PepGlider: Attribute Regularized VAE for Interpretable and Controllable Peptide Design

ICLR 2026 Conference Submission17220 Authors

Keywords: peptide design, antimicrobial peptides, generative model, disentanglement, continuous attribute regularization, controllable generation, VAE

TL;DR: PepGlider uses continuous attribute regularization to structure latent spaces, enabling precise control of peptide properties and independent tuning of correlated features for improved antimicrobial peptide design.

Abstract: Computational peptide design requires precise control over physicochemical properties that often exhibit complex correlations. Existing generative models rely on simplistic discrete conditioning mechanisms rather than precise targeting of specific property values. We present PepGlider, a continuous attribute regularization framework that enables direct control over specific attribute values. The method achieves structured latent space and displays smooth property gradients with superior disentanglement quality. Experimental results demonstrate that PepGlider enables independent control of naturally correlated properties, and supports both unconstrained generation and targeted optimization of existing peptides. PepGlider applied to antimicrobial peptide design allows generation of candidates with desired antibacterial activity profile. Unlike existing approaches, PepGlider provides precise control over continuous property distributions while maintaining generation quality, offering a generalizable solution for therapeutic and materials applications requiring exact property specifications.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 17220