Use SF-Cluster Online

Commercially Available SF-Cluster No-Code Web Server

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

Predict alternative protein conformations, sample distinct structural basins, and unlock hidden functional states beyond standard sequence similarity.

Overview

Deep-learning structure predictors are highly sensitive to their multiple sequence alignment (MSA) inputs, making MSA subsampling a powerful method for discovering alternative protein conformations. However, traditional sequence-space clustering methods (like AF-Cluster) struggle when minor sequence variations heavily impact fold energetics without significantly altering global sequence similarity.

SF-Cluster solves this by shifting from sequence space to a structure-relevant pattern space. It subsamples MSAs based on patterns of predicted local energetic frustration—a representation that directly captures structural preferences independent of sequence distance. Tested across a robust benchmark of 48 challenging cases, SF-Cluster consistently outperforms sequence-based methods in capturing alternative protein states.

Key Features & Scientific Validation

  • State-Directed Conformational Sampling: SF-Cluster reliably recovers alternative conformations across fold-switching, allosteric, oligomerization-coupled, and intrinsically disordered systems.

  • Significant Performance Gains: Achieves a massive +15.5 percentage point improvement in target-state recovery for allosteric systems compared to AF-Cluster.

  • Architecturally Transferable Signals: The conformational signals captured by SF-Cluster live in the MSA composition itself. Alignments selected by SF-Cluster successfully transfer from AlphaFold2 to architecturally distinct predictors like Boltz-1.

  • Biophysically Anchored Signal: Highly frustrated residues selected by the framework match functional and structural switching regions confirmed by independent experimental data, including Deep Mutational Scanning (DMS) and NMR two-state exchange.

  • Depth-Controlled Mosaic Regime: Unlike sequence clustering that can fragment large alignments into shallow subsets, SF-Cluster reliably maintains deep, diverse subsets ($Neff_{80} \approx 30$) crucial for successful structural predictions.

How It Works

SF-Cluster reformulates MSA subsampling into a representation-guided reweighting workflow:

  1. Frustration Mapping: Each homolog sequence within a filtered raw MSA is processed using FrustrAl-Seq to predict its per-residue local energetic frustration profile.

  2. Region-Level Compression: The raw profiles are compressed into fixed-dimensional region-level embeddings capturing spatial frustration averages, variances, and directional contrasts.

  3. Mosaic Selection: Homologs are mapped based on their geometry in frustration-pattern space. A single mosaic selection draws fixed-size subsets designed to maximize within-subset frustration variance, exposing alternative structural basins to the deep learning model.

  4. Screen-and-Refine Protocol: Subsampled alignments are evaluated in a two-stage protocol to identify high-confidence structural states.

What is Tamarind Bio?

Tamarind Bio is a platform dedicated to making advanced structural biology, deep learning prediction pipelines, and bioinformatics tools accessible to all researchers. By hosting state-of-the-art open-source algorithms—such as SF-Cluster and interactive notebooks—the platform removes computational friction, enabling scientists to seamlessly run complex sequence-to-structure tasks, manage model parameters, and visualize alternative molecular configurations directly within a unified ecosystem.

How to Use SF-Cluster on Tamarind Bio

Running SF-Cluster on Tamarind Bio allows you to quickly explore alternative protein landscapes directly from your browser:

  • Input Requirements: Provide the query protein sequence or a pre-generated multiple sequence alignment (MSA) in A3M format.

  • Automated Filtering: The platform automatically filters gap characters over aligned columns to ensure clean input data.

  • Subsampling Configuration: The algorithm seamlessly executes frustration-pattern modeling and mosaic sampling to output 12 optimized subsets of 32 sequences per arm.

  • Prediction Output: The subsets are passed through a controlled screen-and-refine pipeline with your chosen structure predictor backend to generate an ensemble of predicted structures capturing the dominant and target conformational basins.

Source

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma