eduKateSG · Why Science?
A protein model is an invitation to ask better questions
Read AlphaFold DB predictions with sequence identity, pLDDT confidence, PAE, domains and experimental evidence kept clearly separate.
Protein structure helps explain how molecules bind, move and carry out biological roles. AlphaFold DB makes predicted structures widely accessible, linking models to protein sequences and displaying confidence information that belongs beside the coordinates.
The database is a partnership between Google DeepMind and EMBL-EBI. Its records are predictions, not experimental structures. The official FAQ explains pLDDT confidence bands and predicted aligned error, while the download page warns that predictions vary in confidence and are not validated for clinical use.
Did you know? A single predicted model can contain very confident folded regions and very uncertain linkers or tails. Reading one global picture without local confidence can hide that difference. Colour and error plots are part of the scientific evidence, not optional decoration.
This article uses AlphaFold DB to teach a transferable habit: separate what the sequence establishes, what the model predicts, what confidence estimates indicate and what experiments would still be needed.
Chapter index
Chapters 1–9: sequence, model and confidence
- 1. Sequence identity comes first
- 2. A model is a prediction
- 3. pLDDT reports local confidence
- 4. Colour bands carry meaning
- 5. Low confidence may signal flexibility
- 6. PAE compares relative placement
- 7. Domains can be strong while orientation is weak
- 8. Coordinates need confidence fields
- 9. Predicted models complement PDBe
Chapters 10–17: PAE, comparison and experimental follow-up
- 10. Ligands are not automatically modelled
- 11. Multimers need separate evidence
- 12. Isoforms can change the model
- 13. Function is not read directly from shape
- 14. Confidence is not biological importance
- 15. Predictions are not clinical advice
- 16. Downloads enable reproducible analysis
- 17. Structure literacy opens many pathways
Protein shapes are easier to explore when prediction and experiment stay clearly labelled. AlphaFold DB records are connected to protein sequences, commonly through UniProt accessions. The sequence defines which amino-acid chain the model attempts to represent. That habit supports biology learning because structure can suggest questions without proving function or clinical meaning.
Try this
Evidence question: Which accession, organism and sequence version anchor this model? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Record the UniProt accession and length before rotating the structure viewer. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Keep the identifier, retrieval date and source page in the note so the route can be repeated.
Evidence checkpoint
Record the evidence object for sequence identity comes first, its identifier or version context, and one limit that must remain visible.
Confidence colours turn a beautiful model into a question about where the prediction is dependable. AlphaFold predicts three-dimensional coordinates from amino-acid sequence. Those coordinates offer a testable structural hypothesis rather than direct observation by crystallography, NMR or cryo-EM. It also connects classroom ideas about forces, folding and shape to modern computational science.
Try this
Evidence question: Which statement describes predicted geometry, and which would require experimental evidence? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Rewrite “the protein has this structure” as a confidence-aware prediction statement. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. If a field is missing, write “not shown” and follow the official documentation rather than inventing a value.
Evidence checkpoint
Record the evidence object for a model is a prediction, its identifier or version context, and one limit that must remain visible.
A predicted coordinate file is a powerful hypothesis object, provided its limits travel with it. Per-residue predicted Local Distance Difference Test scores estimate confidence around each residue. The FAQ groups scores into very high, confident, low and very low bands. The confidence field is part of the result, not decoration around the model.
Try this
Evidence question: Which regions have high or low pLDDT, and are they contiguous? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Map three sequence intervals to confidence bands and describe them without calling low confidence an error. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Before comparing two records, check that their object types, methods, versions and units really align.
Evidence checkpoint
Record the evidence object for plddt reports local confidence, its identifier or version context, and one limit that must remain visible.
This chapter slows the viewer down long enough to connect sequence, model and evidence. The viewer colours residues by pLDDT, allowing a reader to locate confident cores and uncertain segments. Colours summarise thresholds and should be read with the numerical scale. A careful reader can be curious and cautious at the same time.
Try this
Evidence question: What numerical range corresponds to each colour in the current legend? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Copy the legend into a note and use it to describe one domain and one tail. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Finish by naming one supported statement and one stronger statement that still needs another evidence layer.
