General Multimodal Protein Design Enables DNA-Encoding of Chemistry
| Authors | Jarrid Rector-Brooks et al. |
| Year | 2026 |
| HF Upvotes | 30 |
| arXiv | 2604.05181 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
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(sp^3)-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.
Engineering Breakdown
Plain English
This paper introduces DISCO, a deep generative model that designs novel enzymes by co-designing protein sequences and 3D structures together, conditioned only on the reactive intermediates the enzyme should catalyze. Unlike prior protein design methods that require pre-specifying which amino acids should be catalytic residues, DISCO learns to discover novel active-site geometries from scratch. The team demonstrated the approach by designing heme enzymes that catalyze entirely new-to-nature reactions, including carbene-transfer reactions like alkene cyclopropanation and C-H insertion, showing that AI can expand the chemistry accessible to natural enzymes beyond what evolution has explored.
Core Technical Contribution
The core novelty is a multimodal diffusion model that jointly optimizes protein sequence and 3D structure in a unified latent space, rather than treating them as separate design problems. DISCO conditions on reactive intermediates (small biomolecules) rather than requiring explicit specification of catalytic residues, meaning the model learns where and how to position functional chemistry without human priors. The authors also introduced inference-time scaling methods that can optimize multiple objectives (binding, structure validity, catalytic geometry) across both sequence and structure modalities simultaneously. This represents a fundamental shift from prior generative approaches that either design sequences given fixed structures, or design structures given fixed sequences.
How It Works
DISCO encodes both protein sequences and 3D structures into joint multimodal embeddings, then uses a diffusion process that operates in this shared latent space. The model is trained on native proteins with known structures and sequences, learning the distribution of valid protein geometry and sequence patterns. At inference time, the model takes a reactive intermediate molecule as conditioning input and iteratively denoises random noise into a sequence-structure pair that places the reactive intermediate in a stereochemically appropriate position for catalysis. The inference-time scaling involves sampling multiple sequence-structure candidates, then ranking them by objectives like binding affinity, structural plausibility, and active-site geometry quality before selecting the best design.
Production Impact
For teams building synthetic biology or enzyme engineering pipelines, this approach eliminates the need for manual catalytic residue specification and iterative wet-lab validation loops, potentially reducing design-to-test cycles from months to weeks. You could integrate DISCO into a production system by: (1) taking target reactions and reactive intermediates as inputs, (2) generating enzyme candidates in batch, (3) using fast structure prediction validation to filter designs, and (4) synthesizing top candidates for lab testing. The main production trade-offs are significant computational cost (diffusion models require many denoising steps, likely GPU-intensive), uncertainty in whether AI-designed enzymes will express and fold properly in vivo despite good structure prediction, and the need for wet-lab validation infrastructure to test designed enzymes. However, the potential to access entirely new chemistry (100+ possible reactions vs. natural evolution) makes this attractive for pharma, industrial biotech, and materials science applications.
Limitations and When Not to Use This
DISCO assumes that conditioning on reactive intermediates is sufficient to drive functional enzyme design, but real enzymes require many additional properties: proper expression levels, cellular compatibility, cofactor availability, and kinetic parameters beyond binding. The paper doesn't demonstrate that these designed enzymes work in realistic cellular contexts or at the production scales needed for synthetic biology (fermentation, industrial synthesis). The approach requires substantial training data (structure-sequence pairs from protein databases), making it less applicable to entirely novel fold spaces. Additionally, diffusion-based inference is computationally expensive compared to non-generative baselines, and there's no analysis of how many AI-generated designs must be synthesized to find one that actually catalyzes the target reaction at usable rates.
Research Context
This work builds on recent advances in protein language models (e.g., ProtBERT, ESM) and structure prediction (AlphaFold), extending those to the joint sequence-structure design problem. It improves upon prior conditional protein design methods (like RoseTTAFold Diffusion and OmegaFold-based approaches) by unifying sequence and structure in a single generative model rather than alternating between them. The paper contributes to the emerging field of de novo enzyme design guided by AI, following earlier work by Baker, Simonson, and others, but with the key advance of not requiring pre-specified catalytic residues. This opens a research direction: can similar multimodal diffusion approaches be applied to other biomolecular design problems (antibodies, binders, RNA), and how can we make inference fast enough for real-time design workflows?
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::