💻 bioinformatics

How do Co-Folding Models Organize Structural Information?

This paper dissects the Boltz-1 co-folding model to reveal that structural information is organized across three distinct streams—single, intra-chain, and inter-chain representations—that follow a Mix-Compress-Refine trajectory, where intra-chain geometry is largely pre-conditioned while inter-chain arrangements are progressively constructed and reconciled by the diffusion module.

Original authors: Park, M., Kim, S., Moon, S., Kim, H., Jeon, G., Kim, W. Y.

Published 2026-09-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Park, M., Kim, S., Moon, S., Kim, H., Jeon, G., Kim, W. Y.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the microscopic world of life, the shape of a molecule dictates its function. Proteins, the workhorses of biology, are long chains of amino acids that must twist and fold into precise three-dimensional structures to perform their jobs. For decades, scientists have worked to predict these shapes from the genetic code alone, a challenge that has recently been solved for single proteins with remarkable accuracy. However, life rarely happens in isolation. Proteins often work in teams, binding with other proteins or with small drug-like molecules to form complex machines. Predicting how these separate parts come together to form a stable, functional assembly is a much harder puzzle. It requires understanding not just how one chain folds, but how multiple chains arrange themselves relative to one another, and how a drug molecule might force a protein to change its shape to accommodate it.

A new study by researchers at the Korea Advanced Institute of Science and Technology and HITS investigates the inner workings of a cutting-edge computer model designed to solve this complex problem. This model, known as a co-folding system, attempts to predict the entire structure of a molecular complex in one go, rather than folding the pieces separately and then trying to stick them together. The researchers wanted to understand exactly how the model's internal "brain" organizes the information needed to solve this task. They peered inside the model's layers to see where it stores details about individual atoms, where it keeps track of the shape of a single chain, and where it holds the information about how different chains fit together.

The team focused their investigation on a specific model called Boltz-1, which is part of a new generation of tools that have revolutionized structural biology. They treated the model's internal data streams like a set of distinct channels, each carrying a different type of information. By systematically turning off or scrambling specific parts of these channels and watching how the model's predictions changed, they mapped out the model's internal logic. They found that the model does not mix all its information into a single, chaotic soup. Instead, it separates the task into three distinct streams. One stream carries the fine-grained details of individual atoms within a single amino acid. A second stream is dedicated to the overall geometry of a single protein chain, ensuring it folds into the correct shape. The third stream is responsible for the relative positioning of different molecules, acting as the glue that determines how one chain sits next to another.

To understand how this information is built up as the model processes a sequence, the researchers watched the data evolve layer by layer. They discovered a clear pattern in how the model constructs its understanding of distance between parts of the molecule. For the parts of the protein that belong to the same chain, the model relies heavily on evolutionary data provided at the very beginning of the process. This initial data, derived from comparing the genetic sequences of many similar proteins across nature, gives the model a strong head start on how a single chain should fold. However, the information about how different chains interact is not present at the start. Instead, the model builds this understanding progressively, refining the arrangement of separate molecules step by step as it moves through its deeper layers. This suggests that while the model uses nature's evolutionary history to solve the folding of individual parts, it must actively learn to solve the puzzle of how those parts fit together during the calculation itself.

The researchers also examined how well the model handles cases where a molecule changes shape significantly upon binding to a partner, such as a protein that shifts its structure when a drug attaches to it. They found that the model's internal representation is not always perfect before the final step of generating the 3D coordinates. Sometimes, the model holds conflicting instructions: the internal map of a single chain might suggest one shape, while the map of how chains interact suggests another. The final stage of the model, which generates the physical coordinates, acts as a reconciler. It takes these partially inconsistent instructions and smooths them out to produce a final, physically plausible structure. This reveals that the model does not always have a complete, conflict-free picture of the final shape before it starts drawing it; rather, it uses the final step to resolve the tensions between different pieces of information.

These findings offer a clear mechanical view of how modern artificial intelligence organizes structural information. The study shows that these models are not black boxes that simply output a result; they are structured systems that separate local details, chain-level geometry, and inter-molecular arrangement into specialized pathways. The research suggests that future models could be designed more efficiently by respecting these natural divisions, perhaps by giving different parts of the network specific roles rather than asking a single system to do everything at once. By understanding exactly where and how these models store the secrets of molecular shape, scientists can build better tools to design drugs and understand the machinery of life.

Technical Summary: How Do Co-folding Models Organize Structural Information?

