AlphaFold2

AlphaFold2 webserver overview

AlphaFold2 is a deep-learning model for predicting protein structures from amino acid sequences. ProteinIQ runs the ColabFold implementation for single-chain monomer and multi-chain AlphaFold2-Multimer prediction.

The webserver accepts protein sequences, query tables, precomputed A3M alignments, protein structures, and RCSB sequence fetches. Optional files supply templates and an initial guess. It returns ranked PDB structures with pLDDT, pTM, ipTM, PAE, and supporting files. Ligands, DNA, and RNA are outside its input scope.

Pricing

AlphaFold2 jobs start at 50 credits. Single-sequence inference runs without an MSA search, while MMseqs2 MSA modes cost 2x the single-sequence calculation.

The final price also depends on sequence length, chain count, models, recycles, random seeds, ensemble runs, and AMBER relaxation, and is calculated before submission.

Inputs

ProteinIQ displays chain IDs so that multi-chain inputs can be tracked in the viewer and result files. Each job requires at least one protein input and accepts up to 10 protein inputs, with a combined limit of 6,000 residues.

Sequence text must use ungapped one-letter protein codes. A precomputed A3M file must contain a valid protein query sequence and alignment.

Input	Accepted formats	Limits and behavior
Protein sequence	Raw sequence or FASTA text, .fasta, .fa, or .txt file	Up to 10 protein inputs. Sequence text must use ungapped amino acid codes accepted by AlphaFold2.
Precomputed protein MSA	.a3m file	Files up to 10 MB. The A3M query sequence must contain a usable ungapped protein sequence.
Query table	.csv or .tsv	Columns id and sequence; each row is a query. Use colons between complex chains.
Query structure	.pdb, .cif, .mmcif	Upload it as the protein input. ColabFold extracts the protein sequences and predicts them as one query.
Template / Initial guess	.pdb, .cif, .mmcif	Optional structure files. Enable Use templates for custom templates. An initial guess requires one query.
MSA	.a3m	Optional files referenced by exact filename in a table's a3mpath column.
Template hits	.tsv or .txt	Native tab-separated hit records, accompanied by the matching uploaded mmCIF files, named by four-character PDB ID such as 1abc.cif.
RCSB protein fetch	PDB entry fetched as FASTA	One RCSB fetch per input. Upload a structure file instead to use its exact chains.
Job name	Text	Optional label used to identify the run.

AlphaFold2 accepts protein and protein-complex inputs only. A ligand, DNA sequence, or RNA sequence cannot be submitted as a prediction input; use a multi-molecule model such as Boltz-2 when those components are part of the system.

Settings

The settings below follow the current AlphaFold2 webserver definition. Some fields depend on a parent setting, and Not set means the optional value is left to the native or tier-specific default.

Prediction and model

Setting	Default or range	Description
Multiple input files	Chains of one complex	Separate queries runs each file's queries independently. A multi-record FASTA always retains separate queries.
Number of recycles	Native model default; explicit 0-12	Blank preserves the native default: 20 for Multimer v3, 3 for monomer and Multimer v1/v2. 0 disables recycling.
Early stop tolerance	Native model default	Blank or none preserves the model tolerance, including 0.5 for Multimer v3. -1 disables early stopping. Other finite numeric tolerances pass to ColabFold.
Number of models	5; 0-5	0 prepares alignments and templates without predicting structures. Positive values select the number of trained models.
Random seeds	1; 1-5	Number of different random seeds used to generate additional structural diversity.
Ensemble runs	1; 1-3	Number of runs with different MSA cluster samplings. More runs increase diversity and runtime.
Model type	Auto	Auto selects the pTM model for single chains and a multimer model for complexes. Specific options include AlphaFold2 monomer, AlphaFold2-ptm monomer, AlphaFold2-multimer v1, v2, or v3, and DeepFold v1.
Enable dropout	Off	Enables dropout during prediction to generate more diverse structures that can indicate conformational uncertainty.

