How to Use AlphaFlow

Overview. AlphaFlow and ESMFlow are sequence-conditioned generative models for protein conformational ensembles. Instead of returning one best structure, the method fine-tunes AlphaFold- and ESMFold-style predictors under a custom flow-matching framework so they can sample multiple structures from a learned conformational landscape.

Use Cases. This is useful when protein function depends on flexibility, alternative states, or heterogeneous ensembles rather than a single static fold. In the paper, PDB-trained models improved the precision-diversity trade-off relative to MSA subsampling, while MD-trained variants better captured conformational flexibility, positional distributions, and higher-order ensemble observables.

Workflow on Neurosnap. Researchers submit an Input Sequence, choose among Model Weights spanning AlphaFlow or ESMFlow with PDB- or molecular-dynamics-trained weights, and set Number of Conformations to sample. The results page summarizes each conformer with mean pLDDT, uniqueness, RMSD to the best-ranked model, and an interactive multi-model structure viewer.

How It Works. The paper repurposes AlphaFold and ESMFold as denoisers inside a flow-matching generative process. Sampling starts from a polymer-like prior and progressively refines noisy coordinates into plausible structures, which turns a single-state predictor into a model that can generate ensembles from sequence.

Methodological Context. This is not just inference-time perturbation. AlphaFlow explicitly trains sequence-to-structure generative models, and the MD-specialized weights extend that framework beyond experimentally deposited structures to all-atom simulation ensembles. That broader training signal is why the method is relevant to researchers who care about realistic flexibility, not only static fold prediction.

Workflow. On Neurosnap, Model Weights determine which training regime is closest to the scientific question, while Number of Conformations controls how broadly the sequence is sampled. Mean pLDDT is best used as a per-conformer confidence filter, and the uniqueness and RMSD-to-best metrics help separate small fluctuations from structurally distinct states worth downstream simulation or experiment.

• Sequence-to-ensemble modeling: AlphaFlow starts from sequence but returns multiple candidate states instead of one averaged structure.
• Choice of training regime: PDB- and MD-trained weights let researchers bias runs toward experimentally observed heterogeneity or simulation-derived flexibility.
• Quantitative ensemble review: Mean pLDDT, uniqueness, and RMSD-to-best support faster discrimination between consensus-like and genuinely distinct conformers.
• Useful MD precursor: The generated ensemble can help prioritize which states deserve more expensive downstream simulation or biophysical follow-up.