SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

This post introduces SimpleDesign, a novel single-stage end-to-end model for joint protein sequence and structure codesign. Unlike previous multi-stage approaches, SimpleDesign operates directly in the data space, combining discrete cross-entropy for sequences and a regression objective for structures. It utilizes Transformer-based multimodal backbones with modality-specific processing and global self-attention to effectively model the interplay between sequence and structure. Trained on over 2 million sequence-structure pairs, SimpleDesign demonstrates competitive performance in both co-design and unconditional generation tasks.