MSA and templates

Setting	Default or range	Description
MSA mode	Paid default: UniRef+Environmental; free default: No MSA	Selects UniRef+Environmental, UniRef+Environmental env-pair, UniRef only, or No MSA. The three database-search modes cost 2x credits.
Max MSA sequences	Model default	Blank or none preserves the model configuration. A positive integer sets the MSA cluster count.
Max extra MSA sequences	Model default	A positive integer limits extra MSA sequences. Native defaults include 5120 for monomer, 2048 for Multimer v3 and 1152 for Multimer v1/v2.
Max MSA pair	Not set	Advanced override in the form max_seq:max_extra_seq, such as 512:5120. When set, it overrides Max MSA sequences and Max extra MSA sequences.
Use templates	Off	Queries PDB templates through the MSA server for homologous structures.
Max templates	20; options 5, 10, 20, or 50	Maximum number of templates considered when Use templates is enabled.
Max template date	Not set	Optional cutoff in YYYY-MM-DD format. Only templates released before the selected date are used.

Multimer, ranking, and diagnostics

Setting	Default or range	Description
Pair mode (multimer)	Unpaired+Paired	Chooses Unpaired+Paired, Paired only, or Unpaired only MSA pairing for multi-chain predictions.
Pairing strategy (multimer)	Greedy	Chooses Greedy or Complete species matching for multimer sequence pairing.
Ranking metric	Auto	Selects the metric used to order predictions: Auto, pLDDT, pTM, ipTM, or Multimer score. Auto preserves native monomer or multimer ranking behavior.
Model order	1,2,3,4,5	Controls the ColabFold model execution order for reproducible runs or benchmarking.
Stop at score	Not set	Stops after ColabFold reaches a supplied pLDDT or pTM threshold. When not set, all requested models run.
Extra pTM metrics	Off	Calculates additional multimer interface metrics, including actifpTM and chain-wise pTM.
Binary contacts for extra metrics	Off	Uses binary contacts instead of contact probabilities when Extra pTM metrics is enabled.
Save recycle intermediates	Off	Saves structure predictions from each recycle step for inspecting how the model changes during refinement.
Use MSA cluster profiles	On	Passes the native cluster-profile option to ColabFold.
Save raw prediction data	Off	Retains native prediction pickle files for download.
Save single representations / Save pair representations	Off	Retains native NumPy representation arrays. Pair arrays can be large.

Relaxation and reproducibility

Setting	Default or range	Description
AMBER relaxation	Off	Runs OpenMM/AMBER energy minimization on the selected predictions. It can improve side-chain geometry but adds substantial runtime.
Structures to relax	0; 0-5	With AMBER enabled, 0 relaxes all requested models and seeds, matching ColabFold. A positive value selects that many top-ranked predictions.
Random seed	Not set; native default 0	Sets a fixed seed for reproducibility. When not set, ColabFold uses its default seed behavior.
Relax max iterations	2000	Maximum OpenMM/AMBER relaxation iterations when AMBER relaxation is enabled.
Relax tolerance	2.39	OpenMM/AMBER relaxation convergence tolerance.
Relax stiffness	10.0	OpenMM/AMBER relaxation stiffness parameter.
Relax outer iterations	3	Maximum outer iterations for OpenMM/AMBER relaxation.

With the standard settings, AlphaFold2 uses the model's native recycle configuration, 5 models, 1 random seed, 1 ensemble run, Auto model selection, no templates, no dropout, and no AMBER relaxation. Earlier saved jobs retain their original settings.

The default MSA mode depends on account tier: paid runs default to UniRef+Environmental, while free runs default to No MSA and can switch to any MSA mode. Optional fields marked Not set retain their native behavior.

Choosing settings

A No MSA run is a cheap way to confirm chain order, chain count and sequence lengths before paying for an alignment. It is not a fair test of whether a natural protein can be predicted.

AlphaFold2 draws most of its accuracy from coevolution in the alignment: in the original benchmark, accuracy fell sharply when the effective alignment depth dropped below about 30 sequences, and gains beyond about 100 sequences were small (Jumper et al., 2021). Once the input looks right, repeat the job with UniRef+Environmental and compare the two results.

Leave Number of recycles and Early stop tolerance blank on a first run. For a difficult monomer, raising recycles from the native 3 toward the hosted maximum of 12 gives the model more refinement passes. Multimer v3 already defaults to 20 recycles with early stopping at 0.5, so an explicit value of 12 or less shortens a Multimer v3 run rather than extending it. Lowering Max MSA sequences speeds up long proteins at some cost in accuracy on hard targets.

