Artificial Intelligence is able to discover a novel antibiotic

This topic fits the theme of gene-environment interactions: the environmental signal is an antibiotic, and the microbial genomic response by the microbe to that signal can be either “sensitivity” (tocixity or death) or “resistance” (therapeutic failure). Historically, antibiotics have been discovered largely through screening soil-dwelling or marine microbes for secondary metabolites that prevent growth of pathogenic bacteria. This approach has resulted in the majority of clinically used classes of antibiotics — including b-lactams, aminoglycosides, polymyxins, glycopeptides, and others. Semi-synthetic derivatives of these scaffolds have maintained a viable clinical arsenal of antibiotics by increasing potency, decreasing toxicity, and avoiding determinants of microbial resistance. Entirely synthetic antibiotics — of the pyrimidine, quinolone, oxazolidinone, and sulfa classes — have also found prolonged clinical utility, and new derivatives continue to be optimized for the same properties.

Unfortunately, discovery of new antibiotics is becoming increasingly difficult. Natural product discovery is now plagued

by the “dereplication” problem (i.e. the same molecules being repeatedly ‘discovered’). Furthermore, given the rapid expansion of chemical spaces that are accessible by the derivatization of complex scaffolds, engineering next-generation versions of existing antibiotics results in substantially more failures than successes. Therefore, many antibiotic discovery programs have turned to screening a few million molecules, are often prohibitively costly to curate, limited in chemical diversity, and failure to reflect the chemistry that is inherent to antibiotic molecules. Since implementation of high-throughput screening in the 1980s, no new clinical “breakthough” antibiotics have been discovered using this method.

Novel approaches to antibiotic discovery are therefore needed to increase the rate at which new antibiotics are identified (and simultaneously decrease the associated cost of early lead discovery). Given recent advancements in machine learning large synthetic chemical libraries, the field is now ripe for the application of algorithmic solutions for molecular property prediction to identify novel structural classes of antibiotics (these methods are made largely in silico). To address this challenge, authors [see attached article] trained a “deep neural network” capable of predicting molecules having “antibacterial activity”. Authors performed predictions on multiple chemical libraries and discovered a molecule from the Drug Repurposing Hub — halicin — that is structurally divergent from conventional antibiotics and displays bactericidal (killing) activity against a wide phylogenetic spectrum of pathogens including Mycobacterium tuberculosis and carbapenem-resistant Enterobacteriaceae.

Halicin also was effectively bactericidal against Clostridioides difficile and pan-resistant Acinetobacter baumannii infections in mouse models. In addition, from a discrete set of 23 empirically tested predictions from >1013 (>1,000,000,000,000) molecules curated from the ZINC15 database [this is a curated collection of commercially available chemical compounds, developed by John Irwin in the Shoichet Laboratory in the Dept Pharmaceutical Chem at the Univ of California, San Francisco (UCSF); the latest release of the website interface is “ZINC15” (2015)] — the authors’ model identified eight antibacterial compounds that are structurally distant from known antibiotics. This work highlights the utility of deep learning approaches to expand our antibiotic arsenal through discovery of structurally distinct antibacterial molecules.


Cell 20 Feb 2020; 180: 688-702

This entry was posted in Center for Environmental Genetics. Bookmark the permalink.