PepCVAE: Semi-Supervised Targeted Design of Antimicrobial Peptide\n Sequences

Given the emerging global threat of antimicrobial resistance, new methods for\nnext-generation antimicrobial design are urgently needed. We report a peptide\ngeneration framework PepCVAE, based on a semi-supervised variational\nautoencoder (VAE) model, for designing novel antimicrobial peptide (AMP)\nsequences. Our model learns a rich latent space of the biological peptide\ncontext by taking advantage of abundant, unlabeled peptide sequences. The model\nfurther learns a disentangled antimicrobial attribute space by using the\nfeedback from a jointly trained AMP classifier that uses limited labeled\ninstances. The disentangled representation allows for controllable generation\nof AMPs. Extensive analysis of the PepCVAE-generated sequences reveals superior\nperformance of our model in comparison to a plain VAE, as PepCVAE generates\nnovel AMP sequences with higher long-range diversity, while being closer to the\ntraining distribution of biological peptides. These features are highly desired\nin next-generation antimicrobial design.\n