Unified Sequence and Structure Co-Design for Biomolecules

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2171 papers271 conceptsfrozen 2026-10-04

In the traditional pipeline of computational biology, designing a functional molecule like a protein or an antibody is a two-act play. First, a researcher designs a three-dimensional "scaffold" (the structure). Second, they use an inverse-folding algorithm to find a string of amino acids (the sequence) that might fold into that shape.

This cascaded approach suffers from a fundamental "inverse folding gap": it is remarkably easy to generate a beautiful geometric structure that no physical sequence of amino acids can actually inhabit. As Alex Morehead and colleagues argue in MMDiff, separating these steps ignores the physical constraints governing how sequences prefer to fold. To solve this, a new generation of models is moving toward co-design—the simultaneous generation of sequence and structure within a single, unified generative process.

The Intuition of Joint Distributions

The core insight of co-design is that sequence and structure are not independent variables; they are two facets of the same physical reality. In nature, a protein's sequence determines its structure, and its structure dictates its function.

By modeling the joint distribution of sequences and structures, researchers like Andrew Campbell and Jason Yim have shown that models can capture mutual dependencies that cascaded pipelines miss. In their work on Multiflow, they demonstrate that generating discrete sequences and continuous structures together achieves significantly higher "designability" scores. The sequence is essentially "born" into the structure, ensuring that the chemical identity and spatial configuration are optimized in a mathematically consistent process.

Bridging the Discrete-Continuous Divide

The primary technical challenge of co-design is that sequences are discrete (categorical tokens like 'Alanine' or 'Guanine'), while structures are continuous (3D coordinates or torsion angles). Recent research has converged on two main strategies to bridge this gap:

  1. Multimodal Diffusion and Flow Matching: Models like PPFLOW by Haitao Lin and FlowDesign by Jun Wu use parallel generative flows. While one part of the model denoises the 3D atomic coordinates, another part simultaneously denoises the discrete amino acid identities. By coupling these processes, the model ensures the "skin" of the molecule (its surface) and its "skeleton" (the backbone) are physically consistent.
  2. Unified Tokenization: An alternative approach, exemplified by DPLM-2 from Xinyou Wang and Zaixiang Zheng, treats 3D coordinates as discrete tokens. By concatenating sequence tokens and structural tokens into a single string, a standard Transformer can learn the "grammar" of how specific shapes and sequences co-occur.

Why Co-Design Works: Biophysical Consistency

The benefit of co-design is most visible in high-precision tasks like antibody and peptide engineering. In PepFlow, Jiahan Li and colleagues found that modeling side-chain torsion angles and residue types together captures the fine-grained atomic interactions necessary for high binding affinity.

When sequence and structure are generated in isolation, the resulting designs often contain "geometric outliers"—atoms that are too close together or bonds that are physically impossible. Co-design models, such as AbX developed by Tian Zhu and Milong Ren, incorporate "biological guardrails" directly into the joint generation. This ensures that the amino acid sequence is not just a random string, but one that is physically capable of folding into the predicted 3D geometry.

Expanding the Frontier: RNA and Beyond

While much of this work began with proteins, the co-design paradigm is rapidly expanding. Runze Ma introduced RiboFlow, which applies these principles to RNA. Because RNA is highly flexible, its sequence and structure are even more tightly coupled than in proteins. Ma demonstrated that simultaneous generation leads to a 50% increase in structural validity compared to multi-step methods.

Similarly, Daniel Flam-Shepherd has pushed the boundaries by representing proteins as atom-by-atom strings, allowing models to co-design proteins alongside synthetic drug payloads. This suggests that the future of molecular design lies in "omni-design" frameworks—systems that don't just pick a sequence or a shape, but generate a complete, physically viable, and functional biological system in a single pass.

Papers in this version

Showing 60 of 171 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 246 papers
  2. Version 1 · 2026-10-04 · 171 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_joint_sequence_structure_co_design_e2aca53b,
  title = {joint sequence-structure co-design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-sequence-structure-co-design/rdr_e2aca53b910a4df1b930fc438eda5583},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_unified_sequence_and_structure_co_design_e2aca53b_v1,
  title = {Unified Sequence and Structure Co-Design for Biomolecules},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unified-sequence-and-structure-co-design-for-biomolecules/rdr_e2aca53b910a4df1b930fc438eda5583/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}