How to Use the AlphaFold Database?

Quick Index:

Q1 What's in the database · Q2 How to read pLDDT · Q3 What is PAE · Q4 Batch download · Q5 What to do if you can't find a protein · Q6 Can AlphaFold structures be used directly for mutation design · Q7 Linking with other databases · Q8 What other resources are there

Q1 What exactly is in the AlphaFold database?

The AlphaFold Database is an open-access resource for protein structure predictions. Its core data comes from the AlphaFold 2 model predicting each UniProt reference sequence one by one. As of 2026, AlphaFold DB v6 contains over 241 million predicted protein structures, covering humans and over a hundred model organisms. According to the official update in May 2026, the database includes 1.7 million high-confidence homodimers and nearly 80,000 high-confidence heterodimers, plus around 26 million lower-confidence complexes available for bulk download via FTP, making a total of roughly 30 million predicted complexes.

Each entry in the database includes four types of core files: PDB/mmCIF structure files (atomic coordinates), PAE matrix files (predicted alignment errors), pLDDT confidence files (residue-level confidence scores), and sequence & annotation files (corresponding UniProt ID, species, function, and other metadata).

A common misconception: the AlphaFold database ≠ the AlphaFold model itself. The database only contains completed structure predictions and doesn’t provide online prediction services. If you want to predict a new sequence or a mutant, you need to use a locally deployed model or a dedicated prediction service.

Q2 How high does the pLDDT score need to be to be considered 'reliable'?

This is the most commonly asked question when using the AlphaFold database. The answer isn’t a simple threshold—it depends on the range:

- pLDDT > 90 (very high confidence): the backbone is highly reliable, around 80% of side-chain conformations are accurate, but about 7% of side chains (especially in binding pockets/catalytic residues) may still be off;

- pLDDT 70–90 (high confidence): backbone conformation is basically correct, side chains might have deviations, suitable for domain analysis and locating conserved sites;

- pLDDT 50–70 (low confidence): the backbone is only roughly outlined, interpret with caution, usually corresponds to intrinsically disordered regions (IDRs) or flexible loops;

- pLDDT < 50 (very low confidence): basically not reliable, usually corresponds to disordered regions or sequences lacking homologous templates, not recommended for any structure-guided analysis.

Studies estimate that about 30–40% of predicted residues fall into the low-confidence range of pLDDT < 70.

Practical tip: don’t just look at the overall average pLDDT; you must check the regional distribution. Many proteins have a low overall pLDDT, but their core domains can be very reliable. The low average is often caused by terminal tails and flexible loops.

Q3 What is the PAE plot? How is it different from pLDDT?

The difference between pLDDT and PAE

Many people focus only on pLDDT and ignore PAE, which is the second biggest misconception when using the AlphaFold database.

pLDDT measures the local confidence of a single residue and answers "Is the position of this amino acid accurate?"; PAE (Predicted Aligned Error, unit: Å) measures the confidence of the relative position between two residues and answers "Is the relative position between residue A and residue B accurate?" The lower the PAE value, the more reliable the relative positions of the two residues.

The PAE map is especially important for multi-domain proteins—each domain may have high pLDDT, but their relative orientation could be completely off (corresponding to a high PAE value). If your research involves domain-domain interactions or allosteric site analysis, make sure to check the PAE matrix and not just judge based on pLDDT.

Q4 How to download AlphaFold structures in bulk?

If you only want a few sequences, you can simply search for the UniProt ID on the AlphaFold DB website and download. But if you're dealing with dozens or hundreds of sequences, manual download is inefficient. Here are three recommended bulk download methods:

Option 1: Official bulk search interface. The official site provides a batch download tool that supports submitting up to 100 UniProt IDs at a time, returning a compressed file—suitable for medium-scale needs.

Option 2: REST API for programmatic access. The database offers a RESTful API for batch querying and downloading. You can use Python's requests library for automation, which is suitable for researchers with programming skills.

Option 3: Full dataset download. If you need whole-genome or species-wide structure predictions, you can download the entire database from EMBL-EBI FTP or cloud platforms. The complete raw database is tens of TB in size; downloading just a single species proteome is usually in the GB to tens of GB range, so you can choose according to your needs.

Tip: Make sure to remove duplicate sequences before bulk downloading. Each UniProt entry corresponds to a unique AlphaFold model, and downloading duplicates just wastes time and storage.

Q5 What should I do if I can't find my protein in the AlphaFold database?

