Researcher scientists from Massachusetts Institute of Technology (MIT) used artificial intelligence to identify a new type of antibiotic. By analyzing over one hundred million chemical compounds in days which can kill 35 types of potentially deadly bacteria, including some strains resistant to all known antibiotics. Additionally, it cleared infections in two different mouse models. The MIT team has created a platform that may offer drug developers the opportunity to utilize the power of AI to lead in a new age of antibiotic drug discovery. Moving forward, the paradigm shifting approach leverages deep learning during the entire antibiotic drug development Lifecyle.
In what the World Health Organization labels “one of the biggest threats to global health security and development today,” antibiotic-resistant infections have risen in the past years. Moreover, there are few antibiotics in the pipeline and most that are include variants of existing drugs. Furthermore, existing methods for screening new antibiotics are prohibitively costly, represent material investment and are limited to a narrow spectrum of chemical diversity, reports Anne Trafton with MIT News. The growing antibiotic resistance, a disturbing trend threatening public health is intensified by two trends including 1) an increasing number of resistant pathogens and 2) a weak antibiotic drug development pipeline in biopharma industry.
The team was led out of MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health (J-Clinic) and the results were recently published in the journal Cell. The team developed, via machine learning-powered computer models, the ability to explore, in silico, large chemical spaces that otherwise would be prohibitively too expensive.
The utilization of predictive computer models for “in silico” drug screening has been around for a while now but such models have lacked sufficient accuracy to transform drug discovery. In previous models, molecules are represented as vectors reflecting the presence or absence of certain chemical groups. The MIT model, representing new neural networks, reports Ms. Trafton, can learn representations automatically, mapping molecules into continuous vectors which in turn are utilized to predict their properties.
The MIT team designed a model that can look for chemical features that enable molecules effectivity at killing E. coli. To pull this off, they had to train the model on about 2,500 molecules, including approximately 1,700 FDA-approved drugs and a set of 800 natural products with diverse structures and a wide range of bioactivities. The researchers sufficiently trained the model and then worked with the Broad Institute’s Drug Repurposing Hub, a library of approximately 6,000 compounds. They applied the model against the thousands of compounds and the model selected one molecule that it calculated to possess superior antibacterial activity in addition to a unique chemical structure—e.g. it was different than existing antibiotics. The MIT researchers also ran machine learning-based analysis to determine it would likely have low toxicity to human cells.
The Molecule: Halicin
The AI-based model selected was named by the team: halicin (after the AI system from the film “2001: A Space Odyssey” which had been previously identified as a possible diabetes drug. In fact, the team ran a series of tests against dozens of bacterial strains, isolated from patients and grown in lab dishes. They found that this molecule could kill many bacteria currently resistant to treatment, such as Clostridium difficile, Acinetobacter baumannii, and Mycobacterium tuberculosis. In fact, this drug killed all bacteria except for a difficult-to-treat lung pathogen called Pseudomonas aeruginosa.
In very early preclinical animal studies, the team investigated halicin’s ability to treat infected mice with a super resistant strain of A. baumannii, a bacterium known to infect U.S. soldiers in Iraq and Afghanistan. The application of halicin-containing ointment cleared the infection in 24 hours.
It would appear based on the MIT research that the identified molecule disrupts bacteria’ ability to maintain an electrochemical gradient across their cell membranes, reports MIT News. In this way, bacteria cannot produce sufficient molecules to store energy leading to the cells’ death. Researchers believe that bacteria may have a difficult time developing resistance to this form of attack.
When applied to E. coli, the team found the bacteria couldn’t develop resistance during a 30-day treatment period. This was in stark comparison to ciprofloxacin which in a test the researchers found the bacteria had developed resistance within one to three days—by day 30 the bacteria grew its resistance to ciprofloxacin by 200 times.
Platform at Work: Optimized Molecules
The research team used the AI-driven model to screen over 100 million molecules from the ZINC15 database, an online collection of approximately 1.5 billion chemical compounds. The model was able to screen the entire database in only three days—in the process identifying 23 optimized candidates (e.g. structurally dissimilar form existing antibiotics and predicted to be nontoxic to human cells).
Further Lab Tests
The researchers then executed tests against five bacteria species and discovered eight of the 23 molecules exhibited antibacterial activity while two of them were significantly robust.
This research included a number of funding contributors including:
· Abdul Latif Jameel Clinic for Machine Learning in Health
· Defense Threat Reduction Agency
· Broad Institute
· DARPA Make-It Program
· Canadian Institutes of Health Research
· Canadian Foundation for Innovation
· Canada Research Chairs Program
· Banting Fellowships Program
· Human Frontier Science Program
· Pershing Square Foundation
· Swiss National Science Foundation
· National Institutes of Health Early Investigator Award
· National Science Foundation Graduate Research Fellowship Program
· Anita and Josh Bekenstein
MIT researchers plan on developing additional studies centering on the potential of halicin. They will consider approaching either a biopharma company or nonprofit organization to consider drug development studies. Moreover, as the research team, via the powerful AI-driven drug discovery platform, identified several other promising antibiotic candidates, further tests will be conducted. Ultimately, the model can be used to design new drugs.
Now the research team members will continue leveraging the ZINC15 database, using the model to design new antibiotics and optimizing existing molecules. They will consider advanced optimization schemes, such as training the model to add features that make a particular antibiotic target only select bacteria while ensuring it doesn’t kill beneficial bacterial in a patient’s digestive tract for example.
James Collins, the Termeer Professor of Medical Engineering and Science, MIT’s Institute for Medial Engineering and Science (IMES) and Dept. of Biological Engineering
Regina Barzilay, the Delta Electronics Professor of Electrical Engineering and Computer Science in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).
Jonathan Stokes, a postdoc at MIT and the Broad Institute of MIT and Harvard
Tommi Jaakkola, Professor
Note, this research represents a team effort including a number of students.
Call to Action: This MIT breakthrough represents an important milestone in the war against antibiotic resistant bacteria around the world. Those interested in this field should keep an eye on this MIT-based group.