Machine learning identifies powerful antibiotic

For years, medicine has struggled to treat antibiotic-resistant bacteria, like MRSA, which I wrote about in my February 12^th article. However, machine learning could potentially revolutionize—not to mention speed up—the way in which researchers find successful antibiotic candidates. Experts from Massachusetts Institute of Technology, Harvard University, and other organizations developed a machine learning model built to identify potential antibiotics based on their chemical structure. Their February 2020 article in Cell details the discovery of the halicin molecule’s antibiotic properties. Importantly, halicin was shown to be effective against resistant tuberculosis bacteria.

Halicin is a divergent molecule: its chemical structure and function is different from those of traditional ones (antibiotics in this case). Machine learning’s capacity to identify such divergent molecules presents a strong advantage over human methods, which often begin by finding candidates similar to successful antibiotics. While this method might seem logical, it could limit potential discoveries. Further, bacteria could be less likely to develop resistance to divergent antibiotics.

“Neurone” by Hugh Tomkins is licensed under CC BY-NC-ND 4.0

The machine learning model used to analyze halicin is called a deep neural network. It finds relationships within data using complex mathematics. Equipped with algorithms inspired by the human brain, a deep neural network can be trained with an initial data set—almost like a practice run—to then use the learned patterns to predict information about new data. Halicin’s antibiotic potential was realized after training a neural network to comb through a library of over 6,000 molecules. As one might imagine, machine learning can analyze that volume of data much more quickly than humans.

With such powerful algorithms, machine learning is beginning to meet the urgent need for treating resistant bacteria.  As such, halicin is hopefully just one of many molecules that deep neural networks can identify as antibiotics. The outlook is promising. The same neural network that identified halicin analyzed another chemical library called ZINC15, this time with over 100 million entries. The algorithm identified another eight antibiotic candidates, two of which were shown to have broad-spectrum activity.

“Their approach highlights the power of computer-aided drug discovery. It would be impossible to physically test over 100m compounds for antibiotic activity,” Jacob Durrant of the University of Pittsburgh, a drug design researcher who was not part of the study, tells The Guardian.

In addition to killing tuberculosis-causing bacteria, halicin was shown to fight other resistant bacteria in mice models. For example, halicin strongly inhibited the growth of Acinetobacter baumannii bacteria. The World Health Organization states that this highly contagious bacteria, sometimes caught in hospitals, requires urgent new treatments.

Thanks to the efficiency of computers and cleverness of top researchers, the dangers of antibiotic resistant bacteria might become less likely.

Citation