Researchers at the University of Pennsylvania used a deep-learning system called APEX to screen a database of over 40 million venom-encrypted peptides (VEPs), proteins employed by animals for attack or defense, in search of potential antibiotics. In just hours, APEX flagged 368 compounds as potential antibiotics. The team published a study in Nature Communications.
From Adobe Stock
From APEX’s list, the team synthesized 58 peptides for testing. 53 of these killed drug-resistant bacteria at doses that were harmless to human red blood cells. This could have significant implications for antibiotic resistance, a growing concern.
Fighting resistance
According to the CDC, antimicrobial resistance was associated with almost 5 million deaths worldwide in 2019. In the U.S., more than 2.8 million antimicrobial-resistant infections occur each year, and more than 35,000 people die as a result. Despite this, traditional antibiotic discovery has plateaued due to high costs and long timelines, according to the team’s study.
Venom-derived peptides offer several advantages over conventional antibiotics that could help solve this problem. The peptides work by disrupting bacterial membranes, a mechanism that bacteria cannot evade through conventional resistance mechanisms. The peptides also exhibit broad-spectrum activity against both gram-positive and gram-negative bacteria, making them ideal for combating multidrug-resistant bacteria.
VEPs have a flexible structure which is able to be engineered for improved stability and selectivity. The top candidates for antibiotics have high net charge and elevated hydrophobicity, both of which are conducive to the disruption of the bacterial membrane.
Applying AI
APEX is a bacterial strain-specific antimicrobial activity predictor based on PyTorch and freely available on GitLab; it is a multiple-target tasks model that can predict minimum inhibitory concentration values of peptides against 34 bacterial strains. Some of the files on GitLab are two years old, indicating that this research has been growing for a while. The system was trained on a peptide dataset and publicly available antimicrobial peptides from DBAASP.
The platform mapped more than 2,000 entirely new antibacterial motifs, short sequences of amino acids within a protein or peptide that are responsible for the protein’s antibacterial activity.
“By pairing computational triage with traditional lab experimentation, we delivered one of the most comprehensive investigations of venom-derived antibiotics to date,” said co-author Marcelo Torres, PhD, a research associate at Penn.
The team is now taking the best peptide candidates and improving them to possibly create new antibiotics. They believe that venoms are a rich source of hidden antimicrobial scaffolds, and that large-scale computational mining can accelerate the discovery of antibiotics.
