Gains:
- Being able to understand the work of AlphaFold, confidence scores such as pLDDT and PAE, and that a structure is a 'prediction', and interpret the structure correctly with artificial intelligence
- Ability to recognize areas with low confidence scores (flexible/irregular) and avoid over-interpretation and compare with experimental evidence
- Ability to apply the discipline of treating structure prediction as a hypothesis and connecting functional claims to experimental verification.
To understand what a protein does, it is often necessary to know its three-dimensional folded structure; because function arises from structure. For decades, determining protein structure experimentally (X-ray crystallography, cryo-electron microscopy) would take months to years. AlphaFold (an artificial intelligence system developed by DeepMind that predicts protein structure from amino acid sequences) reduced this to minutes and made the predicted structures of hundreds of millions of proteins public. In this unit you will learn how to read AlphaFold output, how to interpret confidence scores, and how to safely use an LLM in this interpretation.
Critical distinction: AlphaFold is a structure prediction tool and its output is a prediction, not experimental truth. LLM (a chat model like ChatGPT) is not AlphaFold itself; It cannot "remember" the coordinates of a protein. Use LLM to interpret and explain AlphaFold output; retrieve the structure itself from the AlphaFold database or tool.
Basic concepts
- Primary structure: Amino acid sequence (letter chain of the protein).
- Secondary/tertiary structure: Helix (alpha-helix), sheet (beta-sheet) and three-dimensional folding of the chain.
- pLDDT: AlphaFold's local confidence score for each amino acid, ranging from 0 to 100. 90+ means very high confidence, 70-90 means good, 50-70 means low, and below 50 means very low confidence. Regions of low pLDDT are generally "disordered" regions (without a distinct fold).
- PAE (Predicted Aligned Error): Predicted error in the relative position of two regions; It shows how reliable the placement of domains is relative to each other.
- PDB: Database and file format in which experimental protein structures are stored.
Reading AlphaFold output: step by step
1. Select the correct protein and sequence. Identify protein with UniProt (protein information database) number; Find the structure with this number in the AlphaFold database.
2. Look at the pLDDT map. See which parts of the structure are predicted with high confidence (blue) and which are predicted with low confidence (orange/yellow). Be cautious about comments in low-trust areas.
3. Read the PAE chart. PAE shows whether the relative arrangement of domains in a multi-domain protein is reliable. A high PAE means that the relative position of the fields is uncertain.
4. Compare with the experimental structure. Match the experimental structure in the PDB if available. If the AlphaFold prediction is close to experimental, your confidence increases.
5. Interpret the variant effect. When interpreting the effect of an amino acid change on the structure, consider the pLDDT and functional significance of that region together (active site, binding pocket).
Tip: Low pLDDT does not always mean "wrong guess"; It is often a sign that the area is actually intrinsically disordered. So low confidence sometimes reflects an accurate biological fact. Also check the sequence with irregularity prediction tools to distinguish this.
three mini cases
Case 1 — Overinterpreting the low confidence zone. One student claimed that a “loop” in the AlphaFold structure fit into a particular pocket, and the AI confirmed this. However, when the student looked at the pLDDT, he saw that the score of that stitch was 42 (very low); so his position was unreliable. The claim was withdrawn. Lesson: always check the local trust score before commenting.
Case 2 — Made-up coordinate. A researcher asked LLM the 3D coordinate of a particular amino acid of a protein; LLM gave convincing numbers to three decimal places. When the researcher compared them to the actual coordinates in the AlphaFold PDB file, he saw that they were completely fabricated. LLM cannot "remember" the coordinate; coordinates must be read from the file.
Case 3 — Confirmation gained. A PhD student wondered whether one variant (p.Gly12Val) was close to the active site of the protein. YZ reminded that active site residues are specified in UniProt; The student opened UniProt and found the active site, then measured with a structure viewer (PyMOL) that Gly12 was 8 angstroms away from this site in the AlphaFold structure. Numerical measurement embodied the "near" claim.
Example: reading pLDDT from a structure file
# In the AlphaFold PDB file, pLDDT is stored in the B-factor column. from Bio.PDB import PDBParserimport numpy as npyapi = PDBParser(QUIET=True).get_structure("AF", "AF-P04637-F1.pdb")plddt = [atom.bfactor for atom in structure.get_atoms() if atom.name == "CA"]print("Average pLDDT:", round(np.mean(plddt), 1))print("Low confidence (<70) residue rate (%):", round(100 * np.mean(np.array(plddt) < 70), 1))
This allows you to see numerically the overall reliability of the structure before commenting.
Four copyable templates
1) Structure interpretation (with confidence score):
Your role: structural biology assistant. Help me interpret the AlphaFold structure of the UniProt [number] protein. Answer these questions but coordinate/score FITTING: which regions do we expect high/low pLDDT, what does PAE show, is there an experimental structure (PDB), how do I compare? I will read the values from the file.
