Beginner's Step-by-Step Guide

The basic workflow

There are three routes to an AlphaFold structure, depending on your protein and your needs:

Start with Route A. Most proteins from common model organisms (human, mouse, yeast, E. coli, Arabidopsis, fly, worm, zebrafish) are already in the database. Only proceed to B or C if your protein isn't there, or you need a complex prediction.

Finding your protein on UniProt

UniProt (Universal Protein Resource) is the primary database of protein sequences and annotations. Every protein has a unique UniProt accession ID — a short code like P00533 (human EGFR) or P68871 (human haemoglobin beta chain). This ID is what the AlphaFold database uses to organise entries.

Checking the AlphaFold database

ColabFold: AlphaFold in a Colab notebook

ColabFold (developed by Milot Mirdita, Martin Steinegger, and colleagues) provides AlphaFold 2 and RoseTTAFold predictions through Google Colab notebooks — free Jupyter notebooks that run on Google's cloud computers with free GPU access.

ColabFold is faster than vanilla AlphaFold (uses a faster MSA search method called MMseqs2) and significantly easier to run — no local installation required.

AlphaFold Server: official predictions including AlphaFold 3

Google DeepMind provides an official prediction server at alphafoldserver.com. This uses AlphaFold 3 and is the only free way to get AlphaFold 3 predictions for complexes (protein + DNA, protein + RNA, protein + small molecule).

Limits: Free accounts are limited to 20 jobs per day. Commercial use of the server is restricted under the terms of service — it's intended for non-commercial academic research.

For protein-only predictions, the database and ColabFold are usually more convenient. Use the server when you specifically need AlphaFold 3 features (complex prediction, ligand binding).

Visualising structures in Mol*

Mol* (pronounced "mol-star") is a modern, web-based 3-D molecular viewer developed by RCSB PDB, PDBe, and others. It runs in your browser with no installation required and handles large structures well.

Alternative visualisation tools

Common beginner mistakes to avoid

• Using the prediction without checking confidence first. Always look at pLDDT before analysing any structural feature. Orange regions are not reliable.
• Forgetting to check if an experimental structure already exists. Before running ColabFold for 30 minutes, check the PDB (rcsb.org) for your protein. An experimental structure is always preferable for high-stakes work.
• Ignoring the PAE for multi-domain proteins. Even if each domain looks blue, the relative orientation of domains may be uncertain. Always download and examine the PAE JSON for multi-domain predictions.
• Comparing AlphaFold structures to each other as if they're experimental. Two AlphaFold predictions for slightly different sequences will differ — but you can't be sure whether the differences are real structural differences or prediction artefacts.
• Not specifying the organism. If you search by protein name without specifying organism, you may get a bacterial or yeast version rather than the human one. Always confirm the UniProt accession before downloading.
• Submitting very long sequences to ColabFold without splitting into domains. Proteins >1,500 residues are slower and sometimes less accurate. Check if the protein has well-defined domains you can predict separately, then compare.