Disco

DISCO — Diffusion for Sequence-Structure Co-design

What this is

DISCO is a multimodal diffusion model that jointly generates a protein sequence and a 3D structure in one shot, optionally conditioned on, and co-folded with, one or more non-protein entities (small-molecule ligands, ions, metalloclusters, DNA, RNA). Unlike sequential pipelines (RFdiffusion → ProteinMPNN → AlphaFold refold), DISCO co-designs both modalities so that sequence-based objectives can shape structure and vice versa.

In the paper, DISCO:

• Is state-of-the-art on 178 / 179 ligands of the Studio-179 benchmark (and best on DNA and RNA conditioning) for co-designability — the fraction of generated designs whose backbone and all ligand centroids re-fold to RMSD < 2 Å with Chai-1.
• Was used to design heme enzymes for new-to-nature carbene transfer chemistry (alkene cyclopropanation, spirocyclopropanation, B–H and C(sp³)–H insertions) — conditioning solely on the reactive intermediate (no catalytic residues pre-specified, no template scaffold). Top designs exceeded the activity of previously-engineered enzymes, and random mutagenesis of one design gave a 4× activity gain (i.e. the designs are evolvable).

The code lives at ~/Repos/DISCO and is built on top of the Protenix codebase.

When to use DISCO vs. alternatives

You want…	Use
One-shot joint sequence + structure around a ligand / nucleic acid	DISCO
Backbone-only generation, then a separate sequence-design step	rfdiffusion (backbone) + proteinmpnn / ligandmpnn / solublempnn (sequence). Or boltzgen for all-atom backbone + side chain
End-to-end binder hallucination against a protein target	bindcraft (joint AF2-based) — DISCO does not design protein–protein complexes
Refold / validate a DISCO design	chai (paper uses Chai-1), boltz, or alphafold
Rank refolded binder designs	ipsae ≫ ipTM ≫ iPAE; cross-reference with protein-qc thresholds
Multi-chain antibody / Fv design	Not DISCO — use ABodyBuilder / IgFold pipelines

DISCO's headline strengths are ligand-conditioned active-site design, enzyme design from reactive intermediates, and nucleic-acid-binding protein design. If you only need a protein structure (no biomolecule conditioning) you can still use DISCO unconditionally, but a pure inverse-folding pipeline is usually cheaper.

Quickstart — three steps

1) Install (one-time)

DISCO uses uv for dependency management.

cd ~/Repos/DISCO
uv sync                 # installs torch + everything in pyproject.toml
source .venv/bin/activate

CUDA backend: if you need a specific CUDA wheel for torch (e.g. cu124):

uv pip uninstall torch
uv pip install torch --torch-backend=cu124

AMD GPUs: DeepSpeed has no AMD support. Remove deepspeed from pyproject.toml before uv sync, and pass use_deepspeed_evo_attention=false on every inference command.

Optional but recommended on NVIDIA Ampere+ (A100 / L40S / H100 / B100): clone CUTLASS and point CUTLASS_PATH at it so DeepSpeed4Science EvoformerAttention can compile its memory-efficient kernels on first call:

git clone https://github.com/NVIDIA/cutlass.git /path/to/cutlass
export CUTLASS_PATH=/path/to/cutlass        # add to shell profile to persist

If you skip CUTLASS, pass use_deepspeed_evo_attention=false — DISCO falls back to a naive attention that materializes the full attention matrix (much more GPU memory).

2) Write an input JSON

The JSON lists jobs; each job describes the entities to model. A fully masked protein chain is a string of - characters of the desired length.

[
  {
    "name": "length_200_aspirin",
    "sequences": [
      {
        "proteinChain": {
          "sequence": "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------",
          "count": 1
        }
      },
      {
        "ligand": {
          "ligand": "CC(=O)Oc1ccccc1C(=O)O",
          "count": 1
        }
      }
    ]
  }
]

200 - characters → 200-residue designed protein. Sequence-letter codes in place of - fix that residue; mix to fix termini / hotspots and design the rest.

See references/inputs.md for every supported entity type and field.