Problem Statement
While co-folding models (e.g., AlphaFold3, Boltz-1) have achieved high accuracy in predicting biomolecular complex structures from sequence, the internal mechanisms by which they organize and process structural information remain unclear. Specifically, it is unknown how these models distinguish between intra-chain folding and inter-chain docking, how pairwise distance information is constructed across network layers, and whether the trunk (the representation generator) fully resolves conformational changes induced by ligand binding before passing information to the diffusion module. Unlike single-chain folding, co-folding introduces inter-chain pair blocks with no direct analogue in monomeric systems, rendering existing interpretability methods insufficient.

Methodology
The authors analyze the "Boltz-1" trunk architecture using a chain-aware factorization of the pair representation into intra-chain ( $Z_{intra}$ ) and inter-chain ( $Z_{inter}$ ) blocks. The study employs four analytical lenses:

Ablation: Activations in the single representation ( $S$ ), $Z_{intra}$ , and $Z_{inter}$ are zeroed, averaged, or replaced with Gaussian noise. The modified representations are decoded by a frozen diffusion module to measure the impact on chain-local geometry, inter-chain arrangement, and protein-ligand contacts.
Representation Geometry: Singular-value statistics (matrix entropy, stable rank, top-1 singular value mass) are computed for activation matrices at every layer and recycling iteration to characterize the flow of information (Mix–Compress–Refine trajectory).
Linear Probing: Low-capacity linear decoders are fitted to hidden activations to quantify the linear accessibility of pairwise $C\beta$ – $C\beta$ distances at different depths for both $Z_{intra}$ and $Z_{inter}$ .
Activation Patching: Activations from "apo" (ligand-unbound) and "holo" (ligand-bound) inference runs are swapped to test whether the trunk fully encodes ligand-induced backbone changes or if the diffusion module compensates for inconsistencies.

Key Results

Functional Specialization of Representations:
- Single Representation ( $S$ ): Contains token-internal atomic details. Ablating $S$ (while preserving statistics) maintains chain-level geometry but results in chemically invalid local structures (e.g., incorrect proline ring geometries and side-chain distances).
- Intra-chain Pairs ( $Z_{intra}$ ): Encodes chain-level geometry. Ablation causes a catastrophic drop (~70–78%) in chain-LDDT, indicating this stream is the primary carrier of intra-chain folding information.
- Inter-chain Pairs ( $Z_{inter}$ ): Controls relative molecular arrangement. Ablation preserves individual chain folds but destroys interface-LDDT and protein-ligand contact quality, confirming its role in docking.
Construction of Pairwise Information:
- Asymmetric Preconditioning: Intra-chain distance content is largely supplied by the MSA module early in the process (reaching $R^2 \approx 0.3$ before Pairformer entry and $\approx 0.8$ after recycling). In contrast, inter-chain distance content is constructed progressively across Pairformer layers in every recycling iteration, starting with near-zero correlation.
- Mix–Compress–Refine (MCR) Trajectory: The Pairformer trunk follows an MCR trajectory similar to large language models: the MSA module mixes information (entropy drop), middle layers compress the representation, and later layers re-expand toward the diffusion handoff. Crucially, distance accessibility is preserved even during compression phases.
Ligand-Induced Changes:
- Patching experiments reveal that the trunk does not fully resolve ligand-induced backbone changes before the diffusion stage. When $Z_{intra}$ from a holo state is patched into an apo representation, the diffusion module must reconcile partially inconsistent intra- and inter-chain constraints to generate the correct structure.

Key Contributions

Decomposition of Information Streams: The paper demonstrates that co-folding models encode distinct aspects of information in three separate streams: local atomic details in $S$ , chain-level geometry in $Z_{intra}$ , and relative orientation in $Z_{inter}$ .
Depth Profile of Pair Representations: The authors map the evolution of pairwise distance information, showing that intra-chain signals are MSA-saturated and refined, while inter-chain signals are actively constructed across the trunk depth.
Trunk-Diffusion Interaction: The study provides evidence that the trunk's representations can be incomplete regarding large conformational changes, with the diffusion module playing a critical role in reconciling constraints that the trunk has not fully resolved.

Significance
This work provides a mechanistic view of how co-folding trunks organize structural information, moving beyond black-box performance metrics. By identifying that inter-chain information is built progressively and that the trunk may not fully resolve ligand-induced changes, the findings suggest that future biomolecular structure prediction models could benefit from "partition-aware" designs. Such designs might explicitly separate the handling of intra-chain co-evolutionary signals from inter-chain interaction construction, potentially improving efficiency and interpretability.

Technical Summary: How Do Co-folding Models Organize Structural Information?

More like this