General Multimodal Protein Design Enables DNA-Encoding of Chemistry
Authors: Jarrid Rector-Brooks, Théophile Lambert, Marta Skreta, Daniel Roth, Yueming Long, Zi-Qi Li, Xi Zhang, Miruna Cretu, +10 more
Organizations: California Institute of Technology · Mila – Québec AI Institute · Université de Montréal · Université Paris-Saclay · McGill University · University of Cambridge · University of Oxford · Imperial College London · Institut Courtois · LawZero · AITHYRA · FutureHouse
Evolution is an extraordinary engine for enzymatic diversity, yet the chemistry it has explored remains a narrow slice of what DNA can encode. Deep generative models can design new proteins that bind ligands, but none have created enzymes without pre-specifying catalytic residues. We introduce DISCO (DIffusion for Sequence-structure CO-design), a multimodal model that co-designs protein sequence and 3D structure around arbitrary biomolecules, as well as inference-time scaling methods that optimize objectives across both modalities. Conditioned solely on reactive intermediates, DISCO designs diverse heme enzymes with novel active-site geometries. These enzymes catalyze new-to-nature carbene-transfer reactions, including alkene cyclopropanation, spirocyclopropanation, B-H, and C(sp3)-H insertions, with high activities exceeding those of engineered enzymes. Random mutagenesis of a selected design further confirmed that enzyme activity can be improved through directed evolution. By providing a scalable route to evolvable enzymes, DISCO broadens the potential scope of genetically encodable transformations. Code is available at https://github.com/DISCO-design/DISCO.
Figures & tables
Figure 1: Multimodal protein design workflow with Disco . ( A ) Inference overview, highlighting the arbitrary molecular context, sequence-structure co-generation with sequence correction, and representative applications. ( B ) Architecture overview. ( C ) Multimodal inference-time steering schematics with Feynman-Kac corrector. ( D ) Design-to-test workflow for generating enzymes with new active sites from reactive intermediates.
Figure 2: Disco generates diverse, designable protein sequences and structures across a wide variety of functions evaluated in silico. ( A ) In unconditional monomer generation, Disco outperforms existing methods in terms of co-designability, novelty, and diversity (400 samples). ( B ) Ablation of unconditional monomer generation with Disco to identify best inference-time practices. ( C ) Disco outperforms existing methods for conditional generation across diverse targets, including natural and non-natural molecules, multi-ligands, DNA, and RNA. ( D ) Examples of designed (blue) and refolded (white) protein complexes show high agreement. ( E ) FKC-MM enables the joint steering of sequence and structure, as demonstrated by two multimodal rewards: disulfide bridges and cation- π interactions. Example generated proteins with high disulfide bridge counts are shown at right. ( F ) FKC-SG generates more proteins with predicted binding specificity (a large separation between binding sites of on- and off-targets) compared to existing filtering methods. Example refolded structures are shown at right (on-target in pink, off-target in yellow).
Figure 3: Disco designs biophysically realistic proteins with novel motifs, flexible ligands, and target-informed sequences. ( A - E ) Unconditional generation results: ( A ) Amino acid type distribution as compared to PDB and inverse-folded samples. ( B ) Secondary structure diversity. ( C ) Ramachandran plot. ( D ) Surface hydrophobicity and net charge. ( E ) Distribution of long-range contact order. ( F - M ) Conditional generation results: ( F ) Folddisco motif similarity searches reveal no close matches to designed residue motifs in AlphaFoldDB. ( G ) Example active-site (purple) compared to the closest known motif (yellow). ( H ) Distribution of global sequence/structure novelty. ( I ) Disco designs binding sites with consideration of ligand properties, as illustrated by matching lipophilicity. ( J ) Example design satisfying ligand constraints, building four disulfide bonds to a [4Fe-4S] cluster. ( K ) Disco -generated conformers are valid as evaluated by PoseBusters, even if they deviate from the reference. ( L ) Disco samples different conformers. ( M ) Examples of sampled conformers that are rare or absent in the training set.
Figure 4: Disco -designed carbene transferases (dCTs) catalyze multiple carbene transfer reactions with high activity at room temperature, including ( A ) p -methoxystyrene cyclopropanation. ( B ) B–H bond insertion of carbenes. ( C ) C(sp 3 )–H alkylation. ( D ) Spirocyclopropanation of an unsaturated exocyclic N-heterocycle. Each panel shows a reaction scheme, results for hemin/previous evolved variants, plate-based screening data, and yield and TTN results from selected validations. ND: Not Detectable.
