AMP-Diffusion

Antimicrobial peptides are short, usually cationic sequences that kill bacteria by disrupting membranes, and they are one of the few antibiotic classes to which resistance evolves slowly. The design problem is combinatorial: a 25-residue peptide has more possible sequences than there are atoms in the observable universe, and the mapping from sequence to potency and toxicity is not smooth. Screening approaches sample a vanishing fraction of that space.

AMP-Diffusion applies denoising diffusion to the problem, but in an unusual place. Rather than diffuse over discrete amino acid tokens or over 3D coordinates, it runs the forward and reverse processes in the continuous per-residue embedding space of ESM-2, then decodes the denoised embeddings back to sequence. The denoising network reuses ESM-2's own attention layers, initialised from the pretrained weights, so the generator inherits the language model's sense of what a plausible protein looks like.

The model was introduced by the Chatterjee lab at Duke University in early 2024 and presented at the GenBio workshop at NeurIPS. Its experimental payoff came from a collaboration with the de la Fuente-Nunez lab at the University of Pennsylvania, which generated 50,000 candidates, ranked them with the APEX minimum-inhibitory-concentration predictor, synthesised 46, and carried the leads through to mouse infection models.

#Key Features

• Latent-space diffusion: The diffusion process operates on continuous protein language model embeddings, sidestepping the discrete-token machinery that sequence diffusion models otherwise require.
• Attention weight reuse: The denoiser is not trained from scratch — its transformer blocks are initialised from ESM-2's attention layers, transferring evolutionary sequence priors into the generative model.
• Length-flexible sampling: Candidates can be drawn at a fixed length or across a mixed 10–39 residue range, matching the size distribution of natural antimicrobial peptides.
• Composition shift toward activity: Sampled peptides are enriched in lysine (+4.4 percentage points over the training distribution) and leucine (+2.7), the cationic and hydrophobic residues that drive membrane disruption.
• End-to-end validated pipeline: Generation, in silico ranking, chemical synthesis, in vitro assays, mechanistic studies, and in vivo efficacy testing were carried through on the same design campaign.

#Technical Details

The public checkpoint has 16.54M parameters — 7.40M inherited from ESM-2 8M layers plus 9.14M of its own — and was trained on 19,670 antimicrobial peptide sequences. Two 50,000-sequence generation runs are benchmarked, one at a fixed 25 residues and one at mixed lengths. Against AmpGAN, HydrAMP, and PepCVAE scored through an identical pipeline, AMP-Diffusion attains the lowest mean ProGen2-medium perplexity (15.81, versus 17.84 for the training set itself and 18.51–20.25 for the baselines), the highest HydrAMP classifier P(AMP) at 0.905, and the highest fraction of sequences predicted below a 128 µM MIC threshold at 94.6% against 76.2% for the next-best generator. Jensen-Shannon divergence of amino acid composition from the training distribution is 0.129, and all 50,000 sampled sequences are unique.

In the experimental campaign, 46 candidates selected for predicted activity, diversity, and novelty were synthesised. The resulting peptides showed broad-spectrum activity against clinically relevant pathogens including multidrug-resistant strains, killed bacteria via membrane permeabilisation and depolarisation, and showed low cytotoxicity in human cell assays. In preclinical mouse infection models, lead peptides reduced bacterial burden with efficacy comparable to polymyxin B and levofloxacin and no detectable adverse effects.

#Applications

The model is a candidate generator for antibiotic discovery programmes facing drug-resistant Gram-negative and Gram-positive pathogens, and the same machinery applies to other short bioactive peptide classes where a training corpus of active sequences exists. In practice it is used upstream of a scoring model: sample broadly, rank with an activity predictor, and synthesise the top of the list, which is the workflow the Penn campaign followed. Because sampling is unconditional, target specificity comes from the filtering stage rather than from conditioning the generator.

#Impact

AMP-Diffusion is one of the earlier demonstrations that continuous latent diffusion over protein language model embeddings is a workable generative recipe for peptides, and the Penn collaboration is among the more complete validation chains published for a generative AMP designer — from sampled sequence to in vivo efficacy. The authors are candid that the modelling work dates to early 2023, before diffusion and flow matching became standard in the biological domain, and note that newer methods may supersede its training approach. Practical caveats matter for reuse: the repository is a demonstration of the architecture plus one small checkpoint rather than a full training stack, and although its README states MIT terms, no LICENSE file is present in the repository.