pLM-bVAE: protein language model-enhanced binary representation improves quantum annealing-based protein sequence design

Quantum annealing (QA) has been examined for protein design in extensive sequence combinations. Converting protein sequences into binary vectors for QA optimization demands both high-quality encoding and precise decoding back into amino acid sequences. The efficiency of binary representations and sequence reconstruction accuracy jointly determine the effectiveness of QA-based optimization.

Binary variational autoencoder (bVAE) has been utilized to transform protein sequences into binary formats in QA-driven protein sequence design. However, bVAE, originally crafted for natural language modeling, has not been thoroughly assessed for its appropriateness in handling proteins. In this study, we introduce pLM-bVAE, which integrates protein representations from a pre-trained protein language model (pLM) into the original bVAE framework.

Experiments conducted on three publicly available protein variant datasets revealed that pLM-bVAE exhibited superior sequence reconstruction accuracy and improved binary encoding capabilities compared to bVAE. Furthermore, we integrated pLM-bVAE into a QA-based protein sequence design framework. Across all three datasets, pLM-bVAE produced more sequences with higher scores and greater divergence from the training and validation data than bVAE.

These findings suggest that pLM-bVAE serves as an effective binary representation to enhance QA-based protein sequence design. The source code utilized in this research is accessible at https://github.com/hueda8/pLM-bVAE.