Use SF-Cluster Online

SF-Cluster: Frustration-Guided MSA Subsampling for Alternative Protein Conformations

SF-Cluster is a method that performs frustration-guided subsampling of multiple sequence alignments (MSAs) so that deep-learning structure predictors such as AlphaFold2 can recover alternative protein conformations. It was developed by Hanqun Cao, Zijun Gao, Chunbin Gu, Ge Liu, Pheng Ann Heng and Pranam Chatterjee, with affiliations including the University of Illinois at Urbana-Champaign, the University of Pennsylvania and The Chinese University of Hong Kong, and was described in a June 2026 arXiv preprint.

Structure predictors are sensitive to their MSA input but tend to favor a protein's dominant conformation, which makes fold-switching, allosteric, oligomerization-coupled and disordered systems hard to model. AF-Cluster addresses this by subsampling MSAs according to sequence similarity, but the authors note that this gives little control over which conformational basin gets sampled.

SF-Cluster instead groups homologs by predicted local energetic frustration patterns, aiming for subsets that span diverse frustration modes.

Why Frustration-Guided Subsampling Matters

Alternative conformations underlie allostery, fold switching and regulation, yet single predictions usually show only one state. By selecting MSA subsets using frustration rather than sequence similarity, SF-Cluster offers a different handle for steering predictions toward other conformational basins. The authors also report that highly frustrated residues align with functional sites from deep mutational scanning and with NMR exchange residues, supporting frustration patterns as a transferable representation for conformational prediction.

How SF-Cluster Works

The Pipeline

1. Frustration profiling: a per-residue frustration index is computed for each homolog in the MSA with FrustrAI-Seq.

2. Region-level embedding: residues are grouped into K regions, and each homolog is described by per-region mean frustration, within-region variances and between-region contrasts.

3. Mosaic subset selection: M fixed-size subsets (n = 32 as deployed) are drawn to span diverse modes of the frustration-pattern distribution, with the query sequence prepended to each.

4. Screen stage: each subset is run with one model and two seeds, with templates, relaxation and dropout disabled, and scored by the mean pLDDT of its best structure.

5. Refine stage: the top subsets, K = max(4, min(16, ceil(0.2 x number of subsets))), are re-run with five models and four seeds and three recycling iterations, then aggregated into a final ensemble.

An optional coverage-aware refinement with a per-basin quota is described in the supplement but was not used for the main benchmark.

Performance and Results

Target-State Recovery Versus AF-Cluster

The benchmark covers 33 two-state cases (11 fold-switching, 12 allosteric, 10 oligomerization-coupled) plus 15 single-state intrinsically disordered cases, scored separately. A two-state success required full-chain C-alpha RMSD of 3 A or less to the target, closer to the target than to the dominant reference, and mean pLDDT of at least 70.

The allosteric gain of 15.5 percentage points was the largest.

Second Predictor and Depth Controls

On a ten-protein subset with Boltz-1, single-sequence input recovered 0 of 10 targets, SF-Cluster MSAs recovered 4 of 10 and depth-matched random MSAs recovered 3 of 10. Mean Neff80 rose from 16.6 (AF-Cluster) to 30.2 (SF-Cluster), and the authors conclude that much of the advantage comes from effective subset depth.

Experimental Support

Across eight ProteinGym deep mutational scanning assays, highly frustrated residues were enriched 3.29-fold on average at functional sites. Pooled recall of NMR exchange residues across four proteins was 62.5%.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with SF-Cluster on Tamarind Bio

• Allosteric systems: sample alternative states of proteins where ligand or partner binding shifts the conformation.

• Fold-switching proteins: probe proteins such as KaiB and RfaH, which the authors analyze as examples.

• Disordered regions: the benchmark includes intrinsically disordered cases, reported separately.

• Hypothesis generation: highlight frustrated residues that tend to coincide with functional sites to guide mutational experiments.

How to Use SF-Cluster on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select SF-Cluster.

2. Provide the query sequence: enter the protein sequence whose alternative conformations you want to explore.

3. Supply or generate an MSA: the method takes an MSA of homologs as input; follow the options the tool page offers.

4. Set parameters: the paper deploys a subset size of 32, with a screen stage and a refine stage; adjust only what the interface exposes.

5. Run the job: submit and wait for subset predictions and refinement to finish.

6. Download outputs: collect the predicted structures and the refined ensemble.

7. Interpret results: compare ensemble members by pLDDT and by RMSD to known reference states.

8. Use downstream: prioritize residues or states for experimental follow-up.

Things to Keep in Mind

• Fold-switching recovery remains low (0.101 in the benchmark).

• Most of the aggregate advantage is explained by effective subset depth; frustration geometry adds a residual benefit only on some allosteric and fold-switching systems.

• Subsampling cannot recover states absent from the MSA, as seen with Mpt53.

• Frustration does not predict state populations (r = 0.06 with NMR occupancies), and the DMS signal is not universal (no enrichment for calmodulin or alpha-synuclein).

• Only AlphaFold2 and Boltz-1 were tested, and the method depends on the FrustrAI-Seq predictor.

Source: Cao H, Gao Z, Gu C, Liu G, Heng PA, Chatterjee P. SF-Cluster: Frustration-Guided MSA Subsampling for Alternative Protein Conformation Recovery. arXiv preprint arXiv:2607.00180 [q-bio.BM], 2026. https://arxiv.org/abs/2607.00180