Multi-Model Analysis of De novo Antimicrobial Peptide Design Via Variational Autoencoder Latent Sampling

Growing concerns over antibiotic resistance have promulgated the development of novel therapeutic agents to treat microbial infections. Anti-microbial peptides (AMP) are short proteins that possess unique antimicrobial properties capable of overcoming antibiotic resistance through unique mechanisms of action. This work develops a comparison of generative deep learning models used widely in the field of chemical compound discovery as an application to de novo AMP generation. We analyze the competency of 5 architectures in generating novel anti-microbial peptides. The models investigated are adversarial autoencoders networks (AAE), variational auto-encoders (VAE), Wasserstein auto-encoders (WAE), VAE’s with an attention layer (AVAE) and VAE-Transformers (VAE-Trans). The VAE framework generates smooth explorable latent spaces for which interpretability mechanisms are presented. New AMP candidates with desirable features are sampled and verified from the continuous latent spaces using a feature prediction network.