Unifying sequence-structure coding for advanced protein engineering <i>via</i> a multimodal diffusion transformer

Modern protein engineering demands integrated sequence-structure representations to tackle key challenges in designing, modifying, and evolving proteins for specific functions. While sequence-based methods are promising for generating novel proteins, incorporating structure-oriented information improves the success rate and helps target corresponding functions. Therefore, rather than relying solely on sequence or structure-based approaches, a consensus strategy is essential. Here, we introduce ProTokens, machine-learned "amino acids" derived from structural databases <i>via</i> self-supervised learning, providing a compact yet information-rich representation that bridges sequence and structure modalities. Instead of treating sequences and structures separately, we build PT-DiT, a multimodal diffusion transformer-based model that integrates both into a unified representation, enabling protein engineering in a joint sequence-structure space, streamlining the design process and facilitating the efficient encoding of 3D folds, contextual protein design, sampling of metastable states, and directed evolution for diverse objectives. Therefore, as a unified solution for <i>in silico</i> protein engineering, PT-DiT leverages sequence and structure insights to realize functional protein design.