AlphaFlow

AlphaFlow webserver overview

Give AlphaFold the 76 residues of human ubiquitin and it returns one structure. Give them to AlphaFlow and it returns as many as you ask for, each an independent draw from the same model. The differences between them are the result: regions that overlap in every sample are rigid in the model's view, and regions that wander are flexible. AlphaFlow turns a structure predictor into a sampler of conformational ensembles.

ProteinIQ runs AlphaFlow from Bowen Jing, Bonnie Berger and Tommi Jaakkola (ICML 2024) through its own predict.py, with all 14 released checkpoints. AlphaFlow is built on AlphaFold and uses a multiple sequence alignment, either one you upload or one generated for the job. ESMFlow is built on ESMFold and needs only the sequence. Both have MD + Template checkpoints that also read a reference structure.

Every job returns the native multi-model PDB with the whole ensemble, each sample as its own PDB file, and, for AlphaFlow models, the alignment used. For one best structure with a confidence score, AlphaFold 2 or ESMFold is the better tool; AlphaFlow answers which parts of a protein move and how far.

Pricing

Jobs start at 50 credits. The quote is calculated before submission from the sequence length, the model variant, the number of samples and the inference schedule.

Model variant (10 samples, 10 steps)	100 aa	300 aa	600 aa	900 aa	1,200 aa	1,600 aa
ESMFlow base (MD, PDB or MD + Template)	68	205	670	2,374	4,812	14,364
ESMFlow distilled	50	50	88	270	518	1,496
AlphaFlow base (MD, PDB or 48-layer MD + Template)	113	438	1,320	4,400	8,573	24,757
AlphaFlow distilled	50	50	121	391	767	2,246
AlphaFlow 12-layer MD + Template base	62	192	546	1,787	3,460	9,952

Cost grows with the square of sequence length and steps up again above 600, 900 and 1,250 residues, which is why the last three columns climb steeply. The number of samples scales the inference part of the quote almost linearly:

Samples, 300 aa, 10 steps	1	10	25	50
ESMFlow MD Base	53	205	459	883
AlphaFlow MD Base	70	438	1,051	2,074

These cells come from the current calculator. More Inference steps cost more and fewer cost less, and a Diversity (tmax) below 1.0 adds one step to the schedule. Turning on No diffusion prices a base model like a distilled one, because both run a single pass per sample. Uploading an MSA, adding a template, and the Self conditioning and Noisy first switches leave the quote unchanged.

Inputs

Input	Accepted	Limits
Protein sequence	One raw sequence or one FASTA record, as text or a .fasta, .fa or .txt file, or a PDB ID such as 1UBQ to fetch the sequence from the RCSB	10 to 1,600 residues, one record
MSA	Optional .a3m alignment, uploaded or pasted; shown in AlphaFlow and AlphaFlow + template modes	50 MB
Reference structure	One PDB file, uploaded or fetched from the RCSB; required in both + template modes	One protein chain, one residue per sequence position, 50 MB

Letters outside the 20 standard amino acids, such as B, Z or U, become X at the same position, so the sequence length never changes. Other characters, and FASTA input with more than one record, are rejected (how to make a FASTA file).

Without an uploaded alignment, AlphaFlow modes build one with AlphaFlow's own ColabFold MMseqs2 query script before sampling. An uploaded .a3m is used as it is, and it must pass these checks:

Check	Requirement
Records	Every record has a non-empty > header and at least one sequence line.
Characters	Letters and - only; lowercase letters are insertions.
Columns	With lowercase insertions removed, every record has the same length.
Query	The first record, with insertions and gaps removed, is exactly the submitted sequence, using X where the sequence has an ambiguous residue.

The reference structure is read position by position, so it must have exactly as many residues as the sequence and no gaps. AlphaFlow's PDB reader counts every residue that has a standard protein atom name, including HETATM records.

Crystallographic waters (HOH) and calcium ions (CA) therefore count as positions: RCSB entry 1UBQ has 76 ubiquitin residues and 58 waters, and AlphaFlow reads 134. Remove waters, ions and ligands with the structure editor before submitting. Modified residues written as HETATM, such as selenomethionine (MSE), are kept as positions and are accepted.

Settings

Mode and model

Parameter	Type	Default	Description
AlphaFlow input mode	enum	AlphaFlow	ESMFlow, ESMFlow + template, AlphaFlow or AlphaFlow + template; decides which inputs and checkpoints are offered.
Model variant	enum	AlphaFlow MD Base	Checkpoint for the selected mode; see the table below.
Number of samples	integer	10	Conformations to generate, from 1 to 50.

