💻 bioinformatics

ConforFlux: Particle-Guided Trunk Repulsion for Diverse Protein Conformations

ConforFlux is an inference-time procedure that enhances the diversity of protein conformational sampling in deep learning models like Boltz-2 by introducing a particle-guided repulsion mechanism on trunk embeddings, significantly improving the discovery of multiple states without requiring model retraining.

Original authors: Suzuki, S., Amagasa, T.

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Suzuki, S., Amagasa, T.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Proteins are the workhorses of life, tiny molecular machines that fold into intricate three-dimensional shapes to perform tasks like digesting food, sending nerve signals, or building cells. For decades, scientists believed that to understand how a protein works, they only needed to see its final, resting shape, much like taking a single photograph of a person to understand their personality. However, biology is rarely static. Proteins are dynamic; they breathe, twist, and shift between different shapes to function. An enzyme might need to open a pocket to grab a molecule, and a transporter might need to flip from facing the inside of a cell to facing the outside to move substances across a barrier. These different shapes, or conformations, are often just as important as the structure itself, but capturing them in a laboratory is difficult, expensive, and often incomplete.

In recent years, artificial intelligence has revolutionized the ability to predict these protein shapes, achieving accuracy that rivals laboratory experiments. Yet, a significant limitation has emerged: when these powerful computer models are asked to predict a structure, they tend to produce the same answer over and over again, focusing on the single most likely shape and ignoring the other valid forms the protein might take. This is like a camera that only ever takes a picture of a person standing still, never capturing them walking, sitting, or waving, even though those movements are essential to who they are. Researchers have long sought a way to force these models to explore the full range of a protein's movements without needing to retrain the massive, complex systems from scratch.

A team of researchers at the University of Tsukuba has developed a new method called ConforFlux to solve this problem. Instead of trying to teach the AI model new things or changing its internal code, they introduced a simple but clever rule that guides the computer's thinking while it is working. Imagine the computer running several different simulations of the same protein at the exact same time. In the standard approach, each simulation runs independently, often all landing on the same final shape. The new method connects these parallel simulations, allowing them to "talk" to one another. If two simulations start to produce structures that look too similar, the system gently pushes them apart, encouraging each one to explore a different path. This repulsion is applied not to the final 3D shape, but to the internal mathematical representations the model uses to understand the protein's sequence. By nudging these internal views, the method ensures that the final predicted structures are diverse, covering a wider range of the protein's possible movements without ever altering the model's original training.

The researchers tested this approach on a wide variety of proteins known to change shape, including those that act as doors in cell membranes and enzymes that switch between different functional states. They compared the new method against the standard way of using the AI model, as well as against other existing techniques that try to force diversity. The results were clear: the new method successfully uncovered alternative shapes that the standard model missed entirely. In tests involving membrane transporters, which cycle between inward and outward facing states, the new method improved the success rate of finding both states by up to 17 percentage points. It was particularly effective at finding the "hidden" states of proteins that are difficult to capture, such as those with temporary pockets that open only briefly. In one specific case involving the human dopamine transporter, a protein critical for brain function, the standard method failed to find the inward-facing shape in 500 attempts, while the new method found it dozens of times.

The study also looked at how well this technique worked on other leading AI models, not just the one it was designed for. The method proved to be versatile, improving the diversity of predictions on two other major protein-prediction systems as well. This suggests that the strategy of guiding the internal representations of the model is a general solution that can be applied across different platforms. The researchers did note a trade-off: the new method sometimes produced structures with slightly more minor geometric imperfections, such as atoms being slightly too close together, but these were manageable and did not obscure the main benefit of finding the correct shapes. Crucially, the method worked even on proteins whose structures were discovered after the AI models were originally trained, proving that it could access biological realities the models had never seen before.

By allowing these powerful AI tools to explore the full spectrum of a protein's life, rather than just its most common pose, this work opens new doors for understanding how biological machines function. It provides a way to generate a library of possible shapes for a single protein, giving scientists a more complete picture of how these molecules move and interact. This is not just a technical improvement in computer science; it is a step toward understanding the dynamic nature of life itself, showing that even the most advanced artificial intelligence can be guided to see the world in motion, not just in stillness.

Technical Summary: ConforFlux

Problem Statement

Deep learning models for protein structure prediction, such as AlphaFold 3 and Boltz-2, have achieved near-experimental accuracy for individual protein folds. However, a critical limitation remains: their default inference pipelines produce samples that concentrate around a single dominant conformation. This behavior fails to capture the conformational diversity essential for protein function, such as the transitions between catalytic states in enzymes, the inward/outward cycling of membrane transporters, and the transient exposure of cryptic binding pockets. While experimental resolution of these alternate states is costly and often incomplete, existing computational approaches to broaden conformational coverage—such as perturbing Multiple Sequence Alignments (MSA), modifying internal representations, or training separate ensemble generators—often require additional training, specific experimental restraints, or yield inconsistent results across different model architectures.

Methodology: ConforFlux

The authors introduce ConforFlux, an inference-time procedure designed to diversify sampling without retraining the underlying model or altering its weights. The method is built upon the observation that in models like Boltz-2, OpenFold3, and Protenix, a "trunk" (comprising an MSA module and a Pairformer stack) generates single ( $\bar{s}$ ) and pair ( $\bar{z}$ ) representations that condition a separate diffusion module. These representations are computed once per recycling iteration and influence every subsequent denoising step.

ConforFlux couples parallel sampling trajectories by applying a pairwise C $\alpha$ -RMSD repulsion gradient directly to the trunk's single and pair embeddings. The procedure operates as follows:

Parallel Trajectories: $M$ particles (sampling trajectories) are initialized. Each particle receives a copy of the trunk's output representations $(s^{(m)}, z^{(m)})$ .
Denoising and Repulsion: At specific intervals during the reverse-diffusion process, the diffusion module generates candidate structures $\hat{x}^{(m)}_0$ for each particle.
Pairwise Objective: A pairwise repulsion loss $L$ is calculated based on the C $\alpha$ -RMSD between all pairs of decoded structures. The loss utilizes a Radial Basis Function (RBF) kernel:
$L_{ij} = \exp\left(-\frac{d_{ij}^2}{2\sigma^2}\right)$
where $d_{ij}$ is the Kabsch-aligned RMSD between particles $i$ and $j$ , and $\sigma$ is a bandwidth parameter.
Trunk-Level Descent: The loss $L$ is back-propagated to update the trunk embeddings $(s^{(m)}, z^{(m)})$ for each particle. The update step employs RMS-normalized descent with an adaptive scaling factor $\lambda_t$ that accounts for both the maximum pairwise repulsion and the current noise level of the diffusion schedule.
Re-conditioning: The updated embeddings are used to recompute the conditioning signals (AdaLN scale-shift and attention pair-bias) for the diffusion module, effectively steering the subsequent denoising steps toward distinct conformations.

This approach adapts Particle Guidance (Corso et al., 2024) to the conditioning representations of the diffusion model, requiring no additional training and no target-state experimental restraints.

Key Contributions

The paper makes three primary contributions:

Method: A trunk-level particle-guidance procedure for Boltz-2 (and applicable to OpenFold3 and Protenix) that diversifies sampling by coupling trajectories through their conditioning representations. It requires no model retraining.
Benchmark Comparison: A controlled evaluation against nine baselines (including MSA perturbations, coordinate-level guidance, and trained ensemble generators) across four conformational-change categories: domain motion, membrane transporter cycling, cryptic pocket formation, and fold switching.
Post-Cutoff Evaluation: An assessment of the method's ability to recover experimental reference structures released after the models' training cutoffs, demonstrating generalization to unseen data.

Results

The evaluation was conducted on a combined benchmark of 121 target-category pairs, including an expanded set of membrane transporters.

Coverage Improvement: ConforFlux significantly improves the per-state success rate (fraction of reference states reached within specific RMSD thresholds) compared to default Boltz-2. Improvements ranged from 3 to 17 percentage points across categories.
- On membrane transporters, ConforFlux achieved the highest gains, reaching 28 reference states missed by default sampling.
- On cryptic pockets, ConforFlux achieved the highest success rates among all methods.
- On fold switching, gains were modest but consistent with other inference-time baselines.
Dual-State Recovery: The RC50 metric (fraction of targets where both reference states are reached within half their inter-reference RMSD) improved across all categories. For example, on the transporter category, RC50 increased from 0.34 (default) to 0.60 (ConforFlux).
Generalization: The method successfully transferred to OpenFold3 and Protenix, showing consistent improvements in per-state success and RC50 across these architectures.
Post-Cutoff Performance: On targets with reference structures released after the models' training cutoffs, ConforFlux improved the recovery of both states for all three models. For instance, on the human dopamine transporter (DAT), a post-cutoff target, ConforFlux sampled inward conformations that were absent from 500 default samples.
Trade-offs: The increased coverage comes with a trade-off in structural quality. ConforFlux samples exhibit higher MolProbity clashscores and a slightly higher fraction of samples with abnormal consecutive-residue C $\alpha$ distances compared to default sampling. However, these values remain comparable to or better than other intervention methods like CF-random.

Significance and Claims

The paper claims that ConforFlux provides a structured, training-free mechanism to access alternative protein conformations by leveraging the coupling of sampling trajectories at the level of the model's conditioning representations.

Mechanistic Insight: The study demonstrates that the "trunk" representations are effective guidance sites. By updating these embeddings via back-propagation of a pairwise repulsion, the model is steered away from a single dominant mode without needing to specify a target conformation or retrain the network.
Controlled Baselines: The authors show that matched-noise controls (isotropic updates without repulsion) fail to reproduce coverage gains, confirming that the direction of the repulsion is critical. Furthermore, coordinate-level guidance (applying repulsion directly to atoms) yields lower fill ratios and more geometric outliers than trunk-level guidance, suggesting that guiding the latent conditioning space is more effective for maintaining structural integrity while diversifying samples.
Practical Utility: The method enables the discovery of functionally relevant states (e.g., inward-facing transporters, open cryptic pockets) that are missed by standard inference, offering a computationally efficient tool for mechanistic studies and structure-based design without the overhead of training new ensemble generators.

The authors remain modest regarding fold-switching capabilities, noting that gains in this category are limited, likely due to the inherent difficulty of predicting large-scale structural rearrangements. They also caution that state frequencies under repulsion should not be interpreted as equilibrium populations, as the method prioritizes diversity over thermodynamic accuracy.