Use DISCO Online
Commercially Available DISCO No-Code Web Server
DISCO: Multimodal Protein Sequence and Structure Co-Design for DNA-Encoded Chemistry
DISCO (DIffusion for Sequence-structure CO-design) is a multimodal generative model that jointly designs a protein's sequence and 3D structure around arbitrary biomolecules. Conditioned only on reactive intermediates, it produced diverse heme enzymes with new active-site geometries that catalyze new-to-nature carbene-transfer reactions. It comes from Jarrid Rector-Brooks, Theophile Lambert, Marta Skreta, Daniel Roth, Yueming Long, Zi-Qi Li, Xi Zhang, Miruna Cretu, Francesca-Zhoufan Li, Tanvi Ganapathy, Emily Jin, Avishek Joey Bose, Jason Yang, Kirill Neklyudov, Yoshua Bengio, Alexander Tong, Frances H. Arnold and Cheng-Hao Liu, with institutions including Caltech and Mila. Code and model checkpoints are available at github.com/DISCO-design/DISCO; the text we read did not state a license.
Evolution has explored only a small part of the chemistry that DNA could encode. Deep generative models can design proteins that bind ligands, but none had designed enzymes without first specifying the catalytic residues. DISCO addresses this by co-generating sequence and structure with a single network.
Attribute | Detail |
|---|---|
Generation | Masked discrete diffusion for sequence, continuous denoising diffusion for coordinates |
Conditioning | Small molecules, metal cofactors, reactive intermediates, DNA and RNA |
Architecture | AlphaFold 3-style components with a frozen DPLM 650M language model; about 888M parameters, 235M trainable |
Training data | PDB, 2021-09-30 cutoff, no designability filtering |
Steering | Feynman-Kac correctors for multimodal rewards and specificity guidance |
How DISCO Co-Designs Protein Sequence and Structure
One network, two modalities: sequence and 3D backbone are generated together rather than backbone first and sequence second.
Ligand-aware tokens: non-protein molecules are tokenized per heavy atom, so small molecules, cofactors and nucleic acids can all serve as targets.
Inference techniques: sequence self-correction, an entropy-adaptive sequence temperature and noisy guidance, which the authors report as key to co-designability.
FKC-MM: inference-time steering with rewards defined over both sequence and structure, such as disulfide-bond content or cation-pi interactions.
FKC-SG: specificity guidance that favors an on-target molecule while penalizing an off-target decoy, tested on pairs such as biotin versus PLP.
DISCO Results: In Silico Benchmarks and New-to-Nature Heme Enzymes
Evaluation | Result |
|---|---|
Unconditional monomers | About 90% of sequences refold within 2 A RMSD (ESMFold) |
STUDIO-179 benchmark | Highest proportion of diverse, co-designable complexes for 178 of 179 ligands |
Enzyme screen | 90 designs tested; 2.2% to 66% exceeded the TmTrpB control |
B-H insertion | 98% yield, TTN 5,170 (evolved variant 2,490) |
Cyclopropanation | 72% yield, TTN 4,050, 99:1 dr |
C(sp3)-H insertion | 42% yield, TTN 2,360 (evolved P411-CHF 2,030) |
No two designed enzymes share more than 50% sequence identity, and the 90 designs form 75 structural clusters. The closest PDB match to dCT-H11 has 21% sequence identity. One round of error-prone PCR on dCT-H11, screening about 700 mutants, improved around 35 variants, showing the designs can be evolved further.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with DISCO on Tamarind Bio
Biocatalyst discovery: design heme enzymes for carbene-transfer reactions such as cyclopropanation and B-H insertion.
Ligand binder design: create proteins around small molecules, metals, DNA or RNA targets.
Specificity engineering: steer designs toward an on-target molecule and away from a decoy.
Starting points for directed evolution: generate diverse scaffolds, then improve them with mutagenesis.
How to Use DISCO on Tamarind Bio
Open the tool: log in to tamarind.bio and open DISCO.
Define the target: provide the small molecule, cofactor, reactive intermediate or nucleic acid the protein should accommodate.
Set design parameters: if exposed, choose protein length and the number of designs.
Choose steering: if exposed, enable multimodal rewards or specificity guidance with a decoy molecule.
Run the job: submit and wait for designs.
Download outputs: save the designed sequences and predicted structures.
Filter and test: check co-designability with a structure predictor, then test top designs experimentally.
Things to Keep in Mind
The frozen language model was not trained on multimeric complexes, so evaluation focused on non-protein-binding tasks.
Only a small subset of useful new-to-nature reactions was tested, and spirocyclopropanation gave fewer, less active designs.
Background activity from endogenous E. coli hemoproteins complicated screening.
Enantioselectivity was not a design criterion.
Design filtering relied on predicted structures (AlphaFold 3, Chai-1), not experimental ones.
Source: Rector-Brooks J, Lambert T, Skreta M, et al. General Multimodal Protein Design Enables DNA-Encoding of Chemistry. arXiv:2604.05181, 2026, https://arxiv.org/abs/2604.05181. Code: https://github.com/DISCO-design/DISCO.