What is AlphaFlow?

Protein structures in the PDB and in molecular dynamics simulations rarely describe a single rigid shape. Loops open and close, termini wander, domains breathe, and binding sites can appear only in a subset of conformations. AlphaFlow models that variability by turning AlphaFold-style structure prediction into a generative ensemble method.

AlphaFlow, developed by Jing, Berger, and Jaakkola, includes two model families. AlphaFlow uses AlphaFold with multiple sequence alignments. ESMFlow uses ESMFold and runs from sequence alone. ProteinIQ supports both model families: ESMFlow for single-sequence ensemble generation, and AlphaFlow for MSA-backed AlphaFold-style ensemble generation.

For a single best-guess structure, ESMFold or AlphaFold 2 is usually the clearer choice. AlphaFlow is useful when the question is about plausible conformational spread rather than one static model.

How to use AlphaFlow online

AlphaFlow runs online in ProteinIQ by selecting an input mode, providing one protein sequence, choosing a matching model variant, and generating a conformational ensemble as downloadable PDB files. ESMFlow mode needs only a sequence. AlphaFlow mode accepts the same sequence plus an optional .a3m multiple sequence alignment, and generates the alignment automatically when none is uploaded.

Inputs

AlphaFlow on ProteinIQ is designed for single-chain protein sequences. Multimer interfaces, ligand-bound states, membranes, cofactors, and chain-specific contacts are not modeled directly from the input. When a PDB ID is fetched from RCSB, the sequence is used as the input and the deposited coordinates are not used as a structural template.

A3M validation

Uploaded .a3m alignments are checked before the job starts. ProteinIQ accepts permissive A3M content, but rejects files that would make the AlphaFold-style feature pipeline ambiguous.

If no .a3m is uploaded in AlphaFlow mode, ProteinIQ generates an MSA server-side with an MMseqs2 workflow compatible with ColabFold-style A3M output. The generated alignment is returned with the results so the run can be inspected or repeated.

Before running

• Trim expression tags and purification handles: Flexible His-tags or long artificial linkers can dominate the ensemble and make biologically relevant motion harder to see.
• Keep domains interpretable: Multi-domain proteins can be useful, but long flexible linkers may cause large global rearrangements. Running individual domains can make local flexibility easier to inspect.
• Check the sequence alphabet: Ambiguous residues such as X, B, Z, or U are mapped to X without changing sequence positions. Resolve them before inference when possible.
• Start with 10 samples: The default gives enough structures to spot repeated motion without making the first run slow. Increase to 25 or 50 when comparing domain movement, loop states, or binding-site exposure.

Human ubiquitin (1UBQ, 76 residues) is a useful sanity-check target because the beta-grasp fold should remain compact while termini and surface loops vary modestly across samples. A run where the core unfolds or most samples look unrelated is a warning sign to inspect the input sequence and settings.

Model settings

MD models learn conformational variation from molecular dynamics trajectories. PDB models learn diversity across experimental structures. Base models are the full checkpoints. Distilled models are faster and automatically use the distilled inference settings.

Choosing a model variant

Advanced settings

The two controls that most users should change first are Number of samples and Model variant. Inference steps, Diversity (tmax), and the boolean inference flags are better treated as reproducibility controls. Changing them changes the sampling procedure, not just the display.

Results

Both ESMFlow and AlphaFlow return the native AlphaFlow multi-model PDB as the primary result. ProteinIQ also splits ensembles within the derived-output size budget into secondary per-sample PDB files for convenient viewing; very large ensembles remain available in their complete native multi-model file without duplicating the payload. AlphaFlow mode returns the uploaded or generated .a3m, and both modes return source runtime data and the source-entrypoint log when available.

AlphaFlow does not return a pLDDT confidence table or a binding score. The value of the run is in the ensemble itself: how similar the structures are, which regions move, and whether the sampled conformations suggest more than one plausible state.

Results record the reviewed AlphaFlow source commit 796971961368c9bd1504b3d57c5348a3d1d4ecaa, the predict.py entrypoint, the immutable model-repository revision, and the selected checkpoint checksum.

Interpreting AlphaFlow ensembles

An AlphaFlow output should be read as a set of plausible conformations, not as a ranked list where sample 1 is best. The sample index only records generation order.

Large differences across samples usually point to regions where the model sees conformational uncertainty or mobility. In practice, the most useful checks are visual:

• Stable core: Helices and sheets that overlap across samples are more likely to represent a consistent fold.
• Flexible loops and termini: Disordered ends and solvent-exposed loops often spread widely. That spread is expected and should not be treated as a modeling failure by itself.
• Domain motion: If two domains remain internally stable but shift relative to each other, the ensemble may suggest hinge-like motion.
• Binding-site exposure: A pocket that appears in only some samples can be useful for hypothesis generation, especially before docking or mutagenesis planning.
• Outlier samples: A single highly distorted sample should not outweigh the rest of the ensemble. Inspect whether the same movement appears repeatedly.

Quantitative RMSD analysis is useful after the job, but RMSD needs a reference choice. Whole-protein RMSD can be dominated by flexible tails. For many proteins, aligning the stable core first and then measuring loop or domain movement gives a clearer interpretation.

How AlphaFlow works

AlphaFlow trains structure predictors as flow-matching models. During training, the model sees an interpolation between a noisy protein backbone representation and a target structure:

$x_t = (1 - t) \cdot x_0 + t \cdot x_1$

Here, $x_0$ is sampled from a harmonic prior, $x_1$ is the target structure, and $t$ marks progress along the denoising path. At inference time, the model starts from a noisy configuration and repeatedly predicts a cleaner structure along the schedule.

ESMFlow applies the same idea to ESMFold. Because ESMFold is a single-sequence model, ESMFlow can run without MSA generation. That makes it practical for fast web execution and for proteins where homologous sequence coverage is weak.

AlphaFlow mode uses AlphaFold-style MSA features. A supplied .a3m file is used directly after validation. Without an upload, ProteinIQ generates the alignment first, then writes it into the directory layout expected by the AlphaFlow inference code before sampling structures. That extra MSA step is the main reason AlphaFlow mode is slower than ESMFlow mode.

Training data behind the model variants

The distinction matters. An MD-trained model may emphasize physically local fluctuations. A PDB-trained model may sample larger state changes if similar state diversity appears in experimental structures.

When to use AlphaFlow vs alternatives

AlphaFlow fits best between structure prediction and simulation. It can suggest flexible regions worth checking before setting up molecular dynamics, docking, mutagenesis, or construct design. It should not replace experimental dynamics data when NMR, HDX-MS, cryo-EM heterogeneity, or high-quality MD trajectories are available.

Practical limits

• MSA generation can dominate runtime: AlphaFlow modes require MSA features. Uploaded .a3m files make runs more reproducible; automatic MMseqs2 generation depends on external search availability.
• No ligand or multimer conditioning: The input is a protein sequence. Ligands, cofactors, membranes, partner chains, and DNA or RNA binding partners are not included in generation.
• Backbone-centered ensemble interpretation: The method is most useful for backbone conformational diversity. Side-chain packing and small binding-site rearrangements need follow-up validation.
• Long sequences cost more: Memory and runtime increase sharply with sequence length. ProteinIQ accepts up to 1600 residues, but shorter constructs are usually easier to interpret.
• Combined workload is bounded: Sequence length, sample count, and inference steps are checked together before launch. Combinations estimated to exceed one six-hour GPU job are rejected with guidance to reduce the workload.
• Samples are hypotheses: Repeated motion across several samples is more informative than a single unusual conformation. Experimental or simulation evidence is still needed before claiming a functional state.