Number of models set to 1 is enough for a quick look. Keep 5 when the ranking matters, because agreement between independently trained models is part of the evidence for a fold. Increase Random seeds, Ensemble runs or turn on Enable dropout when the aim is to sample alternative conformations; none of them is needed for a standard prediction.

Use templates helps when a close homolog has been solved or a specific domain arrangement is expected. A poor template can pull the model toward the wrong fold, so run the same job without templates as a control whenever the template's relevance is uncertain. Max template date excludes structures released after a cutoff, which matters when benchmarking against a structure the model should not have seen.

Keep Model type on Auto (recommended) unless reproducing a command-line run or comparing weights. AMBER relaxation adds substantial runtime and is worth turning on only when side-chain geometry or clashes matter for docking or simulation.

Outputs

Each run includes Viewer, Data, and Files views. The primary PDB structures are ranked by the selected metric, while supporting files preserve confidence data, MSA and template artifacts, relaxation results, and the curated run record.

Output	Description
Ranked PDB structures	Native filenames retain the query, model, seed and rank. Ranks apply within each query. Earlier saved jobs may use prediction_rank01.pdb.
3D structure viewer	Interactive inspection of monomer or multimer structures with chain IDs and confidence coloring.
pLDDT	Per-residue confidence on a 0-100 scale, stored in the PDB B-factor column. Higher values indicate greater local confidence.
pTM	Global fold confidence on a 0-1 scale where the selected model provides pTM.
ipTM	Interface confidence for multimer predictions where the selected model provides ipTM.
PAE	Predicted aligned error data for assessing relative positioning between residues, domains, or chains.
MSA and template artifacts	ColabFold-generated A3M, JSON, image, and text artifacts when the selected MSA or template workflow produces them.
Relaxed structures	AMBER-relaxed structures and their unrelaxed counterparts when relaxation is enabled.
Recycle intermediates	Per-recycle structures when Save recycle intermediates is enabled.
Run log	A downloadable run.log containing chain and residue counts, selected settings, scientific phases, and ranked confidence summaries.

Native A3M files retain their #lengths<TAB>cardinalities header. Residue counts include query stoichiometry, excluding homolog rows. Tables may reference uploaded templates by filename through templatepath; server filesystem paths are not accepted. ColabFold 1.6.2 cannot read monomer a3mpath table entries, so submit those A3M files as protein queries instead.

If a later prediction or relaxation fails, completed structure/score pairs and diagnostics remain downloadable. Results without a native rank are identified by query, model and seed; their displayed order does not imply confidence. MSA-only runs have supporting files and no predicted structures.

DeepFold prediction remains unavailable pending verification of permission to use its model weights.

Example jobs

Two public example jobs can be opened from the tool page to inspect a finished result before submitting a sequence. Both used Random seed 42.

Example	Settings	Result
Human ubiquitin	One 76-residue chain, AlphaFold2-ptm (monomer), UniRef+Environmental with Max MSA sequences 128, 1 model, 3 recycles, ranked by pTM	Mean pLDDT 94.46, pTM 0.81
GCN4 leucine zipper homodimer	Two 33-residue chains, AlphaFold2-multimer v3, No MSA, 1 model, 3 recycles, Extra pTM metrics on, ranked by Multimer score	Mean pLDDT 92.55, pTM 0.77, ipTM 0.74

Ubiquitin is a small, well-conserved fold, and its mean pLDDT of 94.46 with a pTM of 0.81 is what a confident monomer looks like. The GCN4 dimer shows the opposite trap. Its ipTM of 0.74 looks marginal, yet the model matches the 2ZTA crystal structure with a DockQ of 0.935, as the structure prediction confidence worked example shows.

Interpreting AlphaFold2 results

pLDDT describes local residue confidence. pTM describes confidence in the global fold, while ipTM describes confidence in the relative placement of chains.

Start with the structure rather than the numbers. Open the top-ranked model in the Viewer, check that the expected domains are folded and that chains in a complex form a plausible interface, then compare it with the next two ranks. When the models converge on the same fold, the result does not hinge on one set of weights. When they disagree on domain placement or chain arrangement, the top rank is one of several candidates, and its score alone does not settle which is right.

Values above 90 are very high pLDDT confidence, and values from 70 to 90 generally support a reliable backbone. Values below 50 often identify disorder or flexible regions rather than a failed prediction. PAE complements these scores by showing whether domains or chains are confidently positioned relative to one another.

