Using deep learning, a form of artificial intelligence, MIT researchers have identified a new class of compounds capable of killing methicillin-resistant Staphylococcus aureus (MRSA), a drug-resistant bacterium responsible for over 10,000 deaths annually in the United States. The findings were published in the journal Nature.
The compounds exhibited strong activity against MRSA in lab conditions and in two mouse models, while displaying low toxicity against human cells, marking them as potential drug candidates. Crucially, the researchers also deciphered the deep-learning model’s process of predicting antibiotic potency, providing knowledge which could be central in designing future therapies.
The compounds were discovered using deep learning models trained to identify chemical structures associated with antimicrobial activity. The models then sifted through millions of compounds predicting which of them might exhibit strong antimicrobial action. Despite the success of this strategy, a hindrance was the inability to understand the basis of the model’s prediction, limiting the possibility of identifying or designing additional antibiotics.
To glean understanding, the researchers trained the deep learning model using an expanded dataset created by testing roughly 39,000 compounds for MRSA antibiotic activity. Information on the compounds’ chemical structures was also fed into the model. An algorithm, Monte Carlo tree search, was adapted to understand the model’s predictions. This algorithm allows the prediction not just of a molecule’s antimicrobial activity, but also the substructures of the molecule responsible for that activity.
Working with three additional deep learning models, the researchers screened approximately 12 million compounds, identifying five different classes of compounds bearing potential action against MRSA. The researchers then purchased about 280 of these compounds for testing, identifying two promising antibiotic candidates. These compounds appear to kill bacteria by disrupting their ability to maintain an electrochemical gradient across cell membranes, vital for functions such as ATP production.
Further analysis of the compounds’ chemical properties and potential clinical use will be performed by Phare Bio, a nonprofit connected to MIT’s Antibiotics-AI Project. Collins’ laboratory will continue their exploration by designing additional drug candidates using the predictive models.
