geestaltt/peptide-vae-amp

PeptideVAE: Hyperbolic Antimicrobial Peptide Predictor

PeptideVAE is a Transformer-based Variational Autoencoder (VAE) designed for the analysis and design of antimicrobial peptides (AMPs). It leverages hyperbolic geometry (Poincaré ball) to learn a structured latent space where radial distance encodes antimicrobial potency (MIC).

Model Description

• Architecture: Transformer Encoder-Decoder with Hyperbolic Projection.
• Latent Space: 16-dimensional Poincaré ball (curvature $c=1.0$).
• Input: Peptide sequences (length 10-50 AA).
• Task: Regression (Log10 MIC prediction) and Sequence Generation.

Key Features

• Biological Grounding: Stress-tested for sensitivity to physicochemical properties. Correctly predicts reduced activity when positive charge density is decreased (e.g., K $ ightarrow$ E mutation in LL-37).
• Hyperbolic Hierarchy: Higher antimicrobial activity (lower MIC) correlates with specific radial positioning in the hyperbolic manifold.
• Multimodal Embedding: Combines amino acid identity, 5-adic physicochemical groups, and standard AA properties.

Performance (5-Fold Cross-Validation)

Metric	Mean Value	Max (Best Fold)
Spearman $
ho$	0.633	0.760
Pearson $r$	0.618	0.749
MAE	0.32 Log10 units	0.25
Reconstruction	100%	100%

Usage

import torch
from peptide_vae import PeptideVAE

# Load model
model = PeptideVAE(latent_dim=16, hidden_dim=64)
checkpoint = torch.load("pytorch_model.bin", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Encode sequence
seq = "LLGDFFRKSKEKIGKEFKRIVQRIKDFLRNLVPRTES"
with torch.no_grad():
    out = model.encode([seq])
    z_hyp = out['z_hyp']
    mic_pred = model.predict_mic(z_hyp)

print(f"Predicted Log10(MIC): {mic_pred.item():.4f}")

Citation

If you use this model in your research, please cite:

@software{peptide_vae_2026,
  author = {AI Whisperers},
  title = {PeptideVAE: Hyperbolic Antimicrobial Peptide Predictor},
  year = {2026},
  url = {https://huggingface.co/ai-whisperers/peptide-vae-amp}
}