A group of researchers are using deep learning technology to create an AI model that can identify compounds that can be used to create new antibiotics, an important step during a time when drug-resistant bacteria represent a growing global public health threat.
The researchers, from the Massachusetts Institute of Technology (MIT) and biotech company Integrated Biosciences, used deep learning techniques to discover a class of compounds that can kill Staphylococcus aureus – commonly known as MRSA – a methicillin-resistant bacteria that causes more than 10,000 deaths every day in the United States.
Combined with their low threat of toxicity to human cells, the compounds make good candidates to be developed into drugs to combat the health threat, the researchers explained in a recent article in Nature.
Along with being able to identify compounds that could be used in drugs to push back against MRSA, the deep learning model gave the researchers visibility into how it was making its predictions about the potency of antibiotics, which could make it easier to help other researchers design additional drugs that might work even better than the ones identified by the model, they wrote.
Essentially, the model isn’t a “black box” that makes it impossible for researchers to see how the predictions are being made.
“The insight here was that we could see what was being learned by the models to make their predictions that certain molecules would make for good antibiotics,” James Collins, the Termeer Professor of Medical Engineering and Science at MIT, a lead researcher, and founding chair of Integrated Bioscience’s advisory board, said in a statement. “Our work provides a framework that is time-efficient, resource-efficient, and mechanistically insightful, from a chemical-structure standpoint, in ways that we haven’t had to date.”
New Antibiotics Needed
In a 37-page paper on their work, the researchers – who include Felix Wong, a co-founder of Integrated Biosciences, and Erica Zheng, a former Harvard Medical School graduate student advised by Collins – outlined the growing threat of antibiotic-resistant bacteria driven by the lack of new antibiotics and the widespread use in such areas a health care and agriculture. Bacteria have evolved to become resistant to current antibiotics, with stronger defenses against them.
Without new antibiotics, deaths worldwide due to resistant infections could reach 10 million a year by 2050, they wrote. Over the past decade, antibiotic candidates were discovered via various approaches, and more recently Collins used a deep learning model to show that known compounds halicin – once thought to hold promise for diabetes treatment – and abaucin included antibiotic capabilities.
However, those discoveries were found using a black-box model. The latest model is designed to let researchers understand how its decisions were made.
“What we set out to do in this study was to open the black box,” Wong said in a statement. “These models consist of very large numbers of calculations that mimic neural connections, and no one really knows what’s going on underneath the hood.”
First up were the compounds that can kill MRSA, which infects more than 80,000 people a year in the United States and can cause skin infections or pneumonia or can lead to sepsis, a bloodstream infection that can be fatal, the researchers said.
The deep learning model was trained on data created by testing about 39,000 compounds for antibiotic activity against MRSA. That data, along with information about the chemical structures of the compounds, were then fed into the model.
Wong said any molecule can be represented as a chemical structure and the model can be told if the chemical structure is antibacterial.
“The model is trained on many examples like this,” he said. “If you then give it any new molecule, a new arrangement of atoms and bonds, it can tell you a probability that that compound is predicted to be antibacterial.”
Three other deep learning models were used to predict whether compounds were toxic to human cells. Using all this information, the researchers screened about 12 million compounds and the model identified those that could be used to kill MRSA.
Testing about 280 compounds against MRSA grown in a lab and testing two in experiments with mice – one a MRSA skin infection, the other a systemic infection – the researchers found the compounds reduced the MRSA population by a factor of 10.
No ‘Black Box’ Here
They also adapted an algorithm known as the Monte Carlo tree search to determine how the model was making its predictions, removing the black-box nature found in other models. The search algorithm lets the model generate an estimate of each molecule’s antimicrobial activity, as well as a prediction for which substructures of the molecule likely cause the activity.
“This is one of the first demonstrations that deep learning models can explain what they are predicting, with immediate and far-reaching implications to how drug discovery is done and how efficiently we can find new drugs using AI,” Wong said.
Other researchers also are turning to AI techniques to address issues of drug-resistant diseases, as shown with the research done with halicin and abaucin. In November 2023, scientists at the University of Oxford announced an AI-based method that can detect antimicrobial resistance (AMR) in as few as 30 minutes
“This method relies on training deep-learning models to analyze bacterial cell images and detect structural changes that may occur in cells when they are treated with antibiotics,” the researchers wrote, adding that the “method was shown to be effective across multiple antibiotics, achieving at least 80% accuracy on a per-cell basis.”
The MIT researchers shared their findings with Phare Bio, a nonprofit that is using AI and deep learning to address urgent health threats. It’s part of the Antibiotics-AI Project created by Collins and others in 2020 to create new classes of antibiotics. Phare Bio will run a more detailed analysis of the chemical properties and possible clinical uses of the compounds.
In addition, Collins’ lab is using the findings to try to design additional drug candidates and using the deep learning models to search for compounds that can kill other types of bacteria.