Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time

This paper introduces ConforMix, an inference-time algorithm that enhances diffusion models to efficiently sample and discover diverse biomolecular conformational landscapes, including domain motions and cryptic pockets, without requiring prior knowledge of degrees of freedom or retraining.

The Gist

Imagine you are trying to understand how a complex machine, like a Swiss Army knife, works. You know it has a handle and many tools (blades, screwdrivers, scissors) that fold in and out.

For a long time, scientists have used powerful computer programs to predict what this machine looks like when it is completely closed and sitting on a table. These programs are great at finding that one "default" shape. But biology isn't static; proteins (the machines of life) are constantly folding, unfolding, twisting, and shifting to do their jobs. The problem is that the computer programs usually only show you the "closed" version, missing all the other cool ways the machine can move.

This paper introduces a new tool called ConforMix. Think of it as a "smart guide" that you can attach to these computer programs after they have already been trained. It doesn't require re-teaching the computer; it just changes how the computer looks for answers.

Here is how ConforMix works, using some simple analogies:

1. The Problem: The "Lazy" Search

Imagine the computer program is a hiker trying to find the lowest point in a vast, foggy valley (the "energy landscape"). The hiker's goal is to find the deepest, most comfortable spot to sleep (the most stable protein shape).

• The Default Method: The hiker walks straight down the steepest hill they see and stops at the first valley they find. They might find a nice spot, but they miss the other valleys nearby that are just as good, or even better for specific tasks. They only see one version of the machine.
• The Old Way to Find More: Previously, to find other shapes, scientists had to give the computer different "instructions" or "clues" (like changing the input data) to force it to look elsewhere. This was like telling the hiker, "Go look over there," but it often resulted in the hiker getting lost or finding weird, broken shapes that don't make sense physically.

2. The Solution: ConforMix (The "Guided Tour")

ConforMix is like giving the hiker a magnetic compass and a map that gently pushes them toward new areas without forcing them off the path.

• The "Twisted" Path: Instead of just walking straight down, ConforMix uses a technique called "Twisted Diffusion." Imagine the hiker is walking through a field of tall grass. The default method just walks forward. ConforMix gently nudges the hiker to the left or right to explore different parts of the field, but it keeps checking to make sure they are still walking on solid ground (physically possible shapes).
• The "RMSD" Ruler: One specific version of this tool, called ConforMixRMSD, uses a ruler to measure how different a new shape is from the original "closed" shape. It says, "Okay, let's find a shape that is slightly different, then one that is very different, then one that is really different." It systematically explores the range of motion, from a tiny wiggle to a full flip.

3. What They Found

The researchers tested this on a model called Boltz-1 (which is similar to the famous AlphaFold 3).

• Finding Hidden Doors: They found that even though the computer model was only trained to predict the "closed" state, it actually "knew" about the open states deep inside its logic. ConforMix unlocked these hidden doors.
• Real-World Examples:
  • The Transporter: They looked at a protein that acts like a door, opening to the inside of a cell and then to the outside. The default computer only saw the door closed. ConforMix found the door open both ways, plus the "in-between" states.
  • The RNA: They even tried it on RNA (a cousin of DNA), and it successfully predicted how a small RNA molecule unzips and zips back up, matching what scientists see in high-speed simulations.
• Better than Guessing: They compared ConforMix to methods that try to trick the computer by changing the input data. ConforMix was better at finding the right kind of movement (like a hinge opening) rather than just random, messy jiggling.

4. The Bonus: Calculating the "Cost" of Movement

Beyond just finding shapes, ConforMix helps calculate Free Energy.

• The Analogy: Imagine you want to know how much effort it takes to push a heavy door open. You don't just want to see the door open; you want to know the "cost" of that movement.
• The Result: ConforMix allows scientists to estimate these costs much faster than before. It's like having a speedometer that tells you how hard it is for a protein to change shape, which helps explain how well it might bind to a drug or perform a function.

Summary

In short, ConforMix is a "plug-and-play" upgrade for protein-predicting AI. It takes a model that usually only sees one static picture and turns it into a tool that can explore the entire movie of the protein's life. It finds hidden shapes, explains how proteins move, and does it all without needing to retrain the AI or having prior knowledge of what the protein is supposed to do. It's like giving a static photo album the ability to show you the full dance routine.

Technical Summary

Technical Summary: Unlocking Hidden Biomolecular Conformational Landscapes in Diffusion Models at Inference Time

Problem Statement
Biomolecular function relies on the ability of proteins and other molecules to interconvert between a wide range of structures, or "conformations." While experimental methods struggle to capture these dynamic distributions compared to static folded structures, computational methods like molecular dynamics (MD) and Monte Carlo sampling have historically been limited by speed and accuracy. Recently, machine learning diffusion models (e.g., AlphaFold 3, Boltz-1) have revolutionized static structure prediction. However, these models often produce highly concentrated distributions around a single static structure, failing to efficiently sample the broader conformational landscape or rare states (e.g., unbound states of tight-binding proteins) without prior knowledge of specific degrees of freedom. The central question addressed is: Given a pretrained generative structure model, what strategy should be deployed at inference time to efficiently and accurately sample its conformational landscape?

