Generating Novel Protein Sequences with Self-Supervised Learning

Generative artificial intelligence (AI) represents a major technological advancement, with applications spanning image synthesis to large language models capable of humanlike interaction. Inspired by these developments, we designed a suite of generative deep learning models specifically for protein sequence generation. Leveraging the UniProt database, we evaluated several architectures—including autoencoders with dense and convolutional layers, as well as their variational counterparts. Among these, the convolutional variational autoencoder achieved the lowest reconstruction error. Notably, we found that the condensation level—defined as the size of the latent representation— had a strong impact on sequence identity: greater condensation reduced identity but increased sequence diversity. We identified 80% condensation as a critical threshold that enabled the generation of diverse yet biologically meaningful protein variants. These findings underscore the potential of generative AI to advance protein engineering.