The discovery of novel structural classes of antibiotics is urgently needed to address the ongoing antibiotic resistance crisis. Deep learning approaches have aided in exploring chemical spaces; these typically use black box models and do not provide chemical insights.
A new MIT study used deep learning to find a class of compounds that can kill a drug-resistant bacterium that causes more than 10,000 deaths in the United States every year. The researchers demonstrated that these compounds can kill methicillin-resistant Staphylococcus aureus (MRSA) in laboratory dishes and two mouse models of MRSA infection. Notably, these compounds exhibit low toxicity against human cells, making them promising candidates for drug development.
A noteworthy aspect of this study is that the researchers identified the information utilized by the deep-learning model to predict the antibiotic potency of these compounds. This understanding could aid researchers in designing new drugs that may be even more effective than those initially identified by the model.
The work offers a time-efficient, resource-efficient, and mechanistically insightful framework from a chemical-structure standpoint in ways that we haven’t had to date.
Methicillin-resistant Staphylococcus aureus (MRSA) is a cause of staph infection that is difficult to treat because of resistance to some antibiotics.
MIT researchers have been using deep learning to find new antibiotics in recent years. They’ve discovered potential drugs against bacteria like Acinetobacter baumannii, commonly found in hospitals, and other drug-resistant strains. Deep learning models are trained to recognize chemical structures linked to antimicrobial activity, sifting through millions of compounds to predict effective ones.
However, a challenge is that these models are like “black boxes”—we don’t know what features they use to make predictions. If researchers understood how the models work, finding or designing more antibiotics would be easier.
Researchers in this study aimed to open that black box. They trained a deep learning model using substantially expanded datasets to do so. To train their deep learning model, the researchers tested around 39,000 compounds to see if they had antibiotic activity against MRSA. They provided the model with this testing data, including details about the chemical structures of the compounds.
To understand how the model made predictions, the researchers utilized a Monte Carlo tree search algorithm. This algorithm, previously employed to enhance explainability in other deep learning models like AlphaGo, enables the model to estimate each molecule’s antimicrobial activity. Additionally, it predicts which specific substructures of the molecule are likely responsible for that activity.
The researchers trained three more deep-learning models to check if the compounds harmed different types of human cells. Combining this data with the antimicrobial predictions, they found combinations that could kill microbes with minimal harm to human cells.
They analyzed about 12 million available compounds using these models, identifying active ones from five classes against MRSA. They purchased and tested 280 compounds, discovering two promising antibiotics from the same class. In mouse models of MRSA skin and systemic infections, these compounds significantly reduced MRSA.
Experiments showed that these compounds likely kill bacteria by disrupting their ability to maintain a critical electrochemical balance across cell membranes, vital for various cell functions. This mechanism is similar to the antibiotic halicin discovered in 2020, but these new compounds target Gram-positive bacteria like MRSA.
Felix Wong, a postdoc at IMES and the Broad Institute of MIT and Harvard, said, “We have pretty strong evidence that this new structural class is active against Gram-positive pathogens by selectively dissipating the proton motive force in bacteria. The molecules attack bacterial cell membranes selectively in a way that does not incur substantial damage to human cell membranes. Our substantially augmented deep learning approach allowed us to predict this new structural class of antibiotics and enabled the finding that it is not toxic against human cells.”
The researchers have shared their findings with Phare Bio, a nonprofit started by Collins and others as part of the Antibiotics-AI Project. The nonprofit plans to do a more detailed analysis of these compounds’ chemical properties and potential clinical use.
Journal Reference:
- Wong, F., Zheng, E.J., Valeri, J.A. et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature (2023). DOI: 10.1038/s41586-023-06887-8