Gains:
- Understanding protein structure levels and what structure AlphaFold predicts from the sequence
- Ability to read pLDDT and PAE confidence scores correctly and interpret low confidence areas as 'uncertain/flexible' rather than 'incorrect'
- Understand the necessity of validating the structure prediction with experimental evidence (PDB, activity test) in high-consequence decisions.
Proteins are the working molecules of the cell: enzymes speed up reactions, transporters move molecules around, structural proteins keep the cell going. What a protein does is determined by its three-dimensional folded shape (structure). For decades, predicting the structure of a protein from its sequence was one of the most difficult problems in biology. In 2021, DeepMind's artificial intelligence system called AlphaFold made a historic leap forward in this problem, predicting the structure of hundreds of millions of proteins to near experimental accuracy. In this unit, we will discuss how to use AlphaFold and similar tools from a biologist's perspective and how much you can trust its output.
This is the most concrete achievement of artificial intelligence in biology; But even a predictive tool has its limits and the need for validation.
Levels of protein structure
- Primary structure: Amino acid sequence (order of letters).
- Secondary structure: Local folds; such as alpha helix and beta sheet.
- Tertiary structure: 3D shape of the entire chain.
- Quaternary structure: Combination of multiple chains.
AlphaFold predicts the tertiary structure (3D shape) from the primary structure (sequence). It does this through patterns it learns from the alignment (MSA) of evolutionarily related sequences.
Reading AlphaFold output
AlphaFold doesn't just give a structure; It also gives confidence scores for each prediction. Understanding these is the key to not blindly trusting:
- pLDDT: Local confidence score between 0-100 for each amino acid. Above 90 is very reliable, 70-90 is reliable, 50-70 is uncertain, below 50 is most likely a disordered region.
- PAE (Predicted Aligned Error): Inter-domain relative position confidence; It shows how reliable the placement of domains relative to each other is in large multidomain proteins.
Tip: When you see areas of low pLDDT (non-blue, orange/red) on an AlphaFold model, read them as “vague or flexible” rather than “incorrect.” These regions are often truly disordered protein fragments and have biological significance.
Step by step: examining a protein with AlphaFold
- Find the sequence: Get the sequence of your protein from UniProt.
- Get the structure: Search in AlphaFold DB (ready for most proteins) or run with ColabFold.
- Assess confidence: look at pLDDT and PAE; Which regions are reliable?
- Visualize: Inspect in 3D with PyMOL or py3Dmol.
- Interpret: Evaluate functional regions such as active site, binding pocket.
- Verify: Compare experimental structure (X-ray/cryo-EM in PDB) if available.
The artificial intelligence (LLM) writes the code in these steps (visualization with py3Dmol, summarizing scores) and explains the concepts. AlphaFold itself is a discrete structure prediction model; LLM helps you interpret its results.
Copiable prompt templates
Role: You are a structural biology assistant. Task: Explain the AlphaFold pLDDT and PAE scores to a graduate student. Explain what each score measures, what thresholds are considered reliable, and how low-scoring regions should be interpreted.
How do I run the protein sequence I received from UniProt in ColabFold? Explain the steps and the resource/time limits I need to be aware of. Sequence length: [N] amino acids.
Write a Python code that visualizes a PDB file (predicted.pdb) in 3D and colors it according to the pLDDT score with py3Dmol. Let it run in Jupyter, with comments.
I have the AlphaFold prediction and the experimental structure from the PDB. Give the code and comment that aligns the two and calculates the RMSD (mean square deviation). Explain how many angstroms below RMSD is considered good.
Weak prompt / Strong prompt
Weak: "Tell me the shape of this protein."
Strong: "I examined the AlphaFold model of the UniProt P04637 (human p53) protein. The DNA binding domain is high pLDDT (>90), the N-terminus and C-terminus are low pLDDT (<50). How should I interpret this pattern? Explain the possibility that the low-scoring ends are disordered regions and the functional significance of this; without making a definitive structure claim."
Difference: The powerful prompt has the real protein, real score pattern and comment request. The model interprets the data you provide rather than "remembering" it.
three mini cases
Case 1 — Mistaking the disordered region for an error: A student eliminated the low pLDDT region at the end of his protein because “AlphaFold guessed it wrong.” However, that region was also an area of irregular activation experimentally. The student interpreted the region correctly when the model explained that low pLDDT often reflects true elasticity.
Case 2 — Mutation effect exaggeration: A researcher claimed that a single amino acid change made a "huge difference" in the AlphaFold pattern. However, AlphaFold does not reliably predict the stability effect of single point mutations; differences may be model noise. The model recalled this limit and led to specific tools (e.g. stability prediction tools).
