Predicting Protein Structure with the AlphaFold Server

What you’ll accomplish

By the end of this tutorial, you’ll know how to:

Submit a prediction job on the AlphaFold Server
Interpret the two key confidence metrics (pLDDT and PAE)
Identify regions of the prediction you can trust vs. regions to treat with caution
Understand what AlphaFold 3 can predict that AlphaFold 2 couldn’t

This tutorial uses the web server — no Python, no GPU, no local installation required.

Before you start

What you need:

A Google account (required to access the AlphaFold Server)
Your protein sequence in FASTA format (standard single-letter amino acid code)
An idea of what you’re trying to learn from the structure

Which AlphaFold version should you use?

The publicly available server currently runs AlphaFold 3, which can predict:

Single protein chains
Multi-chain protein complexes (protein–protein interactions)
Protein–DNA/RNA complexes
Protein–small molecule (ligand) complexes

If you need to predict a complex or include a ligand, AlphaFold 3 is the right choice. For very large numbers of single-chain predictions, consider ESMFold (via Hugging Face) which is faster but less accurate.

Step 1: Prepare your sequence

Your protein sequence should be in single-letter amino acid code. Example (first 20 residues):

MKTAYIAKQRQISFVKSHFS

For a full protein, paste the entire sequence. For a specific domain, paste just the domain sequence — AlphaFold will predict the structure of whatever you give it.

Common issues:

Remove any signal peptides or propeptides if you want the mature protein structure
If your sequence contains non-standard amino acids, check the server’s documentation — some are supported, others aren’t
For membrane proteins, be aware that AlphaFold predicts solution-phase conformations by default; membrane-embedded structures may require specialist interpretation

Step 2: Submit to the AlphaFold Server

Go to the AlphaFold Server (alphafoldserver.com) and sign in with your Google account
Click New prediction
Paste your sequence in the sequence input box. For a single chain, this is all you need.
If you want to predict a complex: add additional chains using the “Add entity” button. You can add protein chains, DNA/RNA sequences, or small molecule SMILES strings.
Name your job (optional but useful for tracking multiple predictions)
Click Run prediction

Runtime: Single-chain predictions typically complete in 5–20 minutes depending on sequence length. Complex predictions take longer.

Step 3: Download and open the results

When your prediction completes, you’ll receive an email. Open the job and download the results as a zip file. The key files:

*_model_0.cif (or .pdb) — the 3D structure in Crystallographic Information File format. Open this in a molecular viewer.
*_summary_confidences_0.json — the confidence metrics
*_full_data_0.json — full residue-level pLDDT scores

Recommended free molecular viewers:

Mol Viewer* (web-based, no installation: molstar.org) — supports CIF format, modern interface
PyMOL (free educational version) — the standard for publication-quality figures
ChimeraX — from UCSF, excellent for large complexes

Step 4: Interpret confidence scores

This is the most important step researchers skip. A predicted structure is not equivalent to an experimentally determined structure — confidence scores tell you how much to trust each region.

pLDDT (per-residue Local Distance Difference Test)

pLDDT scores each individual amino acid residue from 0 to 100. In the 3D viewer, residues are colored by pLDDT:

Score	Color	Interpretation
> 90	Dark blue	Very high confidence — trust this region
70–90	Light blue	Good confidence — mostly reliable
50–70	Yellow	Low confidence — treat with caution; may be disordered
< 50	Orange/red	Very low confidence — likely intrinsically disordered or a prediction failure

Critical point: Low pLDDT does not always mean the prediction failed — it often means the region is intrinsically disordered in solution and AlphaFold is correctly signaling that no stable structure exists. Disordered regions are biologically real, not prediction errors.

PAE (Predicted Aligned Error)

The PAE plot is a heatmap showing the predicted position error between every pair of residues. It looks like a matrix — residue i on one axis, residue j on the other.

Dark green (low error) at position (i, j): AlphaFold is confident about the relative positions of residues i and j
Light/yellow (high error): the relative position is uncertain

How to read it:

Dark green diagonal blocks indicate confidently predicted domains
Off-diagonal dark green blocks indicate confidently predicted domain–domain contacts
If the PAE plot shows high error between two parts of your protein, their relative positioning is uncertain — even if both regions individually have high pLDDT

PAE is particularly important for multi-chain complexes: a dark green block at the interface between chain A and chain B means AlphaFold is confident about how the chains interact.

Step 5: Common interpretation mistakes to avoid

Mistake 1: Treating predicted structures as experimental structures Predicted structures are models. For drug design, functional inference, or mechanistic claims, experimental validation (crystallography, cryo-EM, NMR) remains essential.

Mistake 2: Ignoring low-pLDDT regions If you truncate your view to only high-confidence regions, you may miss functionally important disordered regions. Disordered loops often contain post-translational modification sites, interaction motifs, or are required for function.

Mistake 3: Misinterpreting high pLDDT in multi-chain predictions In a complex prediction, high pLDDT in each chain doesn’t mean the interface is correctly predicted. Always check the PAE heatmap across chains to assess interface confidence.

Mistake 4: Using predicted structures to make definitive conclusions about ligand binding without experimental validation Protein–ligand predictions from AlphaFold 3 are useful for hypothesis generation. They are not equivalent to crystallographic binding mode determination.

Step 6: What to do with the structure

For functional annotation: identify the structural fold using DALI or Foldseek to find similar known structures — the fold may suggest function even if sequence similarity to characterized proteins is low.

For mutagenesis design: look at pLDDT in the region of interest. High-confidence structural predictions around an active site or interface residue can inform targeted mutagenesis decisions.

For protein design: use the AlphaFold prediction as a starting hypothesis, then validate engineered variants experimentally or by running the designed sequence back through AlphaFold (self-consistency check).

For reporting: always include the pLDDT coloring in structural figures and note that the structure is “predicted by AlphaFold 3” rather than “determined by X-ray crystallography” or “cryo-EM.”

When to use ESMFold instead

Use ESMFold (via Hugging Face or local install) when:

You need to predict structures for hundreds or thousands of sequences — ESMFold is ~60× faster than AlphaFold per sequence
You’re annotating a metagenomic or genomic dataset
MSA computation is a bottleneck (ESMFold uses no MSA)

Use AlphaFold when:

You need the highest available accuracy for a specific important target
You’re predicting a multi-chain complex or protein–ligand interaction
The target has many known homologs (where MSA information improves AlphaFold predictions significantly)

Tool page: AlphaFold
Tool page: ESMFold — for high-throughput prediction
Tool page: RFdiffusion + ProteinMPNN — for designing new proteins rather than predicting existing ones
Comparison: AlphaFold vs. ESMFold vs. RFdiffusion
Field Guide: Structural Biology