The antimicrobial resistance crisis has a well-known bottleneck: the discovery pipeline. Traditional screening methods are slow, expensive, and have produced few viable candidates for clinical use. A team at the University of California, Davis has now demonstrated a generative AI approach that may accelerate the process dramatically.
ARCADIAMP, Antimicrobial Rapid Candidate Generation through AI-driven Diffusion and Iterative Assessment of Membrane-active Peptides, combines a discrete denoising diffusion probability model (D3PM) with a two-stage ESM2-based classifier, a novelty filter, and an iterative self-training loop. From a pool of 1 million generated candidates, the platform filtered down to just 10 peptides for experimental synthesis. Eight of the 10 showed antimicrobial activity with a minimum inhibitory concentration (MIC) of 32 µg/mL or below, an 80 percent hit rate that vastly exceeds conventional screening.
The lead candidate, designated Peptide-7 and named Arcinin, is a 29-amino-acid alpha-helical peptide with a sequence (GRWRRVGRKLRTLGKSFGKVAHVAGKAIFA) that shows no significant similarity to any known antimicrobial peptide.
Potent against five of six ESKAPE pathogens
Arcinin was tested against the ESKAPE panel, the six bacterial species responsible for the majority of hospital-acquired infections worldwide and known for their multidrug resistance. The results:
| Pathogen | MIC (µg/mL) | Category |
|———-|————-|———-|
| Klebsiella pneumoniae | 8 | Gram-negative, ESKAPE |
| Acinetobacter baumannii | 8 | Gram-negative, ESKAPE |
| Pseudomonas aeruginosa | 8 | Gram-negative, ESKAPE |
| Escherichia coli | 8 | Gram-negative substitute for Enterobacter |
| Staphylococcus aureus | 16 | Gram-positive, ESKAPE |
| Enterococcus faecalis | 64 | Gram-positive, ESKAPE (weak) |
At concentrations of 8–16 µg/mL, Arcinin matches the potency of pexiganan, a clinical-stage antimicrobial peptide, against five of the six species. Against E. faecalis, activity was weaker at 64 µg/mL.
What sets Arcinin apart is its safety profile. Against human red blood cells, the LC₅₀ (the concentration that kills half the cells) was greater than 512 µg/mL, a safety margin comparable to pexiganan and dramatically better than melittin (bee venom), which has an LC₅₀ of just 5.32 µg/mL against HEK293 cells. Arcinin’s HEK293 LC₅₀ was 50.84 µg/mL, roughly ten times less toxic than melittin and comparable to pexiganan’s 37.75 µg/mL.
Mechanism and in vivo results
Arcinin kills bacteria through membrane disruption, inserting into the lipid bilayer within sub-microseconds and forming discrete lesions that cause catastrophic loss of internal homeostasis. Time-kill kinetics show complete eradication at 2× MIC within 30–60 minutes. Electron microscopy reveals different effects by species: S. aureus cells deflate and show abnormal septa; E. coli cells lose their rod shape and become amorphous; K. pneumoniae shows subtle surface wrinkling.
In a murine full-thickness wound infection model at 5 mg/kg topical treatment, Arcinin achieved a 4.56-log reduction in S. aureus CFU (colony-forming units) and a 4.47-log reduction in E. coli CFU, both with p < 0.0001. This corresponds to a greater than 99.99 percent reduction in bacterial burden. Wound closure on Day 6 was approximately 71 percent for S. aureus-infected wounds (vs. 45 percent for PBS control, p = 0.0006) and 66 percent for E. coli-infected wounds (vs. 34 percent, p = 0.0024). Histopathology on Day 8 showed enhanced re-epithelialization and reduced inflammatory infiltration.
Serum stability testing showed that Arcinin maintained an MIC of 32 µg/mL in 50 percent bovine serum for four ESKAPE species, comparable to pexiganan.
How ARCADIAMP works
The platform’s core innovation is its iterative learning design. The generative model, a D3PM trained from scratch (not pretrained), generated an initial candidate pool, which passed through a two-stage classifier. The first stage distinguishes AMPs from non-AMPs (F1 score: 0.86). The second stage identifies strong AMPs with MIC below 8 µg/mL (F1: 0.68). A novelty filter using BLOSUM62 alignment excluded candidates with similarity above 0.45 or identity above 0.65 to any of 27,636 known antimicrobial peptides.
A sample-weighting mechanism based on the therapeutic index, log₁₀(LC₅₀) minus log₁₀(MIC), co-optimized activity and toxicity simultaneously.
The platform then retrained the generative model on its own high-scoring output. Before this augmentation, the median predicted MIC was 495.3 µg/mL. After augmentation, it dropped to 37.8 µg/mL (p < 2.2 × 10⁻¹⁶). Novelty improved from a similarity of 0.63 to 0.49 (p < 2.2 × 10⁻¹⁶).
The translation challenge
The paper, published in Nature Communications on 7 July 2026 (DOI: 10.1038/s41467-026-75030-8), notes that no new antimicrobial peptide has been approved for clinical use since the introduction of the class. In vivo validation was limited to a topical wound model; systemic infection (sepsis) models have yet to be tested. The ARCADIAMP code is available under Apache 2.0 license on GitHub.
Still, the 80-percent hit rate from a computational pipeline that co-optimizes activity and toxicity represents a significant step toward solving the antimicrobial discovery problem. For the 4.95 million deaths associated with antimicrobial resistance each year, approaches like ARCADIAMP offer a path that traditional screening cannot match.
Sources
1. Markakis, K., Kim, S., Tan, C-E. & Tagkopoulos, I., “Discovery of potent low-toxicity antimicrobial peptides through diffusion modeling”, Nature Communications (2026). DOI: 10.1038/s41467-026-75030-8
2. ARCADIAMP GitHub repository: https://github.com/IBPA/ARCADIAMP