Machine learning predicts possible antimicrobial peptides

Researchers say they have come up with two new antibiotic peptide candidates using a combination of machine learning, molecular dynamics, and experiments(Nat. Biomed. Eng. 2021, DOI: 10.1038/s41551-021-00689-x).

Payel Das of IBM Research and colleagues trained a machine learning method called a deep generative model on a database of all known peptides to create 90,000 peptide sequences the computer predicted were plausible antimicrobials and mapped them in virtual space according to their properties. Das says that let them more efficiently tackle the next step, when they used machine learning to screen for peptides with predicted antimicrobial activity and low human toxicity. A second screening of 163 peptides used molecular dynamics simulations; Das describes this step as a way to add the human expertise built into those techniques to their search. The team made the resulting 20 peptides and tested them in lab experiments and in mice. Two candidates emerged that the group says are effective against a range of pathogens—including a multidrug-resistant pneumonia strain—show little ability to induce drug resistance in Escherichia coli, and have low toxicity.

Experts in developing new antibiotics—a process that has stalled for both scientific and economic reasons—say they are skeptical these peptides will become drugs because they may be too toxic in people. Connor Coley, a machine learning expert at MIT, praises the study’s efficient screening method and efforts at experimental validation.

Das says her group will keep pursuing antimicrobials. And she sees potential to use the same approach to search for other drugs and new materials.