Using a type of artificial intelligence (AI) known as deep learning, Massachusetts Institute of Technology (MIT) researchers discovered a new class of antibiotics for drug-resistant Staphylococcus aureus (MRSA) bacteria.
According to Euronews, this AI technology helped scientists to unlock the first new antibiotics in 60 years. The transparent deep learning models reportedly not only identified this new compound that can kill a drug-resistant bacterium but also shed light on the information used by the AI model to make predictions about certain molecules that would make for good antibiotics.
New Class of Antibiotics Identified by AI
The researchers demonstrated that this newly identified compound exhibits the ability to eliminate methicillin-resistant Staphylococcus aureus (MRSA) in laboratory settings and two mouse models of MRSA infection. Importantly, this compound demonstrated low toxicity against human cells, making it a promising candidate for drug development.
The study, part of the Antibiotics-AI Project at MIT, is spearheaded by James Collins, the Termeer Professor of Medical Engineering and Science in MIT's Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering. The project aims to discover new classes of antibiotics targeting seven types of deadly bacteria over a seven-year period.
MRSA, a bacterium responsible for infecting more than 80,000 people in the United States annually, poses a significant health threat, often causing skin infections, pneumonia, and, in severe cases, sepsis, a potentially fatal bloodstream infection.
The researchers used deep learning models capable of learning chemical structures associated with antimicrobial activity. These models analyzed vast datasets, evaluating around 39,000 compounds for their antibiotic activity against MRSA.
Read Also: MIT Researchers Use Synthetic Imagery to Train Machine Learning Models
Monte Carlo Tree Search
The transparency achieved in this study is a notable advancement, as deep-learning models are typically regarded as "black boxes" due to the lack of insight into the features guiding their predictions.
Felix Wong and Erica Zheng, lead authors of the study, adapted the "Monte Carlo tree search" algorithm to unravel the deep-learning model's decision-making process. This algorithm not only estimated each molecule's antimicrobial activity but also predicted the substructures of the molecule likely responsible for that activity.
To refine the selection of potential drugs, the researchers trained additional deep learning models to predict toxicity against three different types of human cells. The combination of information on antimicrobial activity and toxicity yielded compounds capable of killing microbes while minimizing adverse effects on human cells.
In screening approximately 12 million commercially available compounds, the models identified promising candidates from five different classes, each based on distinct chemical substructures. Experimental validation of these predictions involved testing around 280 compounds, ultimately identifying two promising antibiotic candidates from the same class.
"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," Wong said in a statement.
"The molecules are attacking bacterial cell membranes selectively, in a way that does not incur substantial damage in 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," he added.
The findings of the research team were published in the journal Nature.
Related Article: MIT Merges Physics and AI for Image Generation With Poisson Flow Generative Model++