The Data view lists Mean pLDDT, pTM and ipTM for each model and seed. Mean pLDDT averages every residue, so a well-folded domain attached to a long disordered tail can average in the 60s. Read the per-residue coloring in the Viewer before concluding that the fold failed. A low-scoring region is often a finding about the protein, and trimming it from a monomer sequence only hides that.

In the PAE plot, a low-error block along the diagonal marks a rigid domain. High error between two such blocks means the model is unsure how the domains face each other, even when each has high pLDDT; flexible linkers and hinges look like this. In a complex, the off-diagonal blocks between chains carry the interface confidence.

When an MSA mode is used, the coverage plot among the PNG files shows how many aligned sequences cover each position. A low-pLDDT stretch with thin coverage may improve with a deeper alignment or a different MSA mode. A low-pLDDT stretch with deep coverage is more likely disordered or dependent on a binding partner.

For multimers, inspect both the chain interface and the scores used for ranking. With Ranking metric on Auto, multimer predictions are ordered by the AlphaFold-Multimer confidence score, 0.8 × ipTM + 0.2 × pTM (Evans et al., 2021). A high pLDDT can coexist with uncertain relative chain placement, so ipTM and inter-chain PAE are more informative for complex assembly than local confidence alone.

ipTM also drops for reasons unrelated to the interface: disordered tails, extra domains and short chains all lower it, which is why the GCN4 example scores 0.74. When that is the likely cause, read the inter-chain PAE between the contacting residues, turn on Extra pTM metrics for actifpTM, or pass the scores JSON from the Files view to IPSAE as the PAE file. Rerunning with only the interacting domains is another check.

Comparing several ranked models is useful when the target is flexible, lacks homologs, or has multiple plausible oligomeric states. This helps reveal whether the ranking is stable across plausible assemblies.

Common mistakes

The B-factor column of an AlphaFold2 PDB holds pLDDT, not experimental mobility. High values mean high confidence, the reverse of a crystallographic B-factor, so any viewer coloring or filter that reads B-factors is reading confidence.

Judging a natural protein from a No MSA run usually underestimates AlphaFold2, because a weak result there reflects the missing alignment more than the protein. Templates can mislead in the other direction: a confident model that closely matches its template may be reproducing it, and the untemplated control is the test.

Rank 1 is the model with the highest ranking metric, not a validated structure. Match the score to the question: pLDDT for local structure, PAE for domain arrangement, and ipTM with inter-chain PAE for interfaces. AMBER relaxation removes clashes but does not move a wrong fold toward the right one. The unrelaxed GCN4 model, for example, has clean backbone geometry and still a MolProbity clashscore of 23.3.

Before using a model for docking, simulation or design, check its geometry with MolProbity and compare it with any related experimental structure by RMSD and TM-score. FoldSeek finds structural neighbors in the PDB and AlphaFold DB, and DockQ scores a predicted interface when a reference complex exists.

Important limitations

AlphaFold2 is a protein-only predictor and does not model ligand binding, DNA, RNA, cofactors, or small-molecule effects on structure. Its valid prediction inputs are protein sequences and protein complexes.

MSA-assisted modes generally provide more evolutionary information for natural proteins, while No MSA is useful for fast checks and designed or orphan proteins with few detectable homologs.

Templates can bias a prediction toward a known fold, and AMBER relaxation changes coordinates without validating the biological state. Neither option confirms that the predicted state is biologically populated.

Each model is one static conformation. AlphaFold2 does not return ensembles, dynamics or induced-fit changes, and predicted side-chain rotamers are less reliable than the backbone, especially in low-confidence regions.

For protein complexes containing non-protein components, use Boltz-2 or Chai-1. For fast single-sequence protein folding, use ESMFold. For a protein that already has an entry in the AlphaFold Protein Structure Database, the AlphaFold Database Download tool fetches the precomputed model by UniProt accession without running a prediction.

Sources

• Jumper J, et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583-589 (2021). doi:10.1038/s41586-021-03819-2
• Evans R, et al. "Protein complex prediction with AlphaFold-Multimer." bioRxiv (2021). doi:10.1101/2021.10.04.463034
• Mirdita M, et al. "ColabFold: making protein folding accessible to all." Nature Methods 19, 679-682 (2022). doi:10.1038/s41592-022-01488-1