DISCO: A Multimodal Protein Co-Design Model Enabling DNA-Encoded Chemistry

DISCO is an innovative multimodal generative model that for the first time enables simultaneous co-design of protein sequences and 3D structures. It supports conditional generation with various biomolecules such as small molecules, DNA, and RNA, demonstrates exceptional performance in enzyme design and drug development, and breaks the limitations of traditional step-by-step design strategies.

Traditional protein design uses a step-by-step strategy: first generate the backbone structure, then predict the amino acid sequence, which makes it difficult to ensure the optimal match between sequence and structure. The emergence of DISCO breaks this limitation, enabling co-design of sequence and structure and opening up new possibilities for biomolecular engineering.

Based on a diffusion model architecture, it uses the Hydra configuration system to manage parameters; provides two experimental presets: designable (entropy-adaptive temperature scaling + dual-modal noise guidance, high designability but low diversity), diverse (free sampling, high diversity but low designability); two computational intensity modes: fast (100 diffusion steps + 2 cycles, 4x speed improvement), max (200 diffusion steps +4 cycles, full quality); memory optimization uses DeepSpeed4Science EvoformerAttention, supports long sequences, requires Ampere or higher GPUs, and AMD users need to adjust dependencies.

In the Studio-179 benchmark (179 ligands), it achieved the best performance in 178 out of 179 metrics; Catalytic enzyme design: Inputting reaction intermediates generates heme enzymes with novel active sites, catalyzing non-natural carbene transfer reactions (such as olefin cyclopropanation, etc.). Top designs have higher activity than engineered natural enzymes, and activity increases 4-fold after mutagenesis; Evaluation criteria for co-design ability: RMSD between backbone and ligand centroid after refolding <2 Å.

Supports unconditional generation (70-300 residues), ligand-conditional generation (e.g., heme B, warfarin, etc.), and nucleic acid-conditional generation (DNA/RNA complexes); All experimental samples and results are available on Hugging Face for easy reproduction and verification.

Marks the paradigm shift in protein design from "structure-first" to "sequence-structure co-design"; Improves design success rate and expands functional spaces such as novel catalytic activities; Has important applications in drug development (targeted ligand design) and synthetic biology (new biocatalytic systems); Looks forward to model optimization and community benchmark advancement bringing a new round of breakthroughs.