AI may help make new antibiotics from spider and snake venoms

Using an artificial intelligence tool that predicts the antimicrobial capabilities of peptides, researchers at the University of Pennsylvania were able to screen over 40 million peptides from venom—known as venom-encrypted peptides (VEPs)—in just hours. They discovered that 53 VEPs killed drug-resistant bacteria, including Escherichia coli and Staphylococcus aureus, and most of them were harmless to human red blood cells. Some even outperformed standard antibiotics.

Antimicrobial resistance contributes to around 5 million deaths each year, including an estimated 1.27 million directly caused by resistant infections. Those numbers could double by 2050 without new treatments. New antibiotics targeting gram-negative bacteria could save millions of lives by 2050. Yet since the 1980s, very few truly novel classes of antibiotics have been developed, due to high costs and long timelines.

“Venoms have evolved over millions of years to disrupt biological systems with remarkable potency and specificity,” says Marcelo Der Torossian Torres, one of the lead authors of the new study (Nat. Commun. 2025, DOI: 10.1038/s41467-025-60051-6) who previously redesigned wasp venom peptides into safer synthetic antibiotics. “Unlike traditional antibiotics, some venom peptides disrupt bacterial membranes—a mechanism that bacteria struggle to evade through conventional resistance strategies,” he adds.

The team used a deep learning model called APEX, which identifies new antimicrobial peptides (AMPs) by learning key traits found in known AMPs, such as positive charge, amphiphilicity, hydrophobic moments, and the ability to form helices. This allows the model to discover which peptides have similar properties even if their amino acid sequences do not closely resemble known ones.

“Traditional bioinformatics relies heavily on sequence similarity, but deep learning models can identify functional patterns beyond simple homology,” says Torres. APEX predicts how well a peptide can fight bacteria by estimating its minimum inhibitory concentration (MIC)—the concentration of the peptide needed to stop the growth of up to 34 bacterial strains. Peptides with a lower MIC are more effective.

The researchers sourced 16,123 venom proteins from venomous creatures—including snakes, insects, scorpions, and spiders—from four online databases, and then they computationally truncated them to generate 40,626,260 VEPs. They filtered candidates with MIC ≤ 32 µM and excluded those similar to known AMPs, ensuring that the 386 selected candidates represented structurally novel peptides.

Of those, the researchers manually selected 58 for synthesis based on APEX’s prediction of the peptides’ novelty, antimicrobial potency, and diverse animal origins. In vitro testing of the candidates showed that 53 were active against at least one pathogenic strain. Of the 28 peptides tested against Pseudomonas aeruginosa, 26 depolarized the cytoplasmic membrane more effectively than the existing antibiotics polymyxin B and levofloxacin.

Assessing the peptides’ toxicity to human cells showed that they were generally less toxic to human red blood cells than to human embryonic kidney (HEK293T) cells. The top three VEPs with low toxicity—UniProtKB-7, ConoServer-14, and Arachnoserver-5—were then tested on a skin-abscess mouse model infected with Acinetobacter baumannii. A single topical dose of each VEP at its MIC significantly reduced bacterial counts 2 days postinfection with no signs of toxicity.

“From a drug development standpoint, the VEPs’ potencies are low—some showing MICs in the double-digit micromolar range, while most drug leads aim for nanomolar potency,” says Steve Trim, chief scientific officer of Ventera Bio, a company that discovers and develops peptide-based bioinsecticides from venoms. “But if [the team members] model stability and bioavailability in the tissue environment, they’ll be one step closer to making venom peptides viable antibiotic drugs one day,” he adds.

The team now wants to test a wider range of pathogens, evaluate different infection models, and chemically modify the top peptides to improve stability, reduce cytotoxicity, and enhance pharmacokinetics.