A novel category of antibiotics designed to combat drug-resistant Staphylococcus aureus (MRSA) bacteria has been uncovered through the application of more transparent deep learning models.
Artificial intelligence (AI) is reshaping the landscape of medicine, aiding scientists in uncovering the first set of new antibiotics in six decades.
The identification of a potent compound capable of eliminating a drug-resistant bacterium responsible for thousands of global fatalities annually represents a potential breakthrough in the battle against antibiotic resistance.
According to James Collins, a professor of Medical Engineering and Science at MIT and a co-author of the study, the innovative aspect lies in the ability to understand the models’ learning process, predicting molecules suitable for effective antibiotics. He emphasized the efficiency and mechanistic insights achieved, particularly in terms of chemical structures.
Published in Nature and collaboratively written by a team of 21 researchers, the study aimed to demystify the workings of deep learning, employing it to predict the activity and toxicity of a new compound.
Deep learning, utilizing artificial neural networks for automatic learning and representation of features from data without explicit programming, is increasingly pivotal in drug discovery. It accelerates the identification of potential drug candidates, predicts their properties, and optimizes the drug development process.
In this instance, the research targeted methicillin-resistant Staphylococcus aureus (MRSA), a bacterium causing infections ranging from mild skin issues to severe, potentially life-threatening conditions.
With almost 150,000 MRSA infections annually in the European Union, leading to nearly 35,000 deaths from antimicrobial-resistant infections, the stakes are high.
The MIT researchers enlarged a deep learning model using extensive datasets. Around 39,000 compounds were assessed for their antibiotic activity against MRSA, with resulting data and compound chemical structures input into the model to create the training data.
To refine potential drug selections, three additional deep-learning models evaluated compound toxicity on distinct human cell types. By integrating toxicity predictions with previously determined antimicrobial activity, the researchers identified compounds effective against microbes with minimal harm to the human body.
Approximately 12 million commercially available compounds were screened, leading to the identification of promising antibiotic candidates from five different classes. Experiments involving mouse models confirmed the efficacy of two compounds in reducing MRSA populations in both skin and systemic infections.
