Disentangling the Peptide Space: A Contrastive Approach with Wasserstein Autoencoders

Abstract

Antimicrobial peptides (AMPs) have been shown to be promising therapeutic approaches against antibiotic-resistant pathogens. In the ongoing search for new AMPs, data-driven methods, especially generative models, have become indispensable tools for expediting discovery. We introduce a novel architecture, \textbf{Contrastive Wasserstein Autoencoder (C-WAE)}, designed for the \textit{de novo} generation of AMP candidates by establishing a discriminative latent space of amino acid sequences. The architecture combines Wasserstein distance metrics with a contrastive loss function to achieve a highly separable latent space where AMPs and non-AMPs are distinctly classified. Further, a predictive models trained on a separate validation set could correctly classify as antimicrobial >90\% of samples. Empirical evaluations confirm that the C-WAE succeeds in generating high-quality candidate AMPs as predicted by classifier. Our contributions are twofold: 1) A new architecture for candidate AMP generation using contrastive learning, and 2) To the best of our understanding, this is the first study that integrates contrastive learning for the \textit{de novo} synthesis of AMPs.