2) Variant-structure effect:
Help me interpret the possible effect of the following variant on the protein structure: [p. Give me what evidence to look for instead of the outright "destructive" claim.
3) pLDDT/PAE description:
Explain with an example the difference between pLDDT and PAE, which one should I look at and when in a variant analysis. Consider single-domain and multi-domain protein cases separately.
4) Structure comparison plan:
I want to compare the AlphaFold prediction with the experimental PDB structure. How to calculate RMSD (mean deviation between structures), what tool do I do it with (PyMOL,Biopython), what threshold counts as "good overlap"? Give a step by step plan.
Weak prompt / Strong prompt
Weak: "Draw the structure of this protein and tell the effect of p.Arg273His."
Problem: LLM cannot draw structure/fit coordinates; confidence score is not taken into account; The claim of definitive effect would be unfounded.
Strong: "I downloaded the AlphaFold build for UniProt P04637 myself. Looking at the pLDDT of the p.Arg273His residue and its functional annotation in UniProt, tell me step by step how I should interpret its effect; give me what evidence to look for instead of the definitive claim."
Why it is powerful: The structure is taken from the source, the trust score is put at the center, the claim is linked to evidence.
Question
Does AlphaFold deliver?
Where/how to confirm
3D fold prediction
Yes
AlphaFold DB / file
local trust
Yes (pLDDT)
B-factor column
Interdomain location
Yes (PAE)
PAE chart
Experimental real structure
no
PDB (experimental)
Exact functional impact of the variant
no
Functional study + expert
Common mistakes
- Skipping the confidence score. Interpretation in the low pLDDT region is unreliable.
- Mistaking a prediction for empirical fact. AlphaFold output is an estimate; Critical decisions require empirical evidence.
- Asking for coordinates from LLM. Coordinates are read from the file, not memorized.
- Ignoring PAE. In multidomain proteins, the relative position of the domains can be uncertain.
- Immediately considering low confidence as a "mistake". Sometimes that area is really uneven.
Caution: AlphaFold builds present a "static" image; Remember that proteins are actually mobile molecules that can enter multiple conformations. No single structure fully reflects dynamic behavior.
Depth: Where AlphaFold is structurally quiet
AlphaFold is powerful, but knowing what it can't do is half the battle to using it safely. First, the standard AlphaFold folds a single protein in space; It does not model ligands, ions, cofactors and drug molecules. The zinc ion in the active site of an enzyme or a substrate does not appear in the structure; Therefore, a comment such as "this pocket is empty, dysfunctional" may be misleading. Second, it does not directly model the effect of mutation: even if you introduce two variants close to the same sequence, AlphaFold usually produces almost the same structure; You can't prove the claim "this variant breaks the structure" by comparing two predictions from AlphaFold. Third, it does not take into account post-translational modifications (such as phosphorylation, glycosylation) and the actual environment of membrane proteins within the membrane.
Case in point: one team claimed from an AlphaFold image that a variant close to the ATP binding pocket in a kinase enzyme “closes the pocket”; artificial intelligence also confirmed. When an experimental crystal structure (PDB) was opened, it appeared that a cofactor in that region that AlphaFold had not modeled was already shaping the pocket — the variant's effect was from a completely different mechanism. Second case: a researcher hypothesized that a "loop" between two fields connects the two fields; The PAE map showed high error (unclear relative position) between those two areas, so the connection claim was unfounded. Reading the PAE prevented a faulty mechanism story.
5) AlphaFold borders control template:
Before considering the following structural claim, list what limitations of AlphaFold come into play in this question: [claim]. Tell me what additional evidence (experimental structure, functional study) I need, especially in terms of ligand/ion/cofactor deficiency, lack of mutation modeling, and PAE uncertainty.
In summary
- AlphaFold is a powerful tool that predicts structure from sequence, but its output is a guess; should be read together with confidence scores.
- pLDDT indicates local trust, PAE indicates cross-domain location trust; Look at both before commenting.
- LLM cannot "remember" coordinates or structure; get the structure from AlphaFold/PDB, use LLM in comment.
- Empirical evidence and expert judgment are indispensable in critical decisions.
Application task
Download the AlphaFold structure of a protein (e.g., a well-studied tumor suppressor) by UniProt number. Calculate the average pLDDT and low safe residue ratio with the above code snippet. Then choose a variant and ask the AI for an interpretation plan with the 2nd template; Check the pLDDT and UniProt functional annotation of that residue yourself. Tie your claim to a numerical confidence value.
checklist
- [ ] I got the structure from AlphaFold/PDB with the UniProt number.
- [ ] I read pLDDT and PAE before commenting.
- [ ] I did not ask LLM to make up coordinates/scores.
- [ ] I linked the variant interpretation to the confidence score of the corresponding residue.
- [ ] I compared it with the experimental structure, if available.
- [ ] I supported the critical decision with expert judgment and empirical evidence.