Lesson 12 of 14 · 11 min

Limitations: disorder, multimers and conformational change

An AlphaFold model renders every residue with the same crisp geometry, which makes even wrong regions look authoritative. This lesson maps the three places predictions mislead most: intrinsic disorder, conformational dynamics, and unmodeled chemistry. It closes with how to judge predicted complexes rather than trust them on sight.

Disorder as low pLDDT

Intrinsically disordered regions have no fixed fold, so the network cannot place them and reports low confidence. They surface as extended, low-pLDDT spaghetti whose coordinates are meaningless, yet a pLDDT below about 50 is itself a strong predictor of disorder. Treat those stretches as a prediction of disorder, not as a structure to be measured.

bash

# pLDDT is written into the B-factor column of AlphaFold PDB output.
# Extract the per-residue CA pLDDT and flag the low-confidence stretches.
awk 'substr($0,1,4)=="ATOM" && substr($0,13,4)==" CA " {
       res   = substr($0,23,4) + 0
       plddt = substr($0,61,6) + 0
       if (plddt < 50) printf "%d\t%.1f\n", res, plddt
     }' ranked_0.pdb

One static state

A prediction is a single static snapshot, not an ensemble, so alternate functional states are invisible: open versus closed transporters, active versus inactive kinases, apo versus holo enzymes. A deep MSA tends to collapse the model onto the dominant state. Subsampling the MSA raises diversity across seeds and can surface the minority conformation.

bash

# Sample alternate conformations by subsampling the MSA
# (Del Alamo et al., eLife 2022). Shallow MSA + many seeds = more diversity.
colabfold_batch \
  --num-seeds 16 \
  --max-seq 32 \
  --max-extra-seq 64 \
  --num-recycle 3 \
  transporter.fasta  msa_subsampled/

Warning: AlphaFold2 models only the polypeptide. Ligands, cofactors, metal ions, lipids and post-translational modifications are absent, so binding pockets can be collapsed or mis-shaped. Because the network is driven by the MSA, a single point mutation usually leaves the predicted backbone almost unchanged, so AF2 is not a reliable predictor of mutation-induced destabilisation or delta-delta-G. AlphaFold3 does place ligands and ions, but it still returns one static state.

Judging multimers

AlphaFold-Multimer predicts assemblies, but a plausible-looking interface can be entirely wrong. Overall pLDDT says nothing about how chains are placed relative to each other; judge that with ipTM and with the inter-chain block of the PAE matrix. Low off-diagonal PAE means the two chains are confidently positioned relative to one another.

python

import json
import numpy as np

scores = json.load(open(
    "complex_scores_rank_001_alphafold2_multimer_v3_model_1_seed_000.json"))
pae = np.asarray(scores["pae"])   # (L, L) predicted aligned error, in Angstrom
La  = 148                         # residues in chain A; the remainder is chain B

# Off-diagonal blocks are the chain A <-> chain B relationships (the interface).
inter = np.concatenate([pae[:La, La:].ravel(), pae[La:, :La].ravel()])

print(f"ipTM                 : {scores['iptm']:.2f}")
print(f"median interface PAE : {np.median(inter):.1f} A")
print(f"best interface PAE   : {inter.min():.1f} A")

ipTM                 : 0.41
median interface PAE : 24.6 A
best interface PAE   : 11.3 A

Assessing model quality: Ramachandran, clashes and MolProbity Putting predicted structures to work