How to Use AFcluster

Overview. AFcluster is a multiple-conformation prediction workflow built around AlphaFold2 and multiple-sequence-alignment clustering. The central idea from the Nature paper is that a full MSA can mix co-evolutionary signals from more than one structural state; clustering homologous sequences by similarity before running AlphaFold2 can separate those signals and expose alternative conformations.

Use Cases. AFcluster is therefore useful for fold-switching proteins, state-selective mutants, and other targets where one sequence may occupy more than one structural basin. The paper demonstrated this on metamorphic proteins including KaiB, RfaH, and MAD2, and also used the method to identify a candidate alternative state for Mpt53.

Workflow on Neurosnap. Researchers can submit an Input Sequence and let Neurosnap generate the MSA, or upload a curated a3m alignment directly. The results are organized into a structural cluster map, an animation of the sampled conformations, and a table with per-model confidence metrics such as pLDDT and pTM.

How It Works. AF-Cluster first clusters MSA sequences by sequence similarity, then runs AlphaFold2 on cluster-derived subsets instead of the full alignment alone. This is not a generic diversification trick. It is an attempt to deconvolve conflicting evolutionary couplings that may correspond to distinct conformational substates.

Scientific Context. In the original study, this strategy recovered known alternative states with high confidence and supported experimental validation for KaiB state preferences and mutation-driven state switching. That makes AFcluster more scientifically grounded than treating alternate AlphaFold runs as random samples from conformational space.

Workflow. On Neurosnap, Animation Projection Algorithm affects how sampled structures are ordered for visualization. The cluster projection, animation, and confidence table are best interpreted together: coherent clusters with strong pLDDT and pTM are the most plausible candidates for real alternative states worth experimental or mechanistic follow-up.

• Alignment-aware conformational sampling: AFcluster uses the MSA itself to separate state-specific evolutionary signals rather than relying on one default AF2 prediction.
• Flexible starting point: Researchers can begin from a raw sequence or from a user-supplied a3m alignment when they want tighter control over homolog coverage.
• State-family interpretation: Cluster plots and animations make it easier to see whether predictions form coherent conformational groups or isolated outliers.
• Confidence-guided triage: Per-model pLDDT and pTM help distinguish plausible alternative states from low-confidence structural drift.