“Generative latent diffusion language modeling yields anti-infective synthetic peptides”
Today, we share a research article led by Cesar de la Fuente-Nunez's team, published in Cell Biomaterials. This study developed a generative artificial intelligence platform named AMP-Diffusion, which enables the de novodesign of antimicrobial peptides (AMPs) by integrating latent diffusion models and protein language models (pLMs). This work generates functional peptides directly from the ESM-2 embedding space without requiring predefined motifs or structural priors. Experimental validation demonstrated that 76% of the generated peptides exhibit broad-spectrum antimicrobial activity (including against multidrug-resistant bacteria), with in vivoefficacy comparable to standard antibiotics, providing a scalable, rational design tool to address the antibiotic resistance crisis.
01 Research Background
The accelerating spread of antimicrobial resistance (AMR) has become a global health crisis, and the traditional antibiotic discovery pipeline struggles to meet clinical needs. Antimicrobial peptides (AMPs) are promising alternatives due to their multiple mechanisms of action (e.g., membrane disruption, immunomodulation) and lower propensity to induce resistance. However, the rational design of AMPs faces significant challenges: peptide sequences consist of 20 natural amino acids, resulting in a sequence space of 20^L for a peptide of length L, requiring simultaneous optimization of antimicrobial activity, low toxicity, and biostability. Existing computational methods, such as rule-based design or traditional machine learning, are limited by the discrete and high-dimensional nature of the sequence space, making it difficult to efficiently explore novel structures. The successful application of generative AI (e.g., diffusion models) in protein design offers a new approach to overcome this bottleneck, but adapting it specifically for AMP design and achieving experimental validation remains a challenge to be fully realized.
02 Innovative Highlights
First latent diffusion model for AMPs based on pLM embedding space: AMP-Diffusion operates the diffusion process (Denoising Diffusion Probabilistic Model, DDPM) directly within the continuous embedding space of the pre-trained ESM-2 model, eliminating the need for additional training of a protein latent space. It leverages the biophysical rules already learned by the pLM to ensure the naturalness and functionality of generated peptides.
Denoiser architecture deeply integrated with pLM attention mechanisms: The denoiser reuses the weights of the ESM-2 attention layers, ensuring the generation process strictly adheres to evolutionary constraints captured by the pLM, outperforming models relying solely on statistical generation.
Comprehensive experimental validation and demonstration of in vivoefficacy: 46 peptides were synthesized from 50,000 generated candidates and systematically evaluated for their antimicrobial activity (against 11 pathogens), secondary structure, mechanism of action, cytotoxicity, and efficacy in a murine infection model. The validation rate was significant (76% active, with in vivoefficacy comparable to polymyxin B and levofloxacin).
03 Results and Discussion
3.1 Computational Framework and Characteristics of Generated Peptides
AMP-Diffusion was fine-tuned on 19,670 AMP sequences from the DRAMP, APD3, and DBAASP databases. The distribution of the average minimum inhibitory concentration (MIC) for the generated peptides was highly similar to that of the training set. Furthermore, the perplexity (PPL) calculated via ProGen2 showed no significant difference between generated peptides and natural AMPs (Training set 17.93 vs. Generated set 17.90, p=0.0357), indicating the model successfully captured the sequence distribution of AMPs. The amino acid composition of the generated peptides was consistent with the training set, but the selected subset showed significant enrichment of lysine (K), leucine (L), and arginine (R) – residues known to enhance cationicity and amphipathicity, thereby improving membrane-targeting activity. Physicochemical property analysis further confirmed that the selected peptides possessed higher charge and amphipathicity index (Figure 1).
Figure 1. AMP-Diffusion model and characteristics of the generated peptides.
3.2 In VitroAntimicrobial Activity and Structural Features
Testing of 46 synthesized peptides against 11 clinically relevant pathogens (including ESKAPEE drug-resistant bacteria) showed that 35 peptides (76%) were active against at least one strain, with Acinetobacter baumanniiATCC 19606 and vancomycin-resistant Enterococcus faeciumATCC 700221 being the most susceptible. Circular dichroism (CD) spectroscopy indicated that the peptides adopted α-helical conformations in membrane-mimetic environments (e.g., 60% TFE, SDS micelles), but were random coils in water, consistent with the membrane-induced folding mechanism of typical AMPs. All active peptides (red dots) showed high helical content under helix-inducing conditions, corroborating the structure-activity relationship (Figure 2).
Figure 2. Antimicrobial activity and secondary structure of the peptides generated with AMP-Diffusion.
3.3 Mechanism of Action and Safety Profile
Membrane permeabilization (NPN assay) and depolarization (DiSC3-5 assay) experiments revealed that the generated peptides exert bactericidal effects by disrupting the bacterial membrane. For instance, peptides like AMP-diff2-13, -32, and -33 exhibited membrane permeabilization capacity exceeding that of polymyxin B and levofloxacin, while AMP-diff2-13, -30, and -42 showed strong depolarization effects. Cytotoxicity testing (HEK293T cells) indicated that most peptides showed no significant toxicity at 128 μmol/L, with only 6 peptides having a CC50 below 64 μmol/L (Figure 3), highlighting their excellent safety profile.
Figure 3. Mechanism of action and cytotoxic activity of the peptides.
3.4 In VivoAnti-infective Efficacy
In a murine subcutaneous abscess model, two highly active peptides (AMP-diff2-16 and -43, MIC=1 μmol/L) were administered locally at a dose of 10×MIC. After 4 days of treatment, the peptide groups showed a 2-2.5 log reduction in bacterial load, with efficacy comparable to polymyxin B and levofloxacin, and no observed weight changes or tissue damage. This result confirms, for the first time, the therapeutic potential of AI-generated peptides in a physiological environment (Figure 4).
Figure 4. Anti-infective activity of the peptides in a murine infection model.
04 Conclusion and Future Perspectives
This study successfully developed AMP-Diffusion, an end-to-end AMP generation platform. Its core value lies in the deep integration of generative AI with protein language models, enabling innovation across the entire pipeline from virtual design to in vivovalidation. Experiments demonstrate that this platform efficiently produces diverse, highly active, and low-toxicity antimicrobial peptides, providing a scalable solution to combat antibiotic resistance. Future work could focus on conditional generation (e.g., targeted optimization against specific pathogens), integration of multi-attribute predictors (toxicity, stability), and extension to the design of other functional peptides (e.g., anticancer peptides, antiviral peptides), further advancing the application of generative AI in the biomedical field.
Original Article:
Torres MDT, Chen T, Wan F, Chatterjee P, de la Fuente-Nunez C. Generative latent diffusion language modeling yields anti-infective synthetic peptides. bioRxiv [Preprint]. 2025 Feb 1:2025.01.31.636003.