Evidence checkpoint
Record the evidence object for colour bands carry meaning, its identifier or version context, and one limit that must remain visible.
Protein shapes are easier to explore when prediction and experiment stay clearly labelled. Very low pLDDT regions may lack a stable structure in isolation or may be disordered, but low confidence can also reflect missing context or modelling difficulty. That habit supports biology learning because structure can suggest questions without proving function or clinical meaning.
Try this
Evidence question: Which alternative explanations fit this low-confidence region? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: List three possibilities and one experiment or database comparison that could distinguish them. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Keep the identifier, retrieval date and source page in the note so the route can be repeated.
Evidence checkpoint
Record the evidence object for low confidence may signal flexibility, its identifier or version context, and one limit that must remain visible.
Confidence colours turn a beautiful model into a question about where the prediction is dependable. Predicted aligned error estimates confidence in the relative positions of residue pairs. It is especially useful for asking whether domains are confidently oriented with respect to one another. It also connects classroom ideas about forces, folding and shape to modern computational science.
Try this
Evidence question: Are two domains individually confident but mutually uncertain in placement? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Read a PAE plot and mark blocks of low predicted error without treating the axes as genomic coordinates. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. If a field is missing, write “not shown” and follow the official documentation rather than inventing a value.
Evidence checkpoint
Record the evidence object for pae compares relative placement, its identifier or version context, and one limit that must remain visible.
CHAPTER 7 OF 17 · Read local confidence
7. Domains can be strong while orientation is weak
A predicted coordinate file is a powerful hypothesis object, provided its limits travel with it. A model may predict several well-folded domains with low confidence in the hinge or inter-domain arrangement. That difference matters for mechanistic stories. The confidence field is part of the result, not decoration around the model.
Try this
Evidence question: Which conclusion concerns domain folds and which concerns assembly geometry? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Write separate sentences for within-domain confidence and between-domain orientation. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Before comparing two records, check that their object types, methods, versions and units really align.
Evidence checkpoint
Record the evidence object for domains can be strong while orientation is weak, its identifier or version context, and one limit that must remain visible.
This chapter slows the viewer down long enough to connect sequence, model and evidence. Downloaded structure files can include coordinates and confidence values. Moving the file into another viewer does not remove the need to preserve model provenance and confidence. A careful reader can be curious and cautious at the same time.
Try this
Evidence question: Which file field stores confidence and which accession identifies the model? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Open or inspect a model file header and note provenance before making an image. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Finish by naming one supported statement and one stronger statement that still needs another evidence layer.
Evidence checkpoint
Record the evidence object for coordinates need confidence fields, its identifier or version context, and one limit that must remain visible.
Protein shapes are easier to explore when prediction and experiment stay clearly labelled. Experimental structures in the Protein Data Bank and predicted models answer different questions. Agreement can be informative, while differences require attention to construct, state, ligands and experimental conditions. That habit supports biology learning because structure can suggest questions without proving function or clinical meaning.
Try this
Evidence question: Is the comparison using the same sequence region and biological state? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Align a predicted region conceptually with an experimental entry and list the contextual checks first. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Keep the identifier, retrieval date and source page in the note so the route can be repeated.
Evidence checkpoint
Record the evidence object for predicted models complement pdbe, its identifier or version context, and one limit that must remain visible.
CHAPTER 10 OF 17 · Read domain relationships
10. Ligands are not automatically modelled
Confidence colours turn a beautiful model into a question about where the prediction is dependable. A protein may function with cofactors, ligands, membranes or partners that are absent from a monomeric prediction. A pocket-like shape alone does not prove binding. It also connects classroom ideas about forces, folding and shape to modern computational science.
Try this
Evidence question: Which biological components are missing from the model context? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Annotate a model screenshot with “predicted protein only” and a list of candidate contextual partners. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. If a field is missing, write “not shown” and follow the official documentation rather than inventing a value.
Evidence checkpoint
Record the evidence object for ligands are not automatically modelled, its identifier or version context, and one limit that must remain visible.
