Conditional Generation of Paired Antibody Chain Sequences through Encoder-Decoder Language Model

Protein language models (LMs) have been successful in sequence, structural and functional predictions. However, currently, protein LMs are limited to encoder- or decoder-only architectures for single sequences while many biological contexts involve protein-protein interactions. Here, we introduce pAbT5, which models antibody chain pairing as forward- and back-translations using a T5-based architecture. We show that pAbT5 accurately reflects chain pairing through sequence generation and mispairing as unsupervised and supervised classifications. Our protein LM generates variable-length sequences and its next-word prediction probability agrees with position-specific scoring matrix from sequence alignment. Like other works in protein LM, pAbT5 performs state-of-the-art unsupervised prediction on experimental measurements. To the best of our knowledge, pAbT5 is the first encoder-decoder protein LM for protein-protein interactions.