A deep learning model trained to predict antibiotics based on structure has identified a powerful new antibiotic compound, Halicin, according to study results published in Cell.
Few new antibiotics have been developed over the past few decades and the majority of newly approved antibiotics are structural variations of existing drugs. In addition, current methods for screening new antibiotics are often prohibitively costly, require a significant time investment, and are usually limited to a narrow spectrum of chemical diversity.
Accordingly, the study investigators developed a machine-learning computer model to scan for chemical features that make antibiotics effective at killing Escherichia coli. The deep neural network model was trained on 2335 molecules, including 1760 US Food and Drug Administration-approved drugs and a set of 800 natural products with diverse structures and a wide range of bioactivities.
Continue Reading
Once the deep learning model was trained, it was tested on the Broad Institute’s Drug Repurposing Hub, a library of approximately 6000 compounds. The model identified the molecule, Halicin, which was predicted to have strong antibacterial activity and a chemical structure that was distinct from any existing antibiotics. Halicin was tested against various bacterial strains isolated from patients. The investigators discovered that Halicin was able to kill many strains of bacteria that are resistant to treatment, including Clostridium difficile, Acinetobacter baumannii, and Mycobacterium tuberculosis. The drug was effective against every species that they tested, except Pseudomonas aeruginosa, a difficult-to-treat lung pathogen.
In addition, the model was used to screen >107 million molecules selected from the ZINC15 database, an online collection of approximately 1.5 billion chemical compounds. This screen identified 23 candidates that were structurally dissimilar from existing antibiotics and were predicted to be nontoxic to human cells. In laboratory tests against 5 species of bacteria, 8 of the molecules showed antibacterial activity, and 2 were particularly powerful.
Study limitations included the lack of phenotypic screening conditions that enrich for molecules against specific biologic targets, and the fact that the composition of the training data was small and lacked diversity.
The researchers plan to use their deep learning model at all stages of antibiotic development from discovery to improvements in efficacy and toxicity through drug modifications and medicinal chemistry.
Reference
Stokes JM, Yang K, Swanson K, Jin W, Cubillos-Ruiz A, Donghia NM, et al. A deep learning approach to antibiotic discovery. Cell. 2020;180:688-702.