Case 3 — Gain by validation: A team aligned the AlphaFold prediction with the experimental structure in the PDB; RMSD turned out to be 1.2 angstroms (very good fit). This allowed them to rely on the prediction and hypothesize a binding pocket, and designed the experiment accordingly. Verification justified trust.
comparison chart
Score/concept
What measures
Reliable threshold
pLDDT
Local building trust (0-100)
>90 very good, <50 irregular
PAE
Cross-domain location trust
Low (dark) good
RMSD
Agreement with experiment (angström)
<2 Å is generally good
Common mistakes
- Considering low pLDDT as a "fault": It is generally a disordered/flexible region, do not delete it.
- Ensuring single mutation effect with AlphaFold: Not the right tool for the job.
- Ignoring confidence scores: Assuming the entire model is equally reliable.
- Skipping the experimental structure: If there is the real structure in the PDB, be sure to compare it.
- Interpreting complex/multiple chains as a single chain: Interactions require special modes (AlphaFold-Multimer).
Caution: AlphaFold is a remarkable tool, but it is a guess, not empirical reality. Particularly high-consequence decisions such as new drug target, binding site or mutation effect should not be taken without supporting the prediction with experimental evidence (structure, activity test).
From structure to function: is seeing the pocket enough?
Obtaining the 3D structure of a protein is exciting, but structure alone does not tell you function. You can see the active site (the pocket where the substrate binds) of an enzyme; However, the structure does not indicate which molecule it binds, how fast it works, or where it is located in the cell. AlphaFold is a starting point: it generates a hypothesis (“this pocket might be binding an ATP”), the experiment tests it. AI helps you formulate these hypotheses and write code that analyzes the structure, but a function claim requires empirical evidence.
Another powerful use is structural comparison: even if the sequences of two proteins are very different, their structures can be similar; this is a clue to distant evolutionary relatedness or shared function. Tools like Foldseek quickly look for structure similarity. The model helps you interpret the output of these tools and write the comparison code.
Align the AlphaFold structure of two proteins with Bio.PDB, calculate RMSD, and report which regions overlap and which diverge. Comment that structural similarity is a clue to functional relatedness, but not evidence.
Complexes, interactions and boundaries
Proteins do not work alone, but often with partners (other proteins, DNA, small molecules). AlphaFold-Multimer can predict protein-protein complexes; but it alone is not sufficient for the power and reality of interactions. When the model predicts that “two proteins interact,” this is a preliminary hypothesis; should be confirmed by experiments such as co-immunoprecipitation (co-IP). Use AI to understand where these tools are appropriate and assess output confidence; Confidence scores (especially interface PAE) are more important than ever in complex predictions.
AlphaFold isn't the only option
While AlphaFold is a pioneer, it is not the only tool. ESMFold (Meta) performs rapid prediction from a single sequence, without requiring evolutionary alignment; It is suitable for scanning large sets of sequences, but is generally slightly less accurate than AlphaFold. RoseTTAFold is another powerful alternative. AlphaFold3 and similar newer versions also include protein-ligand and protein-nucleic acid complexes. You can consult the AI which tool is suitable for a construction problem (speed or accuracy, single chain or complex); But remember to read each tool's trust metric on its own scale: what is considered a "good" score in one tool is interpreted differently in another.
In summary
AlphaFold revolutionized biology by predicting protein structure from its sequence. However, its output should be read together with confidence scores (pLDDT, PAE); low confidence zones should be interpreted as "uncertain/flexible" rather than "incorrect". Artificial intelligence (LLM) helps you explain these scores, write visualization code, and compare them to the experimental structure. For high-consequence decisions, prediction must be supported by empirical evidence.
Application task
Choose a protein (e.g. an enzyme you are interested in). Find its array from UniProt and its structure from AlphaFold DB. Have the AI write and run code with py3Dmol that colors the structure according to pLDDT. Identify high and low safe zones. Have the model interpret the possible biological significance of low-confidence regions and check for experimental structures in the PDB.
checklist
- [ ] I evaluated the structure together with the pLDDT/PAE scores.
- [ ] I interpreted low confidence zones as "uncertain/flexible", not "error".
- [ ] I didn't have much confidence in AlphaFold for the single mutation effect.
- [ ] I compared it with the experimental structure (PDB), if available.
- [ ] I thought about the right tool for the polychain/complex.
- [ ] I supported the high-consequence decision with empirical evidence.