In the following situations, the AlphaFold database might not have a corresponding predicted structure:

- The sequence is not in the UniProt reference set: for example, newly discovered metagenomic proteins, artificially designed sequences, or mutant sequences;

- The sequence is very long (thousands of amino acids or more), which limits prediction quality and coverage;

- The database is not fully up-to-date, so newly published sequences might not have been included yet.

If you encounter these cases, there are three possible approaches:

Approach 1: Homology modeling search. Use BLAST to find the structure in the database with the highest homology as a reference scaffold.

Approach 2: Local or online prediction. Use structural prediction tools to predict the new sequence yourself.

Approach 3: Function-oriented design platforms. If your ultimate goal isn’t just to "see the structure" but to do mutation design, stability optimization, or other functional modifications, you can directly use AI protein design agent like MatwingsVenus™ (Xiaowu™) — structure prediction is automatically handled as a built-in step, so there’s no need to manually download PDB files, set up visualization tools, or write analysis scripts.

Q6 Can AlphaFold structures be used directly for mutation design?

Solutions for Missing-Target Queries

This is the practical question that protein engineering researchers care about the most. The answer is: it can be used as a starting point, but not directly as an experimental structure.

There are three reasons:

Limited side-chain accuracy: Even if the pLDDT score is very high, the side-chain conformations, especially the catalytic residues and the fine geometry of binding pockets, may still deviate from the real structure;

Lack of ligands and cofactors: AlphaFold predicts the apo structure, without substrates, metal ions, cofactors, etc., which are actually key components of the active site;

Limitations of static structures: It predicts a single conformation and cannot reflect the dynamic changes of proteins, while function is often closely related to conformational dynamics.

Therefore, the proper way to use AlphaFold structures for mutation design is to first use the predicted structure for preliminary site screening and feasibility assessment, and then cross-validate with molecular dynamics simulations or site-directed mutagenesis experiments. For catalytic pocket optimization that requires high precision, it is recommended to primarily rely on experimentally resolved structures (PDB) and use AlphaFold predictions as a supplement.

For engineering projects that require multi-objective optimization, you can also directly submit sequences on the MatwingsVenus™ (Xiaowu™) platform. The system will automatically complete the full workflow of structure prediction → mutation hotspot identification → combinatorial optimization → experimental verification, which is far more efficient than manually downloading structures and designing mutations yourself.

Q7 How to link and use with other databases?

The value of the AlphaFold database often becomes clear when used together with other databases:

- Linking with UniProt: Each AlphaFold entry is associated with the corresponding UniProt ID, allowing one-click access to functional annotation, subcellular localization, post-translational modifications, etc., helping to interpret the functional significance of structural features;

- Linking with PDB: If a protein has an experimental structure, the AlphaFold predicted model can be overlaid and compared with the PDB structure to assess prediction quality and identify discrepant regions;

- Linking with functional databases: By combining with Pfam, InterPro, GO, and other functional annotations, one can locate functional domains, active sites, disease-related mutations, etc., achieving a complete analysis chain from sequence to structure to function.

A practical workflow: find the target protein in UniProt → obtain the AlphaFold predicted structure → annotate Pfam domains and conserved sites on the structure → make a preliminary assessment of the structural impact of possible mutations.

Q8 Besides AlphaFold, what other structure prediction resources are available?

Although the AlphaFold database has the broadest coverage, it’s not the only option. Depending on your research needs, there are two complementary types of structure prediction resources:

Super-fast structure prediction. Lightweight structure prediction methods based on protein language models are significantly faster than AlphaFold. They’re great for large-scale initial screening but have slightly lower overall accuracy.

Complex structure prediction. The AlphaFold DB mainly contains monomer structures. If you need predictions for protein complexes or protein-ligand complexes, you can use dedicated complex prediction tools.

Overall, the choice of tool really depends on your goal—whether it’s to “see the structure” or “engineer functionality.” If it’s the former, AlphaFold DB along with structure analysis tools is enough. If it’s the latter, structure prediction is just the first step, and using AI platforms focused on functional optimization is often more efficient.

Conclusion: The right way to use the AlphaFold database

Finally, here are three key principles for using the AlphaFold database:

First, use confidence regions wisely. Use high-confidence regions for detailed analysis and treat low-confidence regions as reference. Look at both pLDDT and PAE.

Second, combine tools. Only by combining UniProt, PDB, and functional annotation databases does the structure have real biological meaning.

Third, be clear about your purpose. If it’s to see the structure, the database is enough. If it’s to modify function, treat structure prediction as an intermediate step, and use platforms targeted at functional goals for higher efficiency.

Master these three points, and the AlphaFold database can truly become a powerful tool in your research, rather than just looking like a pretty “electron cloud map”.