MIT researchers have used a form of artificial intelligence called deep learning to identify a new classification of compounds that can effectively kill methicillin-resistant Staphylococcus aureus (MRSA), a drug-resistant bacterium accountable for over 10,000 deaths in the US each year. The researchers have demonstrated that these deep-learning-identified compounds not only have the ability to kill MRSA in laboratory conditions and animal models but also exhibit low toxicity against human cells, making them good candidates for medical drugs.
The novelty of this study, published in Nature, is that the scientists were also able to comprehend what kind of factors the deep-learning model used to predict antibiotic potency. Such understanding could aid researchers in designing more effective drugs. This innovative study is part of the Antibiotics-AI Project at MIT, a seven-year project aimed at identifying new classes of antibiotics against seven types of lethal bacteria.
Deep learning has previously aided the team in identifying potential drugs against hospital-dwelling bacterium Acinetobacter baumannii, along with other drug-resistant bacteria. The model identifies potential compounds among millions based on associated chemical structures that exhibit antimicrobial activity. However, these models work as “black boxes,” making it hard to recognize what features the predictions are based on, hence complicating the design or identification of additional antibiotics. The researchers sought to “open the black box” and shed light on the model’s inner workings.
Using an algorithm known as the Monte Carlo tree search, the researchers improved their understanding of how the model made its antibiotic predictions. They trained the model using data on 39,000 compounds and their antibiotic activity against MRSA, along with information on these compounds’ chemical alignments, then ran the model on any new molecule, predicting its antibacterial properties. The model offers a simultaneous prediction for which substructures of the molecule are likely accountable for its antimicrobial activity.
The researchers further refined their search by training additional models to predict toxicity levels against three different types of human cells and thus identified compounds that could eliminate microbes with minimal side effects on the human body. Approximately 280 of the 12 million commercially available compounds screened by the model were purchased and tested, of which five different classes demonstrated potential activity against MRSA. Two of these appeared to be particularly promising antibiotic contenders, reducing MRSA populations by a factor of ten in laboratory tests. Their mode of operation appears to disrupt bacteria’s ability to maintain an electrochemical gradient across their cell membranes, a mechanism crucial for many cellular functions.
The researchers have made their findings available to Phare Bio, a nonprofit part of the Antibiotics-AI Project, for further analysis of the compounds’ chemical properties and potential clinical uses. Meanwhile, they are using the successful model to seek compounds effective against other bacterial types and working on designing more drug compounds informed by these study’s outcomes.