Figure 5: Novelty of selected, designed enzymes and evaluation of evolvability. ( A ) dCT-H11 design (green/pink) structure and the AlphaFold 3/Chai predicted structures (grey), alongside the closest PDB match by TM-score. The inset shows the five-residue active-site arrangement predicted by AlphaFold 3 (purple), compared to the closest motif in AlphaFoldDB (yellow). ( B ) Mutational landscape of error-prone PCR variants in plate-based screening of dCT-H11 for the spirocyclopropanation reaction. ( C ) Designed and predicted structures of dCT-F9, dCT-G9, dCT-H10, alongside the closest PDB matches by TM-score. The five closest active-site residues yielded no full match in AlphaFoldDB.
Appendix figures & tables92 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Co-designability increases while structural diversity decreases as training progresses.
Category
Hyperparameter
Value
Global Dimensions & Training
Global
Single Representation Dimension ( cs )
384
Global
Pair Representation Dimension ( cz )
128
Global
Gradient Clipping Norm
10.0
Global
EMA Decay Rate
0.999
Global
Sequence Batch Size
32
Appendix
Table S1 : Model and Training Hyperparameters
Figure S2 : Various noise schedules used as inference settings. The leftmost plot displays a naïve schedule representing the default AlphaFold 3 settings paired with a linear sequence schedule. The middle plot is the schedule used in our most co-designable settings, while the right is our most diverse settings (using mask_proportion=t1.5 for sequence).
Figure S3 : DFT-calculated reaction mechanism for the carbene-insertion reactions using Fe-porphyrin as the cofactor. All design campaigns used intermediate 1 reoptimized with heme b.
Method
Co-designable
Struct. Diversity
Seq. Diversity
Struct. Novelty
Seq. Novelty
pLDDT
ESM3
0.145
0.040
0.113
0.697
0.806
0.533
ProteinGenerator
0.355
0.118
0.273
0.173
0.998
0.779
La-Proteína
0.895
0.293
0.790
0.173
0.879
0.917
BoltzGen
0.798
0.065
0.310
0.132
0.865
0.882
RFDiffusion3
0.460
0.135
0.450
0.166
0.926
0.766
Disco (ours)
0.878
0.343
0.855
0.211
0.961
0.879
Appendix
Table S4 : Unconditional generation metrics. Bold indicates best per column; underline indicates second best. Co-designable structural and sequence diversity are the most informative metrics because they reflect both sample quality and the absence of mode collapse, whereas one should not pursue the maximization of co-designability alone, as folding models cannot refold every PDB protein within 2 Å RMSD. Note that only Disco , BoltzGen and RFDiffusion3 can condition on other biomolecules.
Figure S4 : Co-designability and number of structure clusters for different numbers of steps and recycles. Performed on the unconditional task with the settings of Table S2 and noisy guidance disabled.
Figure S5 : Distribution of unconditional generation co-designability and the proportion of co-designable clusters across protein lengths.
Figure S6 : Amino acid distribution for unconditional generation using our sequence–structure model. Shown are sequences designed by our model, sequences obtained by inverse folding our generated structures with ProteinMPNN, and sequences from 50,000 randomly sampled PDB structures from the training data.
Figure S7 : Distributions of long-range contact types in unconditionally generated samples. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S8 : Distributions of radius of gyration in unconditionally generated samples. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S9 : Distribution of the number of disulfide bonds present in each protein chain of the PDB training data, where the length of the chain is less than or equal to 100.
Figure S10 : Benchmark performance across the remaining set of Studio-179 , ranging from natural to non-natural ligands. We observe that for all ligands Disco generates the highest proportion of diverse, co-designable sequence-structure pairs.
Figure S11 : Distribution of average conditional generation co-designability and the proportion of co-designable clusters across protein lengths, where each scatter point corresponds to a ligand type.
Figure S12 : Distribution of amino acids for conditional generation using our sequence–structure model. Results are aggregated from all benchmark ligands. Shown are sequences designed by our model and sequences from 50,000 randomly sampled PDB structures from the training data. Top left shows all generated samples, top right shows only co-designable samples, and bottom shows all generated samples, but split into specific amino acids. Note that despite reduced co-designability, there is no significant change in the distribution of the types of amino acids.
