How to Use DISCO

DISCO (Diffusion for Sequence-structure CO-design) is a multimodal generative model that designs both a protein's amino-acid sequence and its three-dimensional atomic structure in one diffusion pass. The method removes the usual backbone-first, inverse-folding-second handoff by learning a joint distribution over sequence tokens and coordinates while conditioning on arbitrary biomolecular partners.

On Neurosnap, researchers choose a Mode and then complete the matching section: Unconditional for de novo protein design, Ligand-Conditioned for small-molecule-centered binder or enzyme design, RNA-Conditioned for RNA-guided complexes, or DNA-Conditioned for DNA-guided complexes. Shared controls such as Experiment Preset, Effort, and Number of Samples define how broadly DISCO explores co-designable sequence-structure space.

The DISCO paper frames protein design as simultaneous denoising of discrete sequence and continuous structure under multimodal conditioning. Cross-modal recycling, entropy-adaptive sequence temperature, and optional noisy guidance help the model keep sequence choices consistent with emerging backbone geometry, which is the central challenge in co-design workflows.

On Neurosnap, Mode determines the scientific setup. Unconditional mode only requires Protein Length and is useful for exploring novel folds without a binding partner. Ligand-Conditioned mode adds an Input Molecule and is the natural entry point for pocket formation, binder design, and chemistry-aware enzyme ideation. RNA-Conditioned and DNA-Conditioned modes accept nucleic-acid inputs together with a shorter protein length range suited to nucleic-acid-binding designs.

Experiment Preset switches between designable and diverse sampling objectives, while Effort controls inference depth. For conditioned campaigns, diverse with max effort is usually the better starting point because it favors broader exploration around the target biomolecule. Researchers should review the exported sequences and structure files as a candidate panel for downstream folding checks, binding validation, or experimental synthesis rather than as a single definitive answer.

• Single-pass co-design: DISCO generates sequence and structure together instead of relying on a separate inverse-folding stage.
• Four conditioning workflows: Unconditional, ligand-, RNA-, and DNA-conditioned modes cover de novo design and biomolecule-guided binder or enzyme campaigns.
• Sampling controls: Experiment Preset, Effort, and Number of Samples let researchers tune diversity versus co-designability.
• Campaign-ready outputs: Co-designed sequences and structure files are exported for comparative triage and downstream validation.