How to Interpret AlphaFold Output: pLDDT Scores, PAE Plots and Model Quality
Getting an AlphaFold prediction is easy. Knowing what to trust — and what to ignore — requires understanding two metrics that most tutorials mention but rarely explain: pLDDT and PAE. This article teaches you to read both, correctly, before using any predicted structure in your research.
Why confidence metrics exist
AlphaFold doesn’t just predict structure — it predicts how confident it is in each part of that structure. This is one of the things that makes it genuinely useful for research rather than just impressive: you don’t have to treat the output as a black box. The confidence scores tell you which regions to trust and which to treat skeptically, before you do any experimental work.
The two primary confidence outputs are pLDDT (per-residue confidence in atomic positions) and PAE (confidence in relative positioning between pairs of residues). They answer different questions and both need to be checked before using any AlphaFold structure downstream.
_scores_rank_001.json file, and PNG plots of both are saved automatically alongside the PDB files.
pLDDT: per-residue confidence explained
pLDDT stands for predicted Local Distance Difference Test. It is a score from 0 to 100 assigned to every residue, estimating how accurately that residue’s atomic positions would match an experimental structure, if one existed. Higher is better. It is stored in the B-factor column of AlphaFold PDB files.
The most important thing to understand about pLDDT: a low score does not always mean AlphaFold made a mistake. It often means a region is genuinely disordered — it has no single stable position in solution, so there is no “correct” structure to predict. AlphaFold correctly recognized the uncertainty and reported it. A disordered loop that flaps freely will always have low pLDDT, even in a perfect prediction.
Visualizing pLDDT in PyMOL
AlphaFold stores pLDDT in the B-factor column of the PDB file. Load the structure in PyMOL and color by B-factor to see confidence mapped onto the 3D structure:
# Load the AlphaFold PDB
load AF-P04637-F1-model_v4.pdb
# Color by pLDDT stored in B-factor column
# Blue = very high confidence, red = very low
spectrum b, blue_cyan_yellow_orange_red, minimum=50, maximum=100
# Show as smooth ribbon
show cartoon
hide lines
set cartoon_fancy_helices, 1
# Visually identify low-confidence regions
# Red/orange regions = exercise cautionFor the AlphaFold Database viewer, the default coloring is already by pLDDT — dark blue is high confidence, orange is low. You can toggle this view on the protein page.
PAE: predicted aligned error explained
PAE (Predicted Aligned Error) is a matrix where each cell (i, j) contains AlphaFold’s estimate of the position error for residue j when the structure is aligned on residue i. The value is in ångströms. Lower PAE = more confident in the relative positions of those two residues.
Where pLDDT answers “how confident is AlphaFold about this residue’s position?”, PAE answers “how confident is AlphaFold that these two residues are in the correct position relative to each other?” These are different questions — and for multi-domain proteins and complexes, PAE is often more informative than pLDDT alone.
The PAE plot is displayed as a colored heatmap: dark blue (low PAE, high confidence) means the relative positions of two residues are well-constrained. Light blue, white, or green (high PAE) means their relative positioning is uncertain — even if each residue’s absolute position (pLDDT) looks fine.
Reading PAE patterns
Three patterns appear repeatedly in PAE plots, each with a clear biological interpretation:
What to do with low-confidence regions
Low pLDDT and high PAE are not failures — they’re information. Here is how to handle them appropriately:
- Trim low-pLDDT terminal regions before docking or MD
- Report which regions have low confidence in your methods section
- Treat low-pLDDT loops as disordered — consider their flexibility in interpretation
- Use PAE to identify which domains are reliably arranged relative to each other
- Validate key structural features with independent experimental data
- Run MD simulation on the high-confidence core to explore dynamics
- Dock ligands to a binding site with pLDDT below 70
- Report specific side chain interactions from low-confidence regions
- Use high inter-domain PAE structures as if domain arrangement is known
- Cite predicted structures without stating confidence scores
- Assume a low-pLDDT region is wrong — it may be correctly identified as disordered
- Use the lowest-confidence model just because it shows a feature you want to see
Handling low-confidence regions in PyMOL
For docking and MD preparation, trim or exclude residues below your confidence threshold. In PyMOL, select and remove low-pLDDT regions before saving:
# Select residues with pLDDT (B-factor) below 70
select low_conf, b < 70
# Inspect which residues are selected
iterate low_conf, print(resi, resn, b)
# Remove low-confidence residues if they are terminal regions
# (only remove loops/termini — not internal low-confidence regions
# which may be genuinely disordered but structurally important)
remove low_conf and (resi 1-15 or resi 485-500)
# Save the trimmed structure
save protein_trimmed.pdbWhich of the 5 models should you use?
ColabFold generates 5 models ranked by a composite confidence score (pTM or ipTM for complexes). In most cases, use rank_1 — the highest-confidence model. But the ranking deserves scrutiny:
| Model | When to use it |
|---|---|
| rank_1 (default) | Use in almost all cases. Highest composite confidence score. This is what “best AlphaFold model” means. |
| rank_2–3 | Compare with rank_1 — if they agree, the prediction is robust. Large differences between rank_1 and rank_2 suggest conformational ambiguity. |
| rank_4–5 | Rarely useful for structural analysis. Can indicate alternative conformational states if dramatically different from rank_1. |
The inter-model agreement is itself a confidence signal. If all five models superimpose well (low pairwise RMSD), AlphaFold is converging on a single answer — the prediction is robust. If the five models are dramatically different, especially in a region of interest, the structural uncertainty is real and should be reflected in how you use and report the prediction.
align to compare them: align rank_002, rank_001 — note the RMSD. If all five align to rank_1 with RMSD below 1 Å, the prediction is highly consistent. If some models differ by 3+ Å in a loop or domain, treat that region’s conformation as uncertain regardless of pLDDT.
How to report AlphaFold confidence in a paper
Reviewers increasingly expect explicit reporting of confidence metrics for any AlphaFold structure used in a study. A minimal complete reporting statement looks like this:
Interpreting AlphaFold output in four steps
First, check pLDDT: color the structure and identify any regions below 70 — especially around your site of interest. Second, check the PAE plot: for multi-domain proteins, look for high inter-domain PAE that would make the relative domain arrangement unreliable. Third, compare the five models: strong agreement means a robust prediction; disagreement signals real conformational uncertainty. Fourth, exclude or caveat low-confidence regions in your analysis and methods section — not as a weakness, but as rigorous reporting of what is and isn’t known.