MIT researchers, using deep learning techniques, have discovered compounds that can effectively combat methicillin-resistant Staphylococcus aureus (MRSA). This drug-resistant bacterium annually leads to over 10,000 deaths in the United States alone. Detailed in a study published in Nature, the compounds not only successfully killed MRSA in laboratory and mice model tests, but also showed significantly low toxicity against human cells, positioning them as promising antibiotic candidates.
This study’s pivotal innovation is the researchers’ ability to understand the informative elements the deep-learning model used for its antibiotic potency forecasts. This understanding enhances the ability to design more effective drugs. Additional drugs that might show higher efficacy than the ones presently being identified by the model.
With MRSA affecting more than 80,000 people in the U.S. annually, often leading to eventual sepsis, the need for effective antibiotics is pressing. Using deep learning, researchers have identified potential antibiotic candidates, their efforts yielding prospects effective against Acinetobacter baumannii, another common hospital bacterium, and several other drug-resistant bacteria. The AI models can perform exhaustive search through millions of compounds, identifying those that likely possess antimicrobial activity.
This study aimed to decipher the ‘black box’ of AI models and gain an understanding of their predictive processes. The researchers trained the model with a dataset of 39,000 compounds and their antibiotic activity against MRSA. The model was then able to predict a molecule’s antimicrobial property based on its atomic structure and bonds. To ascertain how this occurred, the team adapted a search algorithm called Monte Carlo tree search, which allowed the model to not only estimate each molecule’s antimicrobial activity but also provide a prediction for which substructures of the molecule accounted for this activity.
An additional step towards identifying ideal drug candidates was training three more deep learning models to predicate potential toxicity to three human cell types. Screening around 12 million commercially available compounds led to the identification of five classes of compounds predicted to be potent against MRSA.
Testing around 280 of these compounds led to the identification of two particularly promising antibiotics. It was discovered that these compounds kill bacteria by impacting their ability to maintain an electrochemical gradient across cell membranes, necessary for many critical functions. Importantly, these molecules selectively affect bacterial over human cells, meaning overall human toxicity should be minimal.
The nonprofit Phare Bio intends to conduct further analyses on these compounds’ chemical properties and feasible clinical applications. Collins’ lab intends to use the findings from the current study and the models to identify potential antibiotics for different bacteria types, signifying a significant step in the fight against antibiotic resistance.
