OmegAMP

Whether a short cationic peptide kills bacteria is governed largely by three quantities computable straight from its sequence: net charge, which drives electrostatic attachment to the anionic bacterial surface; mean hydrophobicity, which sets how readily the peptide inserts into the lipid bilayer; and length, which constrains secondary structure and synthesis feasibility. Generative models for antimicrobial peptide (AMP) design had conditioned instead on a coarse active/inactive label, leaving those three properties to post-hoc filtering. A researcher who needed a +6-charge, 20-residue design had to sample broadly and hope some outputs landed in range.

OmegAMP is a conditional denoising diffusion model that moves those properties into the sampling loop itself. It was developed by Helmholtz Munich with the Technical University of Munich, the University of Pennsylvania and the University of Warsaw, and introduced in 2025 as a methods and benchmarking paper, with a second preprint in 2026 reporting the wet-lab and animal work.

The framework separates how generation starts from what it is asked to hit. Three generation modes control the diffusion trajectory — de novo from pure Gaussian noise, analog generation from a partially noised prototype, and motif-guided generation in which designated residues stay fixed throughout denoising. Three conditioning strategies populate the property target independently: unconditional, targeted with explicit values or intervals, and prototype-derived from a known peptide. Modes and strategies combine freely, which is what lets a designer fix a functional motif while simultaneously steering charge and hydrophobicity.

#Key Features

• Physicochemical embedding: Each residue maps to a six-dimensional vector — hydrophobicity, isoelectric point, α-helix propensity, transmembrane tendency, an amino-acid selectivity index, and a padding flag. The five scales are jointly injective, so denoised embeddings decode back to sequence by nearest-neighbour distance.
• Masked conditioning at training time: Charge, length and hydrophobicity entries are randomly masked during training, so at inference any subset can be fixed while the rest float.
• Analog optimization with two dials: Exploration strength controls how far the sequence may drift from its prototype; a separate relaxation parameter controls how far the property profile may deviate.
• Motif preservation: Designated positions are held fixed while the surrounding sequence is optimized, which preserves auxiliary functions such as lipopolysaccharide engagement or DNA binding.
• Prototype as a design lever: Because analogs retain sequence identity with their prototype, they inherit properties never optimized for directly — the membrane-disruption mode and the mammalian-cell safety profile both track the prototype family.

#Technical Details

The denoiser is a one-dimensional U-Net with skip connections and attention layers, operating on a 6 × 100 embedding tensor (100 residues being the maximum length). It is trained as a denoising diffusion probabilistic model with a cosine variance schedule and a v-prediction objective. Training used 810,283 peptides: 35,878 curated AMPs aggregated from the AMPScanner v2 training set, dbAMP 2.0 and DRAMP 3.0, plus 774,405 functional peptides drawn from Peptipedia and UniProtKB to teach the distribution of chemically plausible non-antimicrobial sequences. Conditioning accuracy — the fraction of samples falling inside a target range for charge 2–10, length 5–30 and hydrophobicity −0.5 to +0.8 — exceeded 90% under targeted conditioning across single, pairwise and three-property constraints, where baseline generators fell to 70–80% and degraded as constraints were combined. Of 204 designed peptides synthesized and characterized, 95 came from de novo generation, of which 90 reached a minimum inhibitory concentration at or below 4 µM against at least one strain of a 20-strain panel. Across the 11 strains also reported for AMP-Diffusion, 45 of OmegAMP's 95 de novo candidates reached MIC ≤ 2 µM against E. coli, compared with 8 of AMP-Diffusion's 35.

#Applications

The three modes map onto distinct medicinal-chemistry tasks. De novo generation supplies fresh leads against multidrug-resistant Gram-negative pathogens; analog generation performs lead optimization, converting six experimentally inactive prototypes into antimicrobials while holding at least 60% sequence identity; motif-guided generation installs antimicrobial activity onto a scaffold chosen for a different function, demonstrated by redesigning a non-antimicrobial bZIP leucine zipper into a peptide that kills bacteria while retaining DNA-perturbing character.

#Impact

Selected leads were carried into murine skin scarification and neutropenic thigh infection models of Acinetobacter baumannii. The motif-guided design Ω-MT-bZIP-8 reduced thigh bacterial burden by 3.4 log₁₀ by day eight, statistically indistinguishable from the levofloxacin control. The framework's reach is bounded in ways its authors state plainly: it conditions on global properties but not on local features such as residue spacing or hydrophobic-face geometry, it optimizes for activity without conditioning on toxicity, the strain panel omits slow-growing and intracellular pathogens, and the animal work covers only skin and thigh infection. The wet-lab results remain a preprint awaiting peer review. Code is MIT-licensed, with a v1.0.0 source snapshot archived on Zenodo, while the trained generative checkpoint is fetched from a Google Drive link given in the README rather than from a versioned archive — so the conditioning machinery is reusable for other peptide classes whose function is governed by sequence-derived physicochemical properties.