2026-09-16
OmegAMP conditions AMP design on charge, hydrophobicity and length. 82/95 de novo peptides hit MIC ≤ 2 μM; a motif analog matched levofloxacin in mice.
Antimicrobial peptides (AMPs) stick to anionic bacterial membranes through net charge, insert through hydrophobicity, and fold with enough length to form secondary structure. Most generative AMP models either condition on a coarse active/inactive label or filter after sampling. Neither pins charge, mean hydrophobicity, and length, and neither can freeze a short motif that binds LPS or perturbs DNA. Analog generators in the HydrAMP family also cannot hold an analog's physicochemical profile relative to its prototype.
OmegAMP treats those knobs as inputs to one conditional diffusion model.
Each residue becomes a 6-D vector: hydrophobicity, isoelectric point, helix propensity, transmembrane tendency, selectivity, plus a padding bit. A separate conditioning vector carries target activity, charge, length, and mean hydrophobicity. Training noises peptide embeddings and trains a denoiser to reverse the process given the condition. At inference the model denoises an initial noisy embedding into a sequence.
Three trajectories:
Conditions can be blank (sample the learned AMP distribution), filled with explicit ranges, or copied from a prototype and relaxed with σ. Random masking during training lets any subset of properties be specified at inference. Analog and motif modes compose.
The wet-lab set is 215 peptides: 95 de novo, 59 analogs of six inactive prototypes, 50 motif-preserving designs, plus 11 prototypes and polymyxin B / levofloxacin. Leads were picked on strain coverage at MIC ≤ 2 μM, median MIC₅₀, hemolysis HC₅₀, and HEK293 CC₅₀.
Of 95 de novo peptides, 90 (95%) reached MIC ≤ 4 μM on at least one strain and 82 (86%) reached MIC ≤ 2 μM. On E. coli at MIC ≤ 2 μM the count is 45/95, versus 8/35 for AMP-Diffusion, 0/12 for Joker, and 0/21 for CLaSS. Ω-DP-52 hit 18 of 20 strains including MRSA and VRE, with HC₅₀ > 128 μM but CC₅₀ = 2.0 μM. Ω-DP-19 covered all five drug-resistant Gram-negatives, HC₅₀ > 128 μM, CC₅₀ = 19.8 μM. Gram-positive coverage is thin: B. subtilis at 81% of designs, most pathogens largely resistant.
Among 59 analogs held at ≥60% identity, 50/59 were active at MIC ≤ 2 μM on at least one strain (5/10 to 10/10 by family). Prototype family explained most variance in outer-membrane permeabilization, hemolysis, and cytotoxicity. Outer-membrane-dominant peptides were substantially less toxic to mammalian cells.
All 40 LPS-engaging motif analogs hit MIC ≤ 2 μM (39 on an MDR isolate). Seven of ten redesigned leucine-zipper (bZIP) peptides gained antibacterial activity. In a murine skin model, Ω-AMT-cecropin-4 and cecropin-1 cut A. baumannii burden by 3.4 and 3.2 log₁₀ at day 2, versus 3.0 for polymyxin B. In a neutropenic thigh model at day 8, proteolytically stabler Ω-MT-bZIP-8 matched levofloxacin at 3.4 log₁₀; Ω-AMT-pa4-1 reached 2.4 log₁₀.
Targeted conditioning (OmegAMP-DT) exceeded 90% success even when charge, length, and hydrophobicity were constrained together. Unconditional sampling sat around 70–80%.
This is a usable increment, not a new antibiotic class. Physicochemical knobs become inputs, motifs stay put, and analogs inherit membrane mechanism and safety that were never written into the condition. A practical workflow is to screen prototypes for outer-membrane-dominant disruption, then expand analogs on a charge-hydrophobicity grid. Two in vivo models include standard-of-care comparators.
Toxicity is still measured after the fact. Gram-positive coverage remains the weak side.
Analog claims sit above a 60% identity floor; the 30% bZIP case is anecdotal. Membrane assays were taken at MIC against one strain, A. baumannii ATCC 19606, and do not cover Gram-positives that lack an LPS outer membrane. The conditioner omits residue spacing and hydrophobic-face geometry, which CD shows also shape potency. The 20-strain panel skips slow-growing and intracellular pathogens; animals are skin and thigh only. Toxicity, proteolytic stability, and resistance propensity are not yet conditioning targets.