MIT researchers used a type of artificial intelligence known as deep learning to discover a class of compounds that can kill drug-resistant bacteria that cause more than 10,000 deaths each year in the United States.
In a study published today natureResearchers have shown that this compound can kill methicillin resistance. Staphylococcus aureus (MRSA) was grown in laboratory dishes and in two mouse models of MRSA infection. Additionally, this compound exhibits very low toxicity to human cells, making it a particularly good drug candidate.
The key innovation of the new study is that researchers were able to find out what kind of information deep learning models were using to predict antibiotic efficacy. This knowledge could help researchers design additional drugs that may work better than those identified in the model.
“The insight here is that we can look at what the model has learned to make predictions that a particular molecule will make a good antibiotic. Our work provides a time-efficient, resource-efficient, and mechanistically insightful framework from a chemical structure perspective in a way that has never been done before,” says James Collins, Termeer Professor of Medical Engineering and Science. He worked at the MIT Institute for Medical Engineering and Sciences (IMES) and the Department of Biological Engineering.
Felix Wong, a postdoctoral fellow at IMES and the Broad Institute at MIT and Harvard, and Erica Zheng, a former Harvard Medical School graduate student advised by Collins, are lead authors of the study, which is part of the Antibiotics-AI project. At MIT. The mission of the project, led by Collins, is to discover new classes of antibiotics against seven types of deadly bacteria over seven years.
explainable predictions
MRSA, which infects more than 80,000 people in the United States every year, often causes skin infections or pneumonia. In severe cases, it can lead to sepsis, a potentially fatal bloodstream infection.
Over the past few years, Collins and his colleagues at MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic) have begun using deep learning to find new antibiotics. Their research led to a potential drug. Acinetobacter baumanniiBacteria commonly found in hospitals and many other drug-resistant bacteria.
These compounds were identified using deep learning models that can learn to identify chemical structures associated with antibacterial activity. The model then examines millions of different compounds to predict which ones may have strong antibacterial activity.
Although this type of search has proven to be fruitful, one limitation of this approach is that the model is a “black box.” This means there is no way to know what features the model is basing its predictions on. If scientists know how the model makes predictions, it may be easier to identify or design additional antibiotics.
“What we were trying to do in this study was to open the black box,” says Wong. “These models are made up of a very large number of computations that mimic neural connections, and no one really knows what’s going on inside.”
First, the researchers trained a deep learning model using a significantly expanded dataset. They generated this training data by testing about 39,000 compounds for antibiotic activity against MRSA, then fed this data and information about the chemical structures of the compounds into the model.
“You can basically represent any molecule as a chemical structure, and you can also tell the model whether that chemical structure is antibacterial or not,” Wong said. “The model was trained on many examples like this. “Then, given a new molecule, a new arrangement of atoms and bonds, we can tell you the probability that that compound is predicted to be antibacterial.”
To figure out how the model made its predictions, the researchers applied an algorithm called Monte Carlo tree search, which has been used to make other deep learning models like AlphaGo more explainable. This search algorithm allows the model to generate not only an estimate of the antibacterial activity of each molecule, but also predictions about which substructure of the molecule is likely to account for that activity.
powerful activity
To further narrow down the candidate drugs, the researchers trained three additional deep learning models to predict whether the compounds would be toxic to three types of human cells. By combining this information with predictions of antibacterial activity, the researchers discovered a compound that could kill microorganisms with minimal side effects on the human body.
Researchers have screened approximately 12 million compounds using this collection of models, all of which are commercially available. In this collection, the model identified five different classes of compounds predicted to be active against MRSA based on the chemical substructure within the molecule.
By purchasing about 280 compounds and testing them on MRSA grown in laboratory dishes, the researchers were able to identify two that appeared to be very promising antibiotic candidates from the same class. In tests on two mouse models of MRSA skin infection and MRSA systemic infection, each compound reduced MRSA populations by 10-fold.
Experiments have shown that this compound kills bacteria by disrupting their ability to maintain electrochemical gradients across their cell membranes. This gradient is required for many important cellular functions, including the ability to produce ATP (a molecule that cells use to store energy). Halicin, an antibiotic candidate discovered in 2020 by Collins’ lab, appears to work by a similar mechanism, but only for Gram-negative bacteria (bacteria with thin cell walls). MRSA is a Gram-positive bacteria with a thicker cell wall.
“We have very strong evidence that this new class of structures is active against Gram-positive pathogens by selectively quenching the proton motive force of bacteria,” says Wong. “This molecule selectively attacks bacterial cell membranes in a way that causes no substantial damage to human cell membranes. “Our substantially enhanced deep learning approach allowed us to predict this new structure of the antibiotic and found that it was not toxic to human cells.”
The researchers shared their findings with Phare Bio, a nonprofit started by Collins and others as part of the Antibiotics-AI project. The non-profit organization now plans to conduct a more detailed analysis of the chemical properties of these compounds and their potential clinical uses. Meanwhile, Collins’ lab is working to design additional drug candidates based on the new findings and use the model to find compounds that can kill other types of bacteria.
“We are already leveraging similar approaches based on chemical substructure to design new compounds, and of course this approach could be immediately adopted to discover new classes of antibiotics against a variety of pathogens,” Wong said. said.
In addition to MIT, Harvard, and the Broad Institute, contributing institutions to this paper include Integrated Biosciences, Inc., the Wyss Institute for Biologically Inspired Engineering, and the Leibniz Institute of Polymer Research in Dresden, Germany. This research was supported by the James S. McDonnell Foundation, the National Institute of Allergy and Infectious Diseases, the Swiss National Science Foundation, the Banting Fellowships Program, the Volkswagen Foundation, the Defense Threat Reduction Agency, the National Institutes of Health, and the Broad Institute. The Antibiotics-AI project is funded by the Audacious Project, Flu Lab, Sea Grape Foundation, Wyss Foundation, and an anonymous donor.