PLUM

PLUM (Peptide modeLs for Understanding and engineering antiMicrobial therapeutics) is a generative model for the controlled design of antimicrobial peptides (AMPs). AMPs are short peptides with activity against bacteria, fungi, and other pathogens, and they are of growing interest as antibiotic resistance spreads. Designing them computationally is challenging because useful candidates must satisfy multiple constraints at once — sequence composition, antimicrobial function, and length — and most generative approaches do not give designers explicit control over these properties.

PLUM is a conditional variational autoencoder (cVAE) that addresses this by disentangling sequence, function, and length within its latent space, so that each can be steered independently during generation. This lets users generate peptides of a specified length (the model supports 5–35 amino acids) with a desired functional label, either de novo or by conditioning on an existing prototype peptide. The model was developed by Banerjee, Friedberg, Rued, and Eulenstein at Iowa State University and released as a February 2026 bioRxiv preprint. The authors report that PLUM outperforms the prior AMP generator HydrAMP across multiple metrics.

PLUM sits within a line of generative AMP design tools, building on the conditional latent-variable approach popularized by HydrAMP while adding explicit disentanglement of length and function for finer controllability.

#Key Features

• Disentangled latent space: PLUM separates sequence, function, and length factors, giving designers independent control over each property during generation.
• Length-controlled generation: The model can target peptides from 5 to 35 amino acids long, matching the practical size range of antimicrobial peptides.
• Two generation modes: PLUM supports both de novo generation of new sequences and prototype-conditioned generation that produces variants of an existing template peptide.
• Built-in classifiers: The framework includes an AMP/non-AMP function classifier and a potency (minimum inhibitory concentration) classifier for active AMPs, supporting property-guided design.
• Released code and weights: An MIT-licensed implementation is available, with pretrained generative and classifier models distributed via Google Drive.

#Technical Details

PLUM is a conditional variational autoencoder whose latent space is structured to disentangle sequence, function, and length, enabling controlled generation rather than unconditioned sampling. It operates in two modes: de novo generation, which produces entirely new peptide sequences, and prototype-conditioned generation, which creates variants of a provided template. The framework couples the generative model with a functional classifier that distinguishes AMPs from non-AMPs and an additional MIC-based potency classifier for active AMPs, allowing generated candidates to be filtered and conditioned on predicted activity. The authors report that PLUM outperforms HydrAMP across multiple evaluation metrics for AMP generation. The codebase is MIT-licensed and requires Python 3.9+, with pretrained classification, potency, and generative models provided via Google Drive; the preprint is released under a CC BY license.

#Applications

PLUM is aimed at researchers designing antimicrobial peptides as potential therapeutics or antibacterial agents. Its controllable generation is useful for producing candidate libraries with targeted lengths and predicted antimicrobial function, narrowing the experimental search space before wet-lab synthesis and testing. The prototype-conditioned mode supports lead optimization, generating variants of a known active peptide, while the de novo mode supports broader exploration. The integrated AMP and potency classifiers help prioritize candidates likely to be both functional and potent.

#Impact

PLUM advances controllable peptide design by making length and function explicitly steerable within a disentangled latent space, and its reported improvements over HydrAMP suggest meaningful gains for AMP generation. The release of code and pretrained weights under an MIT license lowers the barrier for other groups to adopt and extend the approach. As a February 2026 preprint, the reported advantages have not yet been independently validated, and — as with all computational AMP design — the practical value ultimately depends on experimental confirmation that generated peptides are active and non-toxic, which is not established by the model alone.