3) Run inference

python runner/inference.py \
  experiment=designable \
  input_json_path=input_jsons/your_config.json \
  seeds=\[0,1,2,3,4\]

The first run will JIT-compile the pairformer kernels (CUTLASS path), which can take several minutes before any inference step is logged. The model weights (DISCO.pt) are downloaded automatically from HuggingFace (DISCO-Design/DISCO) on first launch.

If a run is interrupted, rerun the same command — DISCO checks each <name>_sample_<seed>.pdb and skips ones that already exist on disk.

Hydra CLI — every documented flag

DISCO's runner is a Hydra entrypoint, so every config key can be overridden inline with key=value. The most useful ones:

Override	Purpose	Default
experiment=designable|diverse	Sampling-strategy preset (see below)	none
input_json_path=PATH	Input JSON describing jobs	input_jsons/unconditional_config.json
seeds=[0,1,2,...]	Random seeds; total samples = len(seeds) × len(jobs)	[101, 102]
num_inference_seeds=N	Shorthand: use seeds [0..N-1]	null
effort=fast|max	Diffusion-step / recycling preset	max
dump_dir=PATH	Output root	./output
load_checkpoint_path=PATH	Custom checkpoint (else fetched from HF)	null
use_deepspeed_evo_attention=true|false	DeepSpeed4Science kernels (needs CUTLASS_PATH)	true
fabric.num_nodes=N	Multi-node DDP	1
bb_only=true|false	Backbone-only token mask for outputs	true
deterministic=true|false	Force deterministic sampling	true
model.N_cycle=N	Trunk recycling cycles (per-effort: 2 / 4)	4
sample_diffusion.N_step=N	Diffusion sampling steps (per-effort: 100 / 200)	200
sample_diffusion.gamma0=F	Initial diffusion noise scale	0.8 / 1.6
sample_diffusion.gamma_min=F	Min gamma	1.0
sample_diffusion.gamma_anneal=none|linear|cosine	Anneal gamma over the trajectory	none
sample_diffusion.noise_scale_lambda=F	EDM noise multiplier λ	1.003 / 0.1
sample_diffusion.step_scale_eta=F	Step-size η	1.5
sample_diffusion.integrator=euler|heun	Diffusion integrator (heun = 2× NFE)	euler
sample_diffusion.noisy_guidance.enabled=true|false	Toggle structure+sequence noisy guidance	true (designable) / false (diverse)
sample_diffusion.noisy_guidance.omega_struct=F	Structure guidance scale	1.5
sample_diffusion.noisy_guidance.omega_seq=F	Sequence guidance scale	2.0
sample_diffusion.noisy_guidance.rescale_phi=F	Rescaling factor φ	0.7
sample_diffusion.noisy_guidance.guidance_start_frac=F	Fraction of trajectory at which guidance starts	0.3
sample_diffusion.noisy_guidance.guidance_end_frac=F	Fraction at which guidance stops	0.8
sequence_sampling_strategy.logits_temp=F	Decoder logits temperature	0.8
sequence_sampling_strategy.entropy_adaptive_temp=true|false	Per-position entropy-adaptive T	true (designable) / false (diverse)
sequence_sampling_strategy.entropy_adaptive_beta=F	Strength of entropy-adaptive scaling	1.0
sequence_sampling_strategy.score_type=random|...	Path-planning score selector	random
sequence_sampling_strategy.sequence_noise_scheduler.power=F	Sequence-noise schedule exponent	1.0 / 1.5
n_seq_duplicates_per_structure=N	Sample N sequences per generated structure	1
inference_noise_scheduler.rho=F	EDM σ-schedule exponent	7.0
model.use_joint_diffusion_module=false	Internal: joint vs disjoint diffusion module	false
need_atom_confidence=true|false	Dump per-atom confidence in output	false
output_format=unconditional_monomer_protein|null	PDB-style output (default) or CIF dump	unconditional_monomer_protein
dtype=fp32|bf16|fp16	Inference precision	bf16

Anything visible in configs/ can be overridden the same way. See references/cli.md for the long-form table and worked examples.