A predicted coordinate file is a powerful hypothesis object, provided its limits travel with it. A monomer prediction does not establish oligomeric state. Interfaces and assemblies require experimental evidence or appropriately evaluated multimer predictions. The confidence field is part of the result, not decoration around the model.
Try this
Evidence question: What evidence supports the number and arrangement of subunits? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Separate monomer fold confidence from assembly evidence in a two-column note. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Before comparing two records, check that their object types, methods, versions and units really align.
Evidence checkpoint
Record the evidence object for multimers need separate evidence, its identifier or version context, and one limit that must remain visible.
This chapter slows the viewer down long enough to connect sequence, model and evidence. Different protein isoforms or sequence variants can add, remove or alter regions. An accession mismatch can therefore invalidate a structural comparison. A careful reader can be curious and cautious at the same time.
Try this
Evidence question: Does the model sequence match the isoform used in the experiment? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Compare sequence lengths and variant positions before transferring an annotation. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Finish by naming one supported statement and one stronger statement that still needs another evidence layer.
Evidence checkpoint
Record the evidence object for isoforms can change the model, its identifier or version context, and one limit that must remain visible.
CHAPTER 13 OF 17 · Move from prediction to evidence
13. Function is not read directly from shape
Protein shapes are easier to explore when prediction and experiment stay clearly labelled. Structural similarity and conserved motifs can suggest function, but functional assignment also needs sequence, biochemical, cellular and evolutionary evidence. That habit supports biology learning because structure can suggest questions without proving function or clinical meaning.
Try this
Evidence question: Which feature suggests a hypothesis, and what result would test it? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Turn one shape-based guess into a testable functional question. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Keep the identifier, retrieval date and source page in the note so the route can be repeated.
Evidence checkpoint
Record the evidence object for function is not read directly from shape, its identifier or version context, and one limit that must remain visible.
CHAPTER 14 OF 17 · Move from prediction to evidence
14. Confidence is not biological importance
Confidence colours turn a beautiful model into a question about where the prediction is dependable. A high pLDDT score estimates structural confidence; it does not mean the region is more essential, active or disease-relevant. It also connects classroom ideas about forces, folding and shape to modern computational science.
Try this
Evidence question: What does the score measure, and what tempting interpretation must be avoided? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Write one correct and one incorrect sentence about a high-confidence residue. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. If a field is missing, write “not shown” and follow the official documentation rather than inventing a value.
Evidence checkpoint
Record the evidence object for confidence is not biological importance, its identifier or version context, and one limit that must remain visible.
CHAPTER 15 OF 17 · Move from prediction to evidence
15. Predictions are not clinical advice
A predicted coordinate file is a powerful hypothesis object, provided its limits travel with it. AlphaFold DB states that its predictions are for theoretical modelling and are not validated or approved for clinical use. Variant or treatment decisions require validated clinical evidence. The confidence field is part of the result, not decoration around the model.
Try this
Evidence question: Would this use cross from research exploration into diagnosis or treatment? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Rewrite a clinical-sounding claim as a research hypothesis and name the validation needed. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Before comparing two records, check that their object types, methods, versions and units really align.
Evidence checkpoint
Record the evidence object for predictions are not clinical advice, its identifier or version context, and one limit that must remain visible.
CHAPTER 16 OF 17 · Move from prediction to evidence
16. Downloads enable reproducible analysis
This chapter slows the viewer down long enough to connect sequence, model and evidence. The database provides model files and bulk downloads. Reproducibility requires accession, release, file, sequence and analysis settings. A careful reader can be curious and cautious at the same time.
Try this
Evidence question: Which release and input model supported the figure or measurement? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Create a structure-analysis receipt before measuring a distance. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Finish by naming one supported statement and one stronger statement that still needs another evidence layer.
Evidence checkpoint
Record the evidence object for downloads enable reproducible analysis, its identifier or version context, and one limit that must remain visible.
CHAPTER 17 OF 17 · Move from prediction to evidence
17. Structure literacy opens many pathways
Protein shapes are easier to explore when prediction and experiment stay clearly labelled. Structural biology connects physics, chemistry, biology, AI, software and medicine. Careers range from experimental structure determination to model evaluation and drug-discovery research. That habit supