How to Use AFcluster

AFcluster is a conformational-state discovery workflow built around AlphaFold2 and multiple-sequence-alignment subsampling. In the Nature paper, the core idea is that a full MSA can mix evolutionary signals for different structural states; clustering the MSA by sequence similarity can separate those signals and let AlphaFold2 sample distinct conformations with high confidence.

On Neurosnap, you can start from an input sequence and let the platform generate an MSA, or provide your own a3m alignment directly for clustering and conformational sampling. The results view then organizes sampled models into a 2D cluster map, an animation, and a sortable table with pLDDT and pTM so alternate state families are easy to compare.

The paper's AF-Cluster pipeline first generates or accepts a target MSA, clusters alignment sequences by sequence similarity using DBSCAN, and then runs AlphaFold2 separately on those cluster-derived sequence subsets rather than on the full alignment alone. The purpose is not to improve a single consensus structure, but to deconvolve conflicting co-evolutionary signals so alternate biologically relevant states can emerge.

This strategy was shown to recover multiple conformations for known fold-switching proteins such as KaiB, RfaH, and MAD2, and it also improved some engineered mutation cases where default AF2 sampling missed the correct state. In other words, AFcluster is most useful when the protein family may encode more than one structural basin and you want to inspect a conformational landscape instead of one averaged prediction.

On Neurosnap, Animation Projection Algorithm controls how the sampled structures are ordered for visualization. The results page combines structural clustering, per-model confidence summaries, and animated traversal of nearby conformations so researchers can quickly identify coherent high-confidence state families.

• Study-fit inputs: AFcluster accepts either an input sequence for MSA generation or a user-supplied a3m alignment, matching both exploratory and curated workflows.
• State-aware sampling: Sequence clustering can separate mixed evolutionary signals in the MSA, helping AlphaFold2 reveal distinct conformational substates rather than collapsing them into one dominant prediction.
• Protocol control: Researchers can tune Animation Projection Algorithm to choose how sampled structures are projected and ordered for inspection.
• Readable evidence: Structural clustering plots, conformational animations, and per-model pLDDT and pTM summaries make it easier to compare candidate state families.
• Faster iteration: Managed execution on Neurosnap removes infrastructure overhead so teams can focus on interpreting conformational landscapes rather than setting up AF2 subsampling and clustering pipelines.