Figure S13 : Secondary structure distribution for conditional generation using our sequence–structure model. Top shows results from all benchmark ligands without filtering, and bottom shows the effect of filtering by co-designability. We include a comparison against 50,000 randomly sampled PDB structures from the training data. Because current folding methods successfully refold only about 60% of the PDB training set, co-designability is likely a biased filter.
Figure S14 : Distributions of relative contact order for conditional generation using our sequence–structure model (top) and long-range contact types (bottom). Results are aggregated from all benchmark ligands without co-designability filtering. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S15 : Distribution of surface hydrophobicity and net charge for conditional generation using our sequence–structure model. Results are aggregated from all benchmark ligands without co-designability filtering. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S16 : Ramachandran plot for conditional generation using our sequence–structure model. Results are aggregated from all benchmark ligands without co-designability filtering. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S17 : Distributions of radius of gyration for conditional generation using our sequence–structure model. Results are aggregated from all benchmark ligands without filtering. We include a comparison against 50,000 randomly sampled PDB structures from the training data.
Figure S18 : PoseBusters validity breakdown for conditional generation results without co-designability filtering.
Figure S19 : Average rate of any steric clash between the designed protein structure and the ligand(s), grouped by ligand type, without co-designability filtering.
Figure S20 : Number of rotatable bonds compared to the median of conformer RMSD (with standard deviation), split by ligand type, without co-designability filtering.
Figure S21 : The strip plot shows the average binding site residue composition, split by ligand type, without co-designability filtering. Note that the charged-residue outliers correspond to metals that require charged coordination.
Figure S22 : Comparison of chemical fidelity indicators of binding motif and ligand interaction, averaged over all ligand types that all three models share (i.e. no multi-ligands). Top shows the correlation of pocket hydrophobicity with ligand lipophilicity. Bottom shows whether the designed proteins have any backbone clashes with the ligand. Together with the co-designability cluster plot shown in the main text, these results establish that Disco yields realistic binding pockets that respond to ligand’s chemical and physical properties.
Figure S23 : Distribution of motif cluster diversity (averaged per ligand type), across K closest residues to the ligand.
Figure S24 : Violin plot showing the full distribution of per-design nearest-neighbour diversity of the 5 closest residues to the ligand, each overlaid strip point shows a per-ligand median.
Figure S25 : Additional conditional design examples. Left shows conditioning on luciferin, a natural luminescent molecule, in a hybrid helix and sheet topology. Right shows a substituted naphthalene diimide, a non-natural electron acceptor, in a helix bundle.
Figure S26 : Additional generated conformer examples, compared to the reference conformer. Left shows bacteriopheophytin A, and right shows Nirmatrelvir.
Figure S27 : An additional generated motif: the bonding geometries of a class I type II Cu center (2 His, 2 Cys, 1 Glu residues in tetrahedral coordination with Cu[2+]). This suggests the model can build informed residues to satisfy chemical constraints.
Figure S28 : Additional generated motifs: two novel motifs (purple) aligned to the closest known motif (beige) in AlphaFoldDB. This indicates the model can build new residue motifs. Note that many generated motifs yield no match via Folddisco.
Figure S29 : Enantiomeric excess (%) of methoxystyrene cyclopropanation product (trans, 1b ). The corresponding enantiomeric ratio of dCT-H11 is 61:39.
Figure S30 : Enantiomeric excess (%) of the product of B–H insertion 2b . The corresponding enantiomeric ratio of dCT-F9 is 63:37.
Figure S31 : Enantiomeric excess (%) of the product of C–H insertion 3b . The corresponding enantiomeric ratio of dCT-H11 is 40:60.
Figure S32 : Diastereomeric excess (%) of methoxystyrene cyclopropanation product (trans vs cis, 1b ) in selected validation results. The corresponding diastereomeric ratio of dCT-H11 is 99:1 (trans:cis).
Primers
Sequence (5’ → 3’)
005
GAA ATA ATT TTG TTT AAC TTT AAG AAG GAG ATA TAC ATA TG
006
GCC GGA TCT CAG TGG TGG TGG TGG TGG TGC TCG AG
007
CAT ATG TAT ATC TCC TCC TTA AAG TTA AAC AAA ATT ATT TC
Table S10 : Mutations of highest-performing variants from error-prone PCR.
Figure S33 : Selected validation results of the evolution of the spirocyclopropanation reaction 4b . The parent (dCT-H11) is shown in gray. The corresponding enantiomeric ratios of dCT-H11, dCT-r1-E12, and dCT-r1-F1 are 68:32, 75:25 and 25:75, respectively.
Figure S34 : Calibration curve of the cis isomer of 1b using 1,2-diphenylethane as IS.
Figure S35 : Calibration curve of the trans isomer of 1b using 1,2-diphenylethane as IS.
Figure S36 : Calibration curve of 2b using 1,2,3-trimethoxystyrene as IS.
Figure S37 : Calibration curve of 3b using 1,2-diphenylethane as IS.
Figure S38 : Calibration curve of Compound 4b using 1,2-diphenylethane as IS.
Figure S39 : Calibration curve of 4b on LC-MS for directed evolution assay.
Figure S40 : Synthesis of 3b .
Figure S41 : 1 H NMR (400 MHz, CDCl3) of 3b .
Figure S42 : 13 C NMR (101 MHz, CDCl3) of 3b .
Figure S43 : HPLC-UV chromatogram of an authentic standard of compound 1b .
Figure S44 : HPLC-UV chromatogram of compound 1b (reaction with dCT-B9).
Figure S45 : HPLC-UV chromatogram of compound 1b (reaction with dCT-F3).
Figure S46 : HPLC-UV chromatogram of compound 1b (reaction with dCT-F9).
Figure S47 : HPLC-UV chromatogram of compound 1b (reaction with dCT-G1).
Figure S48 : HPLC-UV chromatogram of compound 1b (reaction with dCT-G9).
Figure S49 : HPLC-UV chromatogram of compound 1b (reaction with dCT-H5).
Figure S50 : HPLC-UV chromatogram of compound 1b (reaction with dCT-H10).
Figure S51 : HPLC-UV chromatogram of compound 1b (reaction with dCT-H11).
Figure S52 : HPLC-UV chromatogram of compound 1b (reaction with TmTrpB control).
Figure S53 : HPLC-UV chromatogram of compound 1b (reaction with ApePgb 5312 control).
Figure S54 : HPLC-UV chromatogram of an authentic standard of compound 3b .
Figure S55 : HPLC-UV chromatogram of compound 3b (reaction with dCT-B9).
Figure S56 : HPLC-UV chromatogram of compound 3b (reaction with dCT-F3).
Figure S57 : HPLC-UV chromatogram of compound 3b (reaction with dCT-F9).
Figure S58 : HPLC-UV chromatogram of compound 3b (reaction with dCT-G1).
Figure S59 : HPLC-UV chromatogram of compound 3b (reaction with dCT-G9).
Figure S60 : HPLC-UV chromatogram of compound 3b (reaction with dCT-H5).
Figure S61 : HPLC-UV chromatogram of compound 3b (reaction with dCT-H10).
Figure S62 : HPLC-UV chromatogram of compound 3b (reaction with dCT-H11).
Figure S63 : HPLC-UV chromatogram of compound 3b (reaction with TmTrpB control).
Figure S64 : HPLC-UV chromatogram of compound 3b (reaction with ApePgb 5312 control).
Figure S65 : HPLC-UV chromatogram of an authentic product standard of compound 2b .
Figure S66 : HPLC-UV chromatogram of compound 2b (reaction with dCT-B9).
Figure S67 : HPLC-UV chromatogram of compound 2b (reaction with dCT-F3).
Figure S68 : HPLC-UV chromatogram of compound 2b (reaction with dCT-F9).
Figure S69 : HPLC-UV chromatogram of compound 2b (reaction with dCT-G1).
Figure S70 : HPLC-UV chromatogram of compound 2b (reaction with dCT-G9).
Figure S71 : HPLC-UV chromatogram of compound 2b (reaction with dCT-H5).
Figure S72 : HPLC-UV chromatogram of compound 2b (reaction with dCT-H11).
Figure S73 : HPLC-UV chromatogram of compound 2b (reaction with TmTrpB control).
Figure S74 : HPLC-UV chromatogram of compound 2b (reaction with ApePgb 5312 control).
Figure S75 : Chiral GC-FID trace for racemic standard of 4b .
Figure S76 : Chiral GC-FID trace for an enantiopure standard of (R)- 4b .
Figure S77 : Chiral GC-FID trace for 4b (reaction with dCT-H11).