The checkpoints differ in training data and speed. PDB models were trained on experimental structures and sample the kind of diversity seen across crystal and cryo-EM structures of the same protein. MD models were further trained on the ATLAS molecular dynamics trajectories at 300 K and sample thermal motion around the fold.

MD + Template models also take a structure and sample the 300 K ensemble around it. Distilled models run in a single pass per sample, much faster, with some loss of accuracy reported in the paper.

Model variant	Mode	Notes
AlphaFlow MD Base, AlphaFlow MD Distilled	AlphaFlow	MD ensembles from sequence and MSA.
AlphaFlow PDB Base, AlphaFlow PDB Distilled	AlphaFlow	Experimental-state diversity; Self conditioning and per-sample MSA resampling are always on.
AlphaFlow MD + Template Base, AlphaFlow MD + Template Distilled	AlphaFlow + template	48 Evoformer layers.
AlphaFlow 12-layer MD + Template Base, AlphaFlow 12-layer MD + Template Distilled	AlphaFlow + template	12 Evoformer layers; about 2.5 times faster with a small loss in performance, according to the AlphaFlow authors.
ESMFlow MD Base, ESMFlow MD Distilled	ESMFlow	MD ensembles from sequence alone.
ESMFlow PDB Base, ESMFlow PDB Distilled	ESMFlow	Experimental-state diversity from sequence alone.
ESMFlow MD + Template Base, ESMFlow MD + Template Distilled	ESMFlow + template	Template-conditioned, no MSA.

Advanced settings

Parameter	Type	Default	Description
Inference steps	integer	10	Denoising steps, from 2 to 50.
Diversity (tmax)	number	1.0	Where the denoising schedule starts, from 0.2 to 1.0.
Self conditioning	boolean	false	Passes each step's prediction to the next step; always on for AlphaFlow PDB models.
No diffusion	boolean	false	Predicts each sample in a single pass instead of following the schedule; always on for distilled models.
Noisy first	boolean	false	Gives the first pass the random starting coordinates, which it otherwise does not see; always on for distilled models.

The defaults of 10 steps and tmax 1.0 are AlphaFlow's standard inference. The authors suggest tmax 0.2 with 2 steps for more precise samples with less diversity: with a tmax below 1.0, the model first predicts a structure, then samples from it with only that fraction of the noise added back.

The three switches follow the authors' recommendations, Self conditioning with MSA resampling for PDB models and No diffusion with Noisy first for distilled models, and those models use them whatever the switches show. On a base model, No diffusion without Noisy first gives the network no random input, so every sample comes out the same.

Very long sequences limit how many samples fit in one job. At 1,600 residues, AlphaFlow MD Base allows up to 21 samples at 10 steps and ESMFlow MD Base up to 38; at 1,200 residues AlphaFlow allows 37. Up to 900 residues every variant allows 50. Larger requests are refused before launch with a message to reduce the length, samples or steps.

Outputs

The result page shows the ensemble in a 3D viewer and lists every file for download.

File	Contents
<name>.pdb	The native AlphaFlow output: one PDB with every sample as a MODEL record, in generation order, starting at MODEL 0.
sample_01.pdb, sample_02.pdb, ...	Each model of the ensemble as its own PDB file, in the same order.
<name>.a3m	AlphaFlow modes only: the uploaded or generated alignment the job used.
<name>_runtime.json	Seconds spent on each sample, as AlphaFlow reports them.
<name>_alphaflow.log	AlphaFlow's own console output.
run.log	Model, sequence length, samples, steps, tmax, MSA source and whether a template was used.

<name> comes from the input's file name or label. Each result also records the checkpoint, the effective values of all four inference flags and the AlphaFlow release, so two runs can be compared exactly. In workflows, the ensemble file and the individual samples are separate outputs, so the samples can feed a tool that takes one structure at a time.

Understanding results

Samples are not ranked

sample_01.pdb is the first conformation generated, not the best one. AlphaFlow produces no score that orders its samples, and every sample is an equally weighted draw from the ensemble. Read the ensemble as a whole: what stays put across samples, what moves, and how far.

Reading the ensemble

Superpose the samples on the stable core before judging motion. Whole-protein RMSD between samples is dominated by flexible termini; aligning the helices and sheets that overlap first, then measuring a loop or a second domain, shows where the variation really is.

Pattern across samples	What it suggests	Next check
Secondary structure overlaps in nearly every sample	A consistent fold	Look at loops, termini and domain interfaces rather than the core.
A loop or terminus spreads widely	A mobile or weakly constrained region	Compare with B-factors or disorder in experimental structures, or with mutagenesis data.
Two domains keep their own folds but change orientation	Hinge-like motion	Align on one domain and measure the displacement of the other.
One sample differs from all the others	An outlier draw, not a stable state	Generate more samples and see whether the movement recurs.
Most samples disagree globally	The sequence may be largely disordered or outside the model's reliable range	Try single domains, or compare with an ESMFold or AlphaFold 2 prediction.