The two experiment presets

Preset	Sequence T scaling	Noisy guidance	Best for
designable	Entropy-adaptive (per-position)	On (struct + seq, ω≈1.5/2.0)	Unconditional + cases where you want refoldable samples. Paper default.
diverse	Plain (logits_temp=0.8, no entropy adaptive)	Off	Ligand / DNA / RNA conditioning — recovers the paper's headline Studio-179 numbers. Also: higher structural variety.

Cheap-but-near-designable run: keep the other designable knobs and turn guidance off — sample_diffusion.noisy_guidance.enabled=false. This costs ~10% in co-designability but cuts the effective batch size of each forward pass (guidance doubles it).

The two effort presets

Preset	N_step	N_cycle	Speed	Quality
fast	100	2	~4× faster	~10% lower co-designability
max	200	4	baseline (paper)	full quality

Rule of thumb: effort=fast only for unconditional generation. For any ligand- / DNA- / RNA-conditioned design use effort=max. To benchmark against DISCO, always use effort=max.

You can also override the underlying knobs directly: model.N_cycle=3 sample_diffusion.N_step=150.

Input JSON schema — at a glance

Each file is a JSON list of jobs. Each job has a name (used as the output filename prefix) and a sequences list of entities, plus an optional covalent_bonds list.

Supported entity types:

Entity	Key	Sequence alphabet	Masking with -?
Protein chain	proteinChain	20 AAs + X (unknown)	Yes
DNA (single strand)	dnaSequence	A, T, G, C, N	No — fully specified only
RNA (single strand)	rnaSequence	A, U, G, C, N	No — fully specified only
Ligand	ligand	CCD code (CCD_XXX[_YYY...]), SMILES, or FILE_<path> to SDF / MOL / MOL2 / PDB with a 3D conformer	n/a
Ion / metal	ion	CCD code (MG, ZN, FE, CA, NA, ...)	n/a

Field details and an annotated example: references/inputs.md.

Three ways to specify a ligand

1. SMILES — "ligand": "CC(=O)Oc1ccccc1C(=O)O". DISCO generates a 3D conformer at runtime via RDKit (EmbedMolecule). Will assert if RDKit fails to embed — for those, pre-generate a conformer and use option 2.
2. File — "ligand": "FILE_path/to/molecule.sdf". SDF, MOL, MOL2, PDB. Must be 3D (2D will be rejected). Paths can be absolute or relative to the DISCO repo root. XYZ is not supported — convert with obabel input.xyz -O output.sdf first.
3. CCD code — "ligand": "CCD_ATP". For multi-component ligands (glycans, etc.) join CCD codes with _: "CCD_NAG_BMA_BGC".

Multi-strand DNA

dnaSequence is one strand. For double-stranded DNA, add the reverse complement as a second dnaSequence entity in the same job. Same applies if you want explicit base-pairing for RNA.

Partial sequence masking

Mix masked and fixed residues to fix specific positions (e.g. an N-terminal signal, a known active-site motif). Example: "MKTL----VPEG" fixes 4 residues at each end and designs the 4 middle residues.

Covalent bonds

Add a covalent_bonds field to a job to connect specific atoms across entities (e.g. a covalently-attached ligand, a thioester intermediate). Each entry is:

{
  "left_entity": 1, "left_position": 12, "left_atom": "SG",
  "right_entity": 2, "right_position": 1, "right_atom": "C1",
  "left_copy": 1, "right_copy": 1
}

Entities are 1-indexed in the order they appear in sequences. Atom names are CCD-standard. For SMILES ligands, you can use the SMILES atom map number instead of "left_atom": "C1".

Outputs — the file layout

dump_dir/
├── pdbs/
│   ├── <name>_sample_<seed>.pdb              # generated structure (one per sample)
│   └── <name>_sample_<seed>_ligands.txt      # ligand SMILES — only if ligands present
├── sequences/
│   └── <name>_sample_<seed>.txt              # FASTA-like; >cogen_seq 0, then sequence,
│                                             #   then optional dna_sequence / rna_sequence
│                                             #   / ligand_smiles lines
└── ERR/
    └── <name>.txt                            # only for failed samples (with traceback)

By default DISCO writes PDB files (output_format=unconditional_monomer_protein). Set output_format=null to instead write CIF files (and richer per-residue metadata) via the dumper.

