared to have MRSA efficacy.
After purchasing and testing about 280 compounds, the researchers found two compounds from the same class particularly promising. Tests in two mouse models showed the compounds reduced the MRSA population by a factor of ten. The compounds seem to work by disrupting the bacteria’s ability to maintain an electrochemical gradient across their cell membranes, a function vital to cellular activity.
The findings have been shared with Phare Bio, a nonprofit established as part of the Antibiotics-AI Project. The group is planning more detailed analyses of these compounds, while Collins’ lab is working on designing additional drug candidates and exploring how the models can be used to counter other types of bacteria. The significantly improved deep learning approach has made it possible to predict new classes of antibiotics and determine they are not toxic to human cells. The team is using the approach to design new compounds, using chemical substructures, as well as leveraging similar methods to target different pathogens.
Research for the project was sponsored in part by the James S. McDonnell Foundation, the U.S. National Institute of Allergy and Infectious Diseases, the Swiss National Science Foundation, the Volkswagen Foundation, the U.S. National Institutes of Health, and the Broad Institute.
