A powerful new class of antibiotics capable of killing drug-resistant bacteria has been discovered by researchers at the Massachusetts Institute of Technology (MIT), by utilizing a subtype of artificial intelligence (AI) known as deep learning. Results from the study, published in the journal Nature, demonstrate the compound’s effectiveness against Methicillin-Resistant Staphylococcus Aureus (MRSA), a bacterium responsible for over 10,000 deaths annually in the US. The compounds showcased low toxicity against human cells, reinforcing their potential as drug candidates.
The study’s significant innovation is the scientists’ ability to determine the data the deep learning models used to predict the compounds’ antibiotic potency. Greater understanding of this process could enable the design of even more effective drugs. The researchers’ method offers a time and resource-efficient framework that is also able to provide structural insights into the antibiotics. This study is part of the Antibiotics-AI Project, a seven-year mission to discover new antibiotic classes.
MRSA, which infects more than 80,000 US residents annually, causes skin infections or pneumonia and can escalate to sepsis. Over the last few years, Collins and his team have utilized deep learning to discover new antibiotics, leading to potential drug candidates against several drug-resistant bacterium strains. The deep learning models identified these drug candidates by recognising antimicrobial chemical structures before comparing them to millions of others to predict new candidates with high antimicrobial activity.
The researchers trained the deep learning model using extensive datasets, comprising antibiotic activity of about 39,000 compounds against MRSA. To understand the model’s predictions, the scientists adapted an algorithm known as Monte Carlo tree search, which makes deep learning models more comprehensible. This algorithm allows estimation of each molecule’s antimicrobial activity and prediction of molecule substructures that might contribute to this activity.
Further candidate drugs were identified by training additional deep learning models to predict potential toxicity to different human cell types. Combining toxicity predictions with predictions of antimicrobial activity, the researchers discovered compounds with both antimicrobial properties and minimal adverse effects on the human body.
From over 12 million commercially available compounds, the researchers’ models predicted five different compound classes that could be active against MRSA. After purchasing and testing 280 compounds, two were confirmed as highly promising antibiotic candidates. These two compounds killed bacteria by disrupting their ability to maintain an electrochemical gradient across their cell membranes, necessary for critical cell functions.
The findings will be used for further detailed analysis by Phare Bio, an MIT-founded nonprofit part of the Antibiotics-AI Project, to ascertain the compounds’ chemical properties and potential clinical use. Furthermore, the successful methodology applied in this study can be adapted to discover new antibiotic classes for different pathogens, with efforts already underway to design new drug candidates.