If n_seq_duplicates_per_structure > 1, the sequence file contains multiple >cogen_seq i records for the same backbone — useful for generating a sequence ensemble per structure.

If need_atom_confidence=true, the dumper also writes per-atom confidence data alongside summary scores.

See references/outputs.md for full schemas.

Evaluation — co-designability

The paper's headline metric is co-designability: a design counts if, after refolding the generated sequence with Chai-1, the protein backbone and the centroids of all conditioning ligands are within < 2 Å RMSD of the DISCO design. Studio-179 reports the fraction of generated designs that are diverse and co-designable.

To replicate the metric for your own runs, use the chai (or boltz / alphafold) skill to refold each sequences/*.txt with the original ligand / DNA / RNA conditioning, align to the corresponding pdbs/*.pdb, and threshold on backbone + ligand-centroid RMSD. See references/evaluation.md for a worked recipe.

The Studio-179 benchmark

DISCO ships the SDF files for 179 ligands (170 single + 9 multi-ligand) under studio-179/priority_{0,1,2,3}/. The benchmark covers:

• rigid molecules (e.g. tetrachlorodibenzodioxin),
• large or flexible molecules (e.g. CoQ10, 50-heavy-atom isoprenoid),
• metals and metalloclusters (e.g. [4Fe-4S], Fe / Co / Hg / Pb "ideal" metal-centre placeholders),
• organometallic catalysts (Ru(bpy), Ir-biotin, Ir-piano), and
• the reactive carbene intermediate used for the heme-enzyme experiment (in priority_0/).

Pre-split jobs for parallel runs are in input_jsons/all_priorities_ligands_split_{0..3}.json. Each split runs every job at three lengths (150, 200, 250) — the canonical benchmark configuration.

python runner/inference.py \
  experiment=diverse \
  effort=max \
  input_json_path=input_jsons/all_priorities_ligands_split_0.json \
  seeds=\[$(seq -s "," 0 4)\]

For a single ligand, mirror input_jsons/heme_b.json: three jobs at lengths 150 / 200 / 250 each referencing the same SDF. The example file under examples/single_ligand_studio179.json is ready to adapt.

See references/studio-179.md for tiers, priority structure, and per-tier expected pass rates.

Choosing the right reference doc

You want to…	Read
Install (uv, CUTLASS, AMD path, HF weights, custom checkpoints)	references/installation.md
See every Hydra override with examples	references/cli.md
Author an input JSON for protein / ligand / ion / DNA / RNA / covalent	references/inputs.md
Pick the right preset for your use case (designable vs diverse, fast vs max)	references/presets.md
Read PDB / sequences / ERR / _ligands.txt outputs	references/outputs.md
Reproduce Studio-179 or benchmark a custom ligand against it	references/studio-179.md
Compute co-designability with Chai-1 / Boltz / AF2	references/evaluation.md
Design a heme enzyme (or any reactive-intermediate-conditioned design)	references/enzyme-design.md
Design a DNA- or RNA-binding protein	references/nucleic-acid.md
Run unconditional length sweeps for backbones	references/unconditional.md
Run partial-mask jobs (fix hotspots, design the rest)	references/partial-mask.md
Multi-GPU / multi-node / large-scale screening	references/scaling.md
Diagnose JIT-compile, CUTLASS, OOM, MSA, or shape-mismatch failures	references/troubleshooting.md
What DISCO cannot do (and what's planned)	references/limitations.md

Quick decision tree

• Just want a backbone + sequence of a given length, no target → experiment=designable, input_jsons/unconditional_config.json, effort=fast is fine.
• Want a binder around a small molecule (SMILES in hand) → put 1× proteinChain (~150–250 -s) + 1× ligand with the SMILES in a job. experiment=diverse, effort=max. Run multiple lengths.
• Want a binder around a metal cofactor → use CCD_FE (or whatever CCD) as an ion, or the appropriate _ideal.sdf from studio-179/ as a ligand with FILE_. experiment=diverse, effort=max.
• Want a heme enzyme around a transition-state intermediate → see references/enzyme-design.md. Pass the TS SDF as the only ligand; don't pre-specify catalytic residues.
• Want an RNA-binding protein for an aptamer sequence → 1× masked protein chain + 1× rnaSequence (no - allowed). Lengths 50–80 recovered paper results.
• Want a TF-like dsDNA binder → 1× protein chain + 1× dnaSequence (single strand) + 1× dnaSequence (reverse complement). experiment=diverse, effort=max.
• Want to fix the N- and C-terminal motifs but redesign the loop → partial mask the protein chain: "MKTL----VPEG".
• Want many seeds in parallel across GPUs → fabric.num_nodes>1 and launch with srun / torchrun; DDP shards the dataloader automatically.
• Want a sequence ensemble for the same backbone → n_seq_duplicates_per_structure=N. Output sequences file gets N >cogen_seq i records.