Figure S78 : Chiral GC-FID trace for 4b (reaction with evolved variant dCT-r1-B7).
Figure S79 : Chiral GC-FID trace for 4b (reaction with evolved variant dCT-r1-E3).
Figure S80 : Chiral GC-FID trace for 4b (reaction with evolved variant dCT-r1-E12).
Figure S81 : Chiral GC-FID trace for 4b (reaction with evolved variant dCT-r1-F1).
Figure S82 : Chiral GC-FID trace for 4b (reaction with evolved variant dCT-r1-D1).
Proteins perform their biological functions through three-dimensional structures encoded by amino acid sequences, and ligand-binding protein co-design requires models that generate sequence-structure compatible proteins under explicit ligand constraints. Although continuous diffusion and flow-based models support ligand-aware design in coordinate or latent spaces, existing discrete diffusion protein language models mainly operate over sequence or structure tokens without direct small-molecule conditioning. We introduce \textbf{ProtLiD2}, a Protein Ligand-conditioned Discrete Diffusion model for protein sequence-structure co-design. ProtLiD2 jointly generates amino-acid sequence and discrete structure tokens while incorporating ligand chemical and geometric information through geometry-aware cross-attention. Trained on over one million ligand-protein complexes, ProtLiD2 extends masked discrete diffusion to ligand-aware functional protein design. We further propose maximum confidence-margin guided ReMask decoding, an inference-time self-correction strategy that retains confident predictions and remasks uncertain tokens. ProtLiD2 improves global fold confidence over Complexa in whole-protein design, increasing TM-score from 0.672 to 0.802 and pLDDT from 64.55 to 73.00. In pocket co-design, ProtLiD2 reduces active-site BB-RMSD from 3.46/3.40Å for FAIR/PocketGen to 1.97Å, and improves ligand-aware pass rates over PocketGen from 14.86% to 59.73% and from 6.08% to 23.49% under stricter docking thresholds. These results support ligand-conditioned discrete diffusion as an effective token-space framework for functional protein co-design. Code will be available at https://github.com/auroua/ProtLiD.
Chen Wei, Fanding Xu, Minghao Sun +5
Xi’an University of Posts & Telecommunications · National University of Singapore · Xi’an Jiaotong University +1
We present A-CODE, a fully atomic unified one-stage protein co-design model that simultaneously refines discrete atom types and continuous atom coordinates. Unlike predominant two-stage methods that cascade structure design with amino acid-level sequence design, our approach is fully atomic within a unified multimodal diffusion framework, in which residue identities are inferred solely from atom-level predictions. Built upon the powerful all-atom architecture, A-CODE achieves superior designability for unconditional protein generation, outperforming all existing one-stage and two-stage design models. For binder design, A-CODE rivals and even outperforms existing state-of-the-art two-stage design models and, compared with the existing one-stage co-design model, achieves a drastic tenfold improvement in success rate on hard tasks. The inherent flexibility of our atomic formulation enables, for the first time, seamless adaptation to non-canonical amino acid (ncAA) modeling. Our fully atomic framework establishes a new, versatile foundation for all-atom generative modeling that can be naturally extended to complex biomolecular systems.
Chaoran Cheng, Jiaqi Guan, Milong Ren +5
University of Illinois Urbana-Champaign · 2ByteDance Seed
Proteins are fundamental to biological processes, with their function determined by the complex interplay between the amino acid sequence and the three-dimensional structure. Developing generative models capable of understanding this intrinsically multi-modal relationship is crucial for fields like drug discovery and protein engineering. Existing models often rely on a multi-stage training process where autoencoders that tokenize data into latent representations are trained in a first stage. Secondly, a generative model is trained on the latent representation of the autoencoder(s), i.e., generative modeling in a latent space. We hypothesize that this multi-stage training is not necessary to obtain performant co-design models and thus present SimpleDesign, an effective multi-modal protein design model trained directly in the data space. SimpleDesign leverages a single-stage end-to-end objective that combines discrete cross-entropy for sequences and a regression objective for structures. In order to effectively model the difference in sequence and structure modalities, we develop a Mixture-of-Transformer architecture that allows modality-specific processing while keeping global self-attention over both modalities. We train SimpleDesign on over 2M sequence-structure pairs achieving strong performance across co-design and unconditional sequence/structure generation benchmarks.