🧬 biology

SF-Cluster: Frustration-Guided MSA Subsampling for Alternative Protein Conformation Recovery

The paper introduces SF-Cluster, a frustration-guided MSA subsampling method that significantly improves the recovery of alternative protein conformations compared to existing sequence-based approaches by leveraging predicted local energetic frustration patterns to effectively reweight MSA composition.

Original authors: Hanqun Cao, Zijun Gao, Chunbin Gu, Ge Liu, Pheng Ann Heng, Pranam Chatterjee

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanqun Cao, Zijun Gao, Chunbin Gu, Ge Liu, Pheng Ann Heng, Pranam Chatterjee

Original paper licensed under CC BY 4.0 (http://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine you are trying to predict the shape of a protein, which is like a tiny, complex machine made of a long string of beads (amino acids). For decades, scientists have used a tool called AlphaFold to do this. AlphaFold works by looking at a "family album" of the protein, known as a Multiple Sequence Alignment (MSA). This album contains thousands of slightly different versions of the same protein from various species.

The problem is that AlphaFold usually only sees the protein's most common shape (its "dominant state"). But many proteins are like chameleons; they can switch between two or more different shapes to do their jobs (like turning a signal on or off). If you just feed the whole family album to AlphaFold, it gets confused by the noise and sticks to the most common shape, missing the rare, alternative ones.

The Old Way: Sorting by "Look-Alikes"

Previously, researchers tried to find these rare shapes by sorting the family album based on sequence similarity. Imagine you have a huge pile of photos of people, and you want to find photos of people wearing red hats. The old method (called AF-Cluster) would sort the photos by how much the people look like each other (e.g., hair color, eye shape).

The paper argues this is a bad strategy. Two people might look very similar but be wearing completely different hats. Similarly, two protein sequences might look 99% identical but fold into completely different shapes. Sorting by "looks" doesn't guarantee you get the right shape.

The New Way: SF-Cluster (The "Frustration" Guide)

The authors introduce a new method called SF-Cluster. Instead of sorting the family album by how the proteins look, they sort them by how "frustrated" they are.

What is "Frustration"?
Think of a protein like a puzzle. Some pieces fit together perfectly (happy). Other pieces are forced into awkward positions where they don't quite fit well with their neighbors (frustrated).

Low Frustration: The pieces are happy and stable.
High Frustration: The pieces are stressed and unstable.

The paper suggests that these "stressed" or "frustrated" spots are the exact places where the protein is most likely to wiggle, bend, or switch shapes. It's like finding the loose hinge on a door; that's the part that moves.

How SF-Cluster Works:

Map the Stress: For every version of the protein in the family album, the computer calculates a "frustration map." It highlights which beads are stressed.
The Mosaic Selection: Instead of grouping similar-looking proteins, SF-Cluster picks a small, diverse group of proteins that cover a wide range of "stress patterns." It's like picking a group of people for a team not because they look alike, but because they have a mix of different skills and temperaments that cover all the bases.
The Result: When this carefully selected, diverse group is fed back into the prediction tool, the tool is much more likely to "see" the rare, alternative shape because the stress patterns in the group point toward it.

What They Found

The researchers tested this on 48 different proteins, including those that switch shapes, those that act as biological switches (allosteric), and those that are naturally messy (disordered).

Better Results: SF-Cluster successfully recovered the "hidden" alternative shapes much more often than the old method. For proteins that act as switches (allosteric), it improved success rates by 15.5%.
It's the Data, Not the Tool: They proved that the secret wasn't a new AI model. They took the specific groups of proteins selected by SF-Cluster and fed them into a different AI tool (Boltz-1). It worked there too! This means the "secret" was in the selection of the family album, not the computer program itself.
The Real Secret (Depth): Interestingly, when they matched the size of the groups, the advantage of SF-Cluster mostly disappeared. This suggests that the main reason it works is that SF-Cluster is very good at picking large, diverse groups of proteins that have enough "data depth" to be useful. The "frustration" map is just a very reliable way to find those deep, diverse groups.
Biological Truth: The "frustrated" spots the computer found weren't random. They lined up perfectly with real-world experiments showing where proteins actually change shape or where mutations cause diseases.

The Bottom Line

The paper claims that to find a protein's hidden shapes, you shouldn't just look for similar-looking cousins. Instead, you should look for cousins that share specific patterns of "stress" or "frustration." By using these stress patterns to select a diverse, deep group of proteins, you can guide AI tools to discover the alternative shapes that nature uses to function.

Important Limitation: The paper also notes a hard limit: if the family album only contains proteins that have one shape (no hidden shapes exist in the data), no amount of clever sorting will invent a new shape. You can't find a shape that isn't already hinted at in the family history.

Problem Statement

Deep learning-based protein structure predictors, such as AlphaFold2, are highly sensitive to their Multiple Sequence Alignment (MSA) inputs. While these models excel at predicting a single dominant conformation, recovering alternative conformations (e.g., fold-switching, allosteric states, or oligomerization-coupled states) from sequence alone remains challenging. Existing approaches like AF-Cluster attempt to recover alternative states by subsampling MSAs based on sequence similarity. However, the authors argue that sequence space provides limited guidance for selecting subsets associated with specific conformational basins, as homologs with high sequence similarity can still adopt different structural states, and distinct states may not be separable by sequence clustering alone.

Methodology: SF-Cluster

The authors introduce SF-Cluster, a framework that subsamples MSAs using predicted local energetic frustration patterns rather than sequence similarity. The core hypothesis is that frustration patterns offer a representation of conformational preference that is largely independent of sequence similarity.

Key Technical Components:

Frustration Representation: For each homolog in the MSA, a per-residue frustration profile is computed using FrustrAI-Seq (a sequence-based predictor built on ProtT5-XL). These profiles quantify how well residue interactions are satisfied relative to alternative local configurations.
Embedding and Geometry: The per-residue profiles are compressed into region-level embeddings (grouping residues into $K$ regions and calculating mean frustration, variance, and inter-region contrasts). Homologs are then positioned in a "frustration-pattern space."
Mosaic Selection: Instead of clustering, SF-Cluster employs a mosaic selection strategy. It draws fixed-size subsets (default $n=32$ ) that span diverse modes of the frustration-pattern distribution. This ensures high within-subset frustration variance, intended to reduce energetic homogeneity and expose alternative basins to the predictor.
Coverage-Aware Refinement: A refinement procedure preserves low-frequency conformations that might otherwise be discarded by standard confidence-based (pLDDT) ranking, ensuring that rare states are not lost during the selection process.
Prediction Pipeline: Selected MSAs are fed into structure predictors (AlphaFold2 and Boltz-1) using a two-stage screen-and-refine protocol.

Key Contributions

Frustration-Guided Subsampling: The paper proposes shifting MSA subsampling from sequence space to a frustration-pattern space, arguing that frustration captures conformational signals complementary to sequence similarity.
Transferability of Conformational Signal: The study demonstrates that the conformational signal resides in the MSA composition itself, not just in the specific architecture of AlphaFold2, as selected MSAs successfully transfer to the architecturally distinct predictor, Boltz-1.
Mechanistic Insight: The authors identify that the primary driver of recovery gains is the effective depth ( $N_{eff}$ ) of the selected subsets. SF-Cluster reliably generates deep, diverse subsets (saturating near $N_{eff} \approx 30$ ), whereas sequence-based clustering (AF-Cluster) often fragments alignments into shallower subsets.
Biological Validation: The paper links frustration patterns to independent experimental data, showing that highly frustrated residues overlap with Deep Mutational Scanning (DMS) functional sites and NMR two-state exchange residues. Furthermore, frustration covariation identifies state-switching contacts distinct from classical coevolutionary coupling.

Results

The method was evaluated on a benchmark of 48 cases spanning fold-switching, allosteric, oligomerization-coupled, and intrinsically disordered systems.

Performance Improvement: Using a dual-reference RMSD criterion (requiring prediction within 3 Å of the target, closer to the target than the dominant state, and pLDDT $\ge$ $\geq$ 70), SF-Cluster improved target-state recovery over AF-Cluster across all two-state classes.
- Allosteric systems: Recovery increased from 27.3% to 42.8% (+15.5 percentage points).
- Fold-switching: Recovery increased from 3.3% to 10.1%.
- Oligomerization-coupled: Recovery increased from 0.8% to 11.4%.
Cross-Predictor Validation: When applied to Boltz-1, SF-Cluster-selected MSAs recovered 4/10 target conformations, compared to 0/10 from single-sequence inputs. This confirms the signal is in the MSA, not the predictor.
Depth vs. Geometry: Controlled experiments revealed that effective subset depth is the dominant factor. Depth-matched random controls and frustration-index-shuffled controls performed similarly to SF-Cluster, suggesting that the primary advantage of SF-Cluster is its ability to reliably select deep, diverse subsets. However, frustration-pattern selection retained a residual advantage on allosteric proteins and specific fold-switching systems where the geometry aligned well with the selection strategy.
Limits: The method cannot recover alternative conformations if the underlying sequence pool encodes only a single structural basin (e.g., the GB98 T25I/L20A negative control).

Significance and Claims

The paper positions MSA subsampling as a representation-guided reweighting problem. The authors claim that:

Frustration patterns serve as a transferable representation for conformational prediction, identifying functional residues, conformational exchange regions, and state-switching contacts that are distinct from coevolutionary signals.
The success of SF-Cluster is largely explained by its ability to generate deep, diverse subsets that reach the effective depth regime ( $N_{eff} \approx 30$ ) where conformational recovery saturates, a regime often missed by sequence-based clustering.
While frustration geometry provides state-specific information in a subset of systems, the framework's primary utility is as a mechanism to ensure the selected MSA subsets are sufficiently deep and diverse to allow the predictor to sample alternative basins.
The work establishes a practical boundary: MSA subsampling can redistribute probability among conformations already represented in the sequence pool but cannot recover states absent from that pool.

The authors conclude that future improvements may lie in learned state-aware representations or generative MSA augmentation, but emphasize that the central challenge in subsampling is selecting informative subsets rather than simply generating more predictions.