Hard rules / gotchas

1. One protein chain per job. DISCO is single-chain-only — it inherits this from DPLM (the LM that replaces the MSA module). >1 proteinChain per job will raise. Protein–ligand, protein–DNA, protein–RNA, and protein–multi-ligand are all fine.
2. No motif scaffolding (yet). Fixing a structural motif while designing the surrounding scaffold isn't supported — only sequence-level masking is. The maintainers list this as planned.
3. DNA and RNA must be fully specified. Masking nucleic acid positions with - is not supported. Provide complete sequences.
4. Double-stranded DNA is two dnaSequence entries. DISCO does not auto-pair strands; you must add the reverse complement explicitly.
5. Ligand files must have a 3D conformer. SDF / MOL / MOL2 / PDB with 2D coordinates will be rejected. XYZ is not supported — convert first.
6. SMILES that RDKit can't embed will hard-assert. Workaround: pre-generate a conformer with RDKit / Open Babel and pass it as FILE_.
7. effort=fast is unconditional-only. For ligand / DNA / RNA conditioning, use effort=max — otherwise co-designability drops sharply.
8. infer_batch_size=1 is the only tested setting. The runner makes assumptions about this elsewhere. Don't override.
9. - is mask only inside proteinChain. It's a residue letter, not a generic token. In other entity types it's not interpreted.
10. DeepSpeed EvoformerAttention requires Ampere+ NVIDIA and CUTLASS. On AMD or older NVIDIA, run with use_deepspeed_evo_attention=false — slower / much more memory, but works.
11. Run-resume is name-keyed. DISCO skips a sample iff <dump_dir>/<name>_sample_<seed>.pdb exists. Renaming the job re-runs it; changing seeds re-runs only the new seeds.
12. First-launch JIT compile. Pairformer kernels compile lazily — nothing is broken if the first 5-10 minutes are silent.
13. Sequence outputs are FASTA-ish, not strict FASTA. Each file mixes >cogen_seq i records with dna_sequence / rna_sequence / ligand_smiles annotation lines. Parse line-prefix, not BioPython.
14. There is no built-in refolder. Co-designability requires an external structure predictor — use Chai-1 (paper default) via the chai skill.
15. Single-component CCD codes can be entered as ion or ligand. Use ion for monatomic metals (NA, MG, ZN, FE, CA) — the parser routes it to the same builder but tags it as an ion.
16. No protein-protein complexes. This is the model's biggest gap. Use bindcraft (joint AF2 hallucination) for protein-target binders.

Installing this skill

# Symlink (recommended — picks up edits live)
mkdir -p ~/.claude/skills
ln -s "$(pwd)" ~/.claude/skills/disco

# Or copy:
cp -R . ~/.claude/skills/disco

After that, an agent invokes it via Skill(skill="disco").

Citation

@Article{disco2026,
  title   = {General Multimodal Protein Design Enables DNA-Encoding of Chemistry},
  author  = {Jarrid Rector-Brooks and Théophile Lambert and Marta Skreta and
             Daniel Roth and Yueming Long and Zi-Qi Li and Xi Zhang and
             Miruna Cretu and Francesca-Zhoufan Li and Tanvi Ganapathy and
             Emily Jin and Avishek Joey Bose and Jason Yang and
             Kirill Neklyudov and Yoshua Bengio and Alexander Tong and
             Frances H. Arnold and Cheng-Hao Liu},
  year    = {2026},
  eprint  = {2604.05181},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url     = {https://arxiv.org/abs/2604.05181}
}