César de la Fuente's laboratory at the University of Pennsylvania screened 40,626,260 venom-encrypted peptide sequences through a deep-learning model called APEX and reduced that space to 386 candidates in a matter of hours. Researchers then synthesized 58 of those candidates and tested each against drug-resistant bacterial strains. Fifty-three inhibited at least one pathogen at doses that left human red blood cells intact, ScienceBlog's analysis of the study reported. The findings appeared in Nature Communications on August 26, 2026, under a collaboration between Penn's School of Engineering and Applied Science and the Perelman School of Medicine.

The 53 confirmed hits are in vitro observations. No candidate has entered animal toxicity profiling, pharmacokinetic testing, or a regulatory submission pathway. The 91 percent apparent hit rate applies exclusively to the 58 hand-selected peptides chosen after potency prediction and novelty filtering, not to the full 40.6 million sequence space. APEX has not been shown to classify every active and inactive peptide across that library with equivalent accuracy.

From Four Databases to 58 Synthesized Molecules

The 40.6 million sequences were generated computationally from 16,123 venom proteins assembled across four specialist databases covering snakes, spiders, scorpions, cone snails, and sea anemones. APEX predicted minimum inhibitory concentrations against 34 bacterial strains for each fragment. Sequences projected to inhibit growth at 32 micromoles per liter or lower passed into a similarity filter that excluded peptides too close to known antimicrobial compounds, The Scientist's coverage of the screening pipeline detailed. The 386 survivors of that filter constituted the candidate shortlist.

Practical constraints then governed bench selection. Researchers chose 58 peptides for custom peptide synthesis and assayed them against resistant pathogens including Escherichia coli, Staphylococcus aureus, Acinetobacter baumannii, and Pseudomonas aeruginosa. The peptides acted primarily through membrane depolarization—disrupting the electrical gradient across bacterial cell walls—rather than through permeabilization. Changge Guan, a postdoctoral researcher in the De la Fuente Lab, reported that the platform identified more than 2,000 entirely new antibacterial motifs, short amino acid sequences responsible for killing or inhibiting bacterial growth, Penn Medicine's summary of the publication confirmed.

Some of the 53 active compounds outperformed standard antibiotics at the MIC level in controlled bench conditions, Chemical & Engineering News noted in its brief. That comparison holds only for defined strain panels in microplate assays. It does not account for bioavailability, immune clearance, or serum stability in a living organism.

Evidence Stage and Regulatory Distance

Antibiotic resistance directly kills more than a million people annually, and discovery pipelines have produced few new structural classes in recent decades, ScienceBlog's reporting established. That deficit explains the regulatory attention now directed at computational triage of natural peptide libraries. The European Medicines Agency and the U.S. FDA published joint guiding principles for AI in medicine development in January 2026, covering evidence generation from early research through manufacturing. Neither agency, however, currently accepts a peptide identified solely through in silico screening without standard pharmacokinetic and toxicological data packages.

The red-blood-cell hemolysis assay the Penn team performed is a conventional early safety screen. It does not substitute for multi-organ toxicity studies, serum half-life measurements, or distribution profiling required before any first-in-human protocol. Venom peptides evolved to breach biological defenses in prey and predators; the same membrane-disrupting properties that kill bacteria can damage mammalian tissue at different concentrations or exposure routes. The 53 active sequences join a wider category of bioactive peptides under investigation internationally, from gut-microbiome-derived fragments to prion-associated candidates. A successor model, APEX 1.1, has already been applied to 19.3 million fragments from prion and prion-like proteins, yielding 1,179 additional candidates, two of which reduced A. baumannii burden in a mouse skin-infection model at levels comparable to polymyxin B. That in vivo result extends beyond the venom study's in vitro data but remains preclinical.

Marcelo Torres, a research associate at Penn and co-author on the study, stated that the team is now applying medicinal-chemistry adjustments to the top peptide candidates—modifying side chains and backbone structures to improve stability and reduce off-target effects. Those optimizations will determine whether any of the 53 laboratory-active molecules can advance beyond a promising MIC value toward a compound that a regulatory reviewer would accept into a Phase I protocol.

Related Peptides Agora coverage examines cell penetrating peptides.