Bio Structural Biology Alphafold Predictions

Version Compatibility

Reference examples tested with: biopython 1.83+, numpy 1.26+, requests 2.31+, matplotlib 3.8+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

AlphaFold Predictions

"Get the AlphaFold model for my protein and tell me which parts to believe" -> Fetch the precomputed AFDB entry by UniProt accession, then read its confidence files (pLDDT per-residue, PAE per-residue-pair) to decide which regions and which geometry are trustworthy.

• Python: requests.get(f'https://alphafold.ebi.ac.uk/api/prediction/{accession}') returns metadata with cifUrl/pdbUrl/paeDocUrl download links.

Governing Principle

An AFDB entry is ONE AlphaFold2 prediction of a SINGLE UniProt sequence modeled as an isolated chain in vacuum - a per-chain hypothesis of a dominant fold, not an experimental structure and not a biological state. It carries NO ligands, ions, cofactors, metals, or substrates (the pocket is apo even when the protein only folds around a cofactor), NO post-translational modifications, NO quaternary structure or biological assembly (it is a monomer even for obligate oligomers), NO alternative conformations (one static snapshot - kinases render in one activation state, transporters in one gate state), and NO membrane context. The cardinal error is using an AFDB coordinate file as an experimental holo complex instead of as a scored guess whose own confidence files tell the reader which parts to believe.

Three confidence traps ride on top of this. First, pLDDT is written into the B-FACTOR COLUMN but is per-residue CONFIDENCE (0-100, higher=better) with OPPOSITE polarity to a real B-factor - any tool that reads that column as thermal motion inverts the meaning, and feeding raw pLDDT-as-B into crystallographic refinement mis-weights it. Second, a long low-pLDDT stretch is usually a genuine INTRINSICALLY DISORDERED REGION (pLDDT is competitive with dedicated IDR predictors; Akdel 2022 Nat Struct Mol Biol 29:1056; Piovesan 2022 Protein Sci 31:e4466), not a modeling failure - trimming it as "junk" discards real biology. Third, pLDDT is LOCAL and per-residue: high mean pLDDT does NOT certify inter-domain placement. Two rigid domains can each be 95 pLDDT yet float at an unknown relative orientation - reading that is PAE's job. "Confident" bounds the structural self-consistency of the prediction, NOT its biological correctness.

Decision: pLDDT confidence bands

Band cutoffs 90/70/50 are the AFDB-defined thresholds (Jumper 2021 Nature 596:583). A very-low band is a disorder SIGNAL, not proof of error - a conditionally-folded binding region looks disordered in AFDB yet folds on binding a partner AFDB never sees.

Decision: downstream-use suitability

Foldseek structure search over AFDB (van Kempen 2024 Nat Biotechnol 42:243) is the transformative use - it encodes backbone into the 3Di alphabet and finds structural homologs invisible to sequence search (see alignment/structural-alignment). Molecular replacement needs the model PROCESSED first: phenix.process_predicted_model (Oeffner 2022 Acta Cryst D 78:1303) reads pLDDT from the B column, converts pLDDT->pseudo-B, trims below ~0.7 fractional pLDDT, and splits into PAE-defined domains. Docking into an AFDB pocket gives confidently wrong poses when the backbone is confident but the rotamers, gate state, or cofactor are not.

Decision: use the AFDB entry vs run a new prediction

Anything needing complexes, ligands, mutants, custom MSA depth, or a specific state points to modern-structure-prediction. AFDB is monomer-only and FIXED at deposition - a newer method or deeper MSA is not reflected.

Fetch the AFDB entry via REST metadata

Goal: Download the coordinate file and PAE for a UniProt accession without hard-coding a version suffix.

Approach: Query the prediction metadata endpoint, which returns the current download URLs (cifUrl, pdbUrl, paeDocUrl); the version token drifts (v4 -> v6 as of 2025) so the URLs are discovered, never assembled by hand.

import requests

def afdb_metadata(accession):
    url = f'https://alphafold.ebi.ac.uk/api/prediction/{accession}'
    r = requests.get(url)
    r.raise_for_status()
    return r.json()  # list; one object per fragment/isoform, empty if no model exists

def fetch_afdb(accession, out_dir='.'):
    entries = afdb_metadata(accession)
    if not entries:
        return None  # >2700-aa non-human proteins and non-UniProt sequences are often absent
    entry = entries[0]
    cif_text = requests.get(entry['cifUrl']).text
    pae_json = requests.get(entry['paeDocUrl']).json()
    cif_path = f"{out_dir}/AF-{accession}.cif"
    with open(cif_path, 'w') as f:
        f.write(cif_text)
    return cif_path, pae_json

result = fetch_afdb('P04637')  # human p53

Long proteins split into fragments AF-{accession}-F1-..., -F2-... (human > 2700 aa, ~1400-aa windows shifted by 200); afdb_metadata returns one entry per fragment. Relative placement ACROSS fragments is independent and must not be trusted.

Read pLDDT from the B-factor column

Goal: Extract per-residue confidence and classify each residue into a band.

Approach: Parse the model, read the B-factor field of the CA atom (that is where AFDB stores pLDDT), and map the score through the 90/70/50 cutoffs.

from Bio.PDB import MMCIFParser

def extract_plddt(cif_file):
    parser = MMCIFParser(QUIET=True)
    structure = parser.get_structure('af', cif_file)
    plddt = {}
    for residue in structure[0].get_residues():
        if residue.id[0] == ' ' and 'CA' in residue:
            # pLDDT rides in the B-factor column but is CONFIDENCE (0-100, higher=better),
            # opposite polarity to a thermal B-factor - never read it as motion
            plddt[residue.id[1]] = residue['CA'].get_bfactor()
    return plddt

def plddt_band(score):
    if score > 90:                       # 90: side-chain-trustworthy core (AFDB very-high cut)
        return 'very_high'
    if score >= 70:                      # 70: backbone-reliable fold (AFDB confident cut)
        return 'confident'
    if score >= 50:                      # 50: coarse topology only below this
        return 'low'
    return 'very_low'                    # <50: usually an intrinsically disordered region

plddt = extract_plddt('AF-P04637.cif')
mean_plddt = sum(plddt.values()) / len(plddt)
disordered = [res for res, s in plddt.items() if s < 50]  # candidate IDR, not error

A high mean does not license inter-domain claims - a globally 90-pLDDT model can still have two domains at an unconstrained orientation. Check PAE before measuring any inter-domain distance.