Antibiotic resistance is a major global health threat, but very few new antibiotics have been developed over recent decades, and most of those newly approved antibiotics are variants of existing drugs. Current methods for screening new antibiotics are also often prohibitively costly, require a significant time investment, and are commonly limited to a narrow spectrum of chemical diversity.
The idea of using predictive computer models for in silico screening is not new, but until now, these models were not sufficiently accurate to transform drug discovery. Previously, molecules were represented as vectors reflecting the presence or absence of certain chemical groups. New neural networks can learn these representations automatically, mapping molecules into continuous vectors that are subsequently used to predict their properties.
To try to find completely novel antibiotic compounds, James Collins, PhD, the Termeer professor of medical engineering and science in MIT’s Institute for Medical Engineering and Science (IMES) and department of biological engineering teamed up with Regina Barzilay, PhD, Tommi Jaakkola, PhD, and their students Kevin Yang, Kyle Swanson, and Wengong Jin, who have previously developed machine-learning computer models that can be trained to analyze the molecular structures of compounds and correlate them with particular traits, such as the ability to kill bacteria. Barzilay is the Delta Electronics professor of electrical engineering and computer science and Jaakkola is a professor, at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). Barzilay and Collins are faculty co-leads for MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health.
The researchers designed their model to look for chemical features that make molecules effective at killing E. coli. To do so, they trained the model on about 2,500 molecules, including about 1,700 FDA-approved drugs and a set of 800 natural products with diverse structures and a wide range of bioactivities. The resulting machine learning model, which can screen more than a hundred million chemical compounds in a matter of days, is designed to pick out potential antibiotics that kill bacteria using different mechanisms than those of existing drugs. The powerful antibiotic that has been discovered by MIT researchers is reported in Cell, in a paper titled, “A Deep Learning Approach to Antibiotic Discovery.
The trained model was tested on the Broad Institute’s Drug Repurposing Hub, a library of about 6,000 compounds. The model picked out one molecule in particular that was predicted to have strong antibacterial activity and had a chemical structure that was different from any existing antibiotics. Using a separate machine-learning model, the researchers also showed that this molecule would likely have low toxicity to human cells.
The newly identified molecule, a c-Jun N-terminal kinase inhibitor, SU3327—which the team renamed halicin—had previously been investigated as a possible diabetes drug. Tests showed that halicin effectively killed many of the dozens of lab-grown, patient-derived bacterial strains against which it was tested, including drug-resistant Clostridium difficile, Acinetobacter baumannii, and Mycobacterium tuberculosis. The drug worked against every species tested, with the exception of Pseudomonas aeruginosa, a difficult-to-treat lung pathogen.
Experiments also showed that E. coli did not develop any resistance to halicin during a 30-day treatment period. In contrast, the bacteria started to develop resistance to the antibiotic ciprofloxacin within one to three days, and after 30 days, the bacteria were about 200 times more resistant to ciprofloxacin than they were at the beginning of the experiment.
Preliminary analyses suggested that halicin kills bacteria by disrupting their ability to maintain an electrochemical gradient across their cell membranes. This gradient is necessary, among other functions, to produce ATP, so if the gradient is disrupted, the cells die. The researchers suggested that it could be difficult for bacteria to develop resistance to drugs that target this mechanism.
The team also used their computational model to screen more than 100 million molecules selected from the ZINC15 database, an online collection of about 1.5 billion chemical compounds. The screen, which took three days, identified 23 candidates that were structurally dissimilar to existing antibiotics and were also predicted to be nontoxic to human cells. The researchers found that eight of these molecules showed antibacterial activity in tests against five species, with two demonstrating particular efficacy. “Additionally, from a discrete set of 23 empirically tested predictions from >107 million molecules curated from the ZINC15 database, our model identified eight antibacterial compounds that are structurally distant from known antibiotics,” the investigators noted. They also plan to test these molecules further, and to screen more of the ZINC15 database.
As well as identifying antibiotic activity in existing compounds, the team aims to use the model to help design new antibiotics and to optimize existing molecules, based on what it has learned about chemical structures that enable drugs to kill bacteria. The potential exists to train the model to add features that would make a particular antibiotic target only certain pathogenic bacteria, and so prevent it from killing beneficial bacteria in a patient’s digestive tract. The investigators indicated that “the time is ripe” for the application of machine learning approaches to antibiotic discovery, and potentially help to outpace the spread of antibiotic resistance.
Stokes JM, Yang K, Swanson K, Jin W, Cubillos-Ruiz A, Donghia NM, MacNair CR, French S, Carfrae LA, Bloom-Ackerman Z, Tran VM, Chiappino-Pepe A, Badran AH, Andrews IW, Chory EJ, Church GM, Brown ED, Jaakkola TS, Barzilay R, Collins JJ. A Deep Learning Approach to Antibiotic Discovery. Cell. 2020 Feb 20;180(4):688-702.e13. doi: 10.1016/j.cell.2020.01.021.Go To Cell