Lab 1 · Structure · about 20 minutes
Protein Structure Explorer
AI tools like AlphaFold can now predict a protein's 3D shape from its sequence. But a prediction isn't a measurement. In this lab you'll put a structure measured in the lab next to the AI's prediction and learn to read the AI's own confidence score.
- Compare an experimental structure (Protein Data Bank) with an AI-predicted one (AlphaFold).
- Read pLDDT, AlphaFold's per-residue confidence score.
- Explain why low confidence often points to biology (flexible or disordered regions), not just AI error.
Key ideas · a 2-minute primer before you start
- Protein structure
- A protein is a chain of amino acids (called residues) that folds into a 3D shape. The shape decides what the protein does, including where a drug might bind.
- How structures are measured
- Scientists measure structures with X-ray crystallography, cryo-electron microscopy (cryo-EM), or NMR. The results go into the Protein Data Bank, which holds more than 200,000 structures.
- Resolution
- How sharp a measured structure is, in ångströms (Å). Smaller is sharper: at about 2 Å most atom positions are clear, while at 4 Å you mainly see the overall fold.
- AlphaFold
- An AI system that predicts a protein's 3D shape from its amino-acid sequence alone. Its database holds over 200 million predictions. A prediction is a strong hypothesis, not a measurement. Database models are single chains, without drugs, DNA, or partner proteins.
- pLDDT
- AlphaFold's confidence for each residue, from 0 to 100. Above 90 is very high. 70–90 usually means the overall fold is right. 50–70 is low. Below 50 often means the region has no fixed shape.
- Disordered regions
- Some parts of proteins are naturally floppy and never hold one shape. Low pLDDT there is often the AI correctly signaling flexibility, not simply failing.
Go deeper in the free course lesson Data Analysis and Discovery Models.
Loading…
Loading…
Very high (>90)
Confident (70–90)
Low (50–70)
Very low (<50)
Drag to rotate, scroll or pinch to zoom. Colors on the right show pLDDT, AlphaFold's confidence for each residue.
Your tasks
Start with lysozyme. Compare the two shapes. How much of the prediction is dark blue? Then look at the counts: the AI model has 147 residues but the lab structure has 129. Find the orange tail. (Hint: the first 18 residues are a signal sequence that's cut off before the enzyme does its job.)
Switch to p53. The lab structure covers only part of the protein. Look at the AI's prediction: where is it confident, and where is it orange?
Switch to insulin. The AI predicts the whole precursor protein (110 residues), but the lab structure has two short chains. Which parts are low confidence, and why might that be? (Hint: parts of the precursor are cut away in the mature hormone.)
For each protein, note the mean pLDDT and the share of residues below 50. Which protein would you trust the AI on most for planning an experiment?
Reflect and discuss
- AlphaFold's confidence is about its own prediction, not about whether the protein binds your drug. What would you still need to test in the lab?
- Low pLDDT often marks regions that really are floppy in the cell. How would you explain that to a colleague who says "the AI got it wrong"?
- Where in your own work could a predicted structure save time, and where would it be risky to rely on one?
Go deeper: the lesson Data Analysis and Discovery Models in the free course AI for Life-Science Professionals covers how to read the confidence measures of discovery models like AlphaFold.
CityUx prototype · Content CC BY 4.0, code MIT. Not yet reviewed by a life-science domain expert.