Ten samples show the dominant motions; the AlphaFlow authors evaluated their ensembles with 250 structures per protein. When a rare state matters, such as a pocket that opens in only some conformations, run 25 to 50 samples before drawing conclusions.

The B-factor column

Each atom's B-factor field holds the per-residue pLDDT, on a 0 to 100 scale, that the network's confidence head wrote for that sample. The AlphaFlow README does not describe or evaluate this value. Treat it as the model's confidence in that particular sample, not as a flexibility measure: a loop can move a lot across samples and still score well in each one.

MD and PDB ensembles answer different questions

An MD-model ensemble estimates what the protein does at 300 K in solution around its fold, so its spread is comparable with the RMSF of a short simulation. A PDB-model ensemble estimates the range of states seen across experimental structures, which can include larger changes such as ligand-induced domain closure. Running both on the same sequence and comparing the spread is a quick way to separate thermal breathing from state changes.

Checking a run

Ubiquitin is a good first run, and it is available as the public Human Ubiquitin example. Expect its compact core to overlap in every sample and most of the variation to sit in the C-terminal tail, which is flexible in solution.

A run in which the core of a well-folded protein unfolds in most samples points to a problem with the input, such as a wrong sequence, a long expression tag or, in AlphaFlow modes, a poor alignment; the .a3m returned with the result shows how many homologs the job found.

How AlphaFlow works

AlphaFlow fine-tunes AlphaFold and ESMFold as flow-matching generative models. During training the model sees structures interpolated between a noise sample $x_0$, drawn from a harmonic prior over the chain, and the target structure $x_1$:

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

The network learns to predict the clean structure from $x_t$. At inference it starts from noise and steps along the schedule set by Inference steps and Diversity (tmax), each step moving the coordinates toward its prediction. Different starting noise gives different samples.

AlphaFlow models compute AlphaFold's MSA features for every sample, which makes them slower than ESMFlow and adds the alignment step when no .a3m is uploaded. Template checkpoints read the reference coordinates as an extra input, and the 12-layer variants trim the Evoformer from 48 blocks to 12.

When to use AlphaFlow vs alternatives

Tool	Best use	Difference from AlphaFlow
ESMFold	One fast structure with pLDDT from sequence alone	Returns a single prediction, not an ensemble.
AlphaFold 2	The most accurate single structure, with pLDDT and PAE	Not designed to sample conformational spread.
AF-Cluster	Alternative states from clustered subsets of an MSA	Finds discrete states hidden in the alignment rather than sampling a continuous ensemble.
BioEmu	Equilibrium ensembles from sequence	A different generative model trained on simulation and experimental stability data.
MDGen	Learned molecular dynamics trajectories	Generates time-ordered trajectories instead of independent samples.
GROMACS	Physics-based molecular dynamics	Explicit physics with solvent and ligands, at far higher compute cost.

AlphaFlow sits between structure prediction and simulation. Its ensembles are a cheap way to find the regions worth simulating, and the individual samples can serve as receptor conformations for ensemble docking. When trajectories already exist, MD trajectory analysis measures the same flexibility from real simulation data.

Limitations

AlphaFlow models one protein chain. Ligands, cofactors, partner chains, nucleic acids and membranes are not part of the input, so motions they drive or suppress are missing from the ensemble. A template conditions the samples on one chain's coordinates only.

The ensembles are generated, not simulated. MD checkpoints reproduce the flexibility patterns of ATLAS simulations at 300 K, which are short, so slow transitions and rare states are underrepresented. Side-chain positions are predicted but are less reliable than the backbone for pocket-level conclusions.

AlphaFlow modes depend on their alignment, and a shallow one with few homologs gives less reliable samples. Generated alignments come from the public ColabFold server; if the search does not finish in time, the job stops with a message to retry later or upload an .a3m.

Sources

• Jing B, Berger B, Jaakkola T. AlphaFold Meets Flow Matching for Generating Protein Ensembles. ICML 2024. doi:10.48550/arXiv.2402.04845
• AlphaFlow source code and model weights: github.com/bjing2016/alphaflow
• Vander Meersche Y, Cretin G, Gheeraert A, Gelly JC, Galochkina T. ATLAS: protein flexibility description from atomistic molecular dynamics simulations. Nucleic Acids Research 2024. doi:10.1093/nar/gkad1084