AlphaFold DB Reference

pLDDT (predicted Local Distance Difference Test) is a per-residue confidence score from 0 to 100. Scores above 90 (blue) indicate very high confidence where the backbone and side chains are reliable. Scores 70-90 (cyan) mean the backbone is well-predicted. Scores 50-70 (yellow) suggest low confidence where only the fold topology may be correct. Scores below 50 (orange) often indicate intrinsically disordered regions that genuinely lack stable structure, rather than prediction errors.

The PAE matrix is an NxN grid where entry (i,j) represents the expected positional error (in Angstroms) of residue j when the structure is aligned on residue i. Low values (below 5 Angstroms) between two residues mean their relative position is predicted with high confidence. Look for dark blue diagonal blocks indicating well-defined domains, and light/red off-diagonal regions indicating uncertain inter-domain orientations. PAE is critical for assessing whether domain-domain arrangements are meaningful.

pTM (predicted TM-score) estimates the overall topology accuracy of a single-chain prediction, where values above 0.8 indicate high confidence in the global fold. ipTM (interface predicted TM-score) is specific to AlphaFold-Multimer and assesses confidence in the protein-protein interface. The ranking confidence for multimer predictions is calculated as 0.8*ipTM + 0.2*pTM. Both metrics help determine whether the predicted structure is trustworthy enough for downstream applications like drug design or interaction analysis.

Query the API at https://alphafold.ebi.ac.uk/api/prediction/{UniProt_accession} (e.g., /api/prediction/P04637 for p53). The JSON response includes direct download URLs for the PDB file, mmCIF file, and PAE JSON. You can also construct file URLs directly: https://alphafold.ebi.ac.uk/files/AF-P04637-F1-model_v4.pdb for PDB, replacing .pdb with .cif for mmCIF or -predicted_aligned_error_v4.json for PAE data.

Use ColabFold on Google Colab, which combines MMseqs2-based MSA generation (no large database download needed) with AlphaFold2 prediction. Install with pip install colabfold[alphafold] and run colabfold_batch input.fasta output_dir/. A typical single-chain protein takes about 10 minutes on a Colab GPU. For the simplest approach, use the ColabFold notebook directly at github.com/sokrypton/ColabFold, which provides a GUI-like interface for sequence input and result visualization.

AlphaFold repurposes the B-factor (temperature factor) column in PDB files to store pLDDT confidence scores, since predicted structures do not have experimental B-factors. The value 72.50 in the last numeric column of an ATOM record represents pLDDT=72.50, not crystallographic thermal motion. This is important to remember when using analysis tools that interpret B-factors, as you need to treat these values as confidence scores. The mmCIF format stores pLDDT in dedicated _ma_qa_metric fields instead.

AlphaFold3 extends predictions beyond single proteins and protein complexes to include protein-DNA complexes, protein-RNA complexes, protein-ligand interactions, ions, and modified residues. It uses a diffusion-based architecture instead of the structure module used in AlphaFold2. Access is available through the AlphaFold Server (golgi.sandbox.google.com). This enables modeling of transcription factor-DNA binding, ribonucleoprotein complexes, and drug-target interactions that were not possible with AlphaFold2.

Load the PDB file with "load AF-P04637-F1-model_v4.pdb", then apply pLDDT coloring with "spectrum b, blue_white_red, minimum=50, maximum=100". This maps high-confidence regions (pLDDT near 100) to blue and low-confidence regions to red. In ChimeraX, use "open alphafold:P04637" followed by "color bfactor palette alphafold" which applies the standard AlphaFold blue-cyan-yellow-orange palette. For PAE visualization, use Python with matplotlib: load the JSON, extract the predicted_aligned_error array, and display with plt.imshow using a blue-white-red colormap.