Methodology: ConforMix
The authors introduce ConforMix, an inference-time algorithm designed to enhance sampling in diffusion-based biomolecular structure prediction models without requiring additional model training. The method combines three core components:

1. Twisted Diffusion Sampling (Sequential Monte Carlo): Instead of standard i.i.d. sampling or classifier guidance (which requires retraining or architectural changes), ConforMix employs Twisted Diffusion Sampling. This acts as a particle filter that uses time-independent guidance functions to steer the denoising process toward specific regions of the probability distribution. It provides an asymptotic guarantee of sampling the true target conditional distribution.
2. Automated Exploration (ConforMixRMSD): To enable undirected exploration without user-defined bias potentials, the authors propose ConforMixRMSD. This instantiation biases sampling away from the default prediction ($x_d$) using a Root Mean Square Deviation (RMSD) metric.
   • It generates samples across a range of target RMSD values (0–20 Å) relative to the default structure.
   • To avoid sampling unphysical fluctuations in disordered loops, it applies a "rigid-element mask," computing RMSD only on secondary structure elements ($\alpha$-helices and $\beta$-sheets) while allowing all atoms to move.
   • A filtering step rejects physically implausible samples (e.g., those with significant clashes or low pLDDT in sliding windows).
3. Sample Reweighting and Free Energy Estimation: Since samples are drawn from conditional distributions, the authors utilize the Multistate Bennett Acceptance Ratio (MBAR) algorithm to reweight the mixture of conditional samples back to the unconditional distribution. This allows for the reconstruction of the unbiased free energy landscape and the estimation of free energy differences, a capability previously unapplied to diffusion models in this context.

Key Contributions

• Novel Algorithm: The implementation of a twisted sequential Monte Carlo approach combined with automated landscape exploration and statistically optimal sample reweighting (MBAR) for diffusion models.
• Improved Inference-Time Sampling: Demonstration that ConforMix can be applied to models like Boltz-1 (an AlphaFold 3-like model trained only on static structures) to generate realistic, diverse conformations without modifying input data (e.g., Multiple Sequence Alignments) or requiring prior knowledge of degrees of freedom.
• Efficient Free Energy Estimation: The application of ConforMix to BioEmu (a model trained for conformational sampling) to accelerate the convergence of free energy estimates compared to default sampling.

Results
The authors evaluated ConforMix across several benchmarks:

• Recovery of Experimental Conformations: On datasets involving domain motions, membrane transporters, cryptic pockets, and fold-switching proteins, ConforMixRMSD outperformed default sampling and MSA-modification baselines (AFCluster, AFSample2, CF-random). It recovered a higher fraction of experimentally observed alternate conformations (e.g., 0.69 coverage for domain motions vs. 0.33 for default Boltz).
• Biological Relevance of Sampling: Principal Component Analysis (PCA) revealed that ConforMixRMSD samples concentrate variance along biologically meaningful transition paths (e.g., the open-to-closed motion of the dipeptide binding protein dppA), whereas MSA-based methods often produced diffuse, uninterpretable structural noise.
• Multi-Chain and RNA Systems: The method successfully recovered all three states of the SemiSWEET sugar transporter (inward-open, outward-open, occluded) and recapitulated MD-observed transitions in the hammerhead ribozyme, demonstrating applicability beyond single-chain proteins.
• Free Energy Convergence: In BioEmu, ConforMix enabled significantly faster convergence of free energy estimates for local unfolding transitions compared to unbiased sampling, achieving accurate estimates with fewer samples.

Significance and Claims
The paper claims that ConforMix is orthogonal to improvements in model pretraining, meaning it enhances the utility of existing models and would benefit even a hypothetical model that perfectly reproduces the Boltzmann distribution. The authors argue that while perfect sampling of the Boltzmann distribution is a long-term goal, enhanced sampling algorithms are still necessary to efficiently explore rare states (e.g., unbound structures with probabilities as low as $10^{-14}$) and characterize transition mechanisms.

The significance lies in unlocking "hidden" conformational landscapes within models trained for static prediction, enabling the discovery of domain motions, cryptic pockets, and transporter cycling without user intervention. The authors position this as a step toward improved understanding of basic biology and potential applications in medicine, noting that the method is flexible, applicable to various biomolecular systems (proteins, RNA, complexes), and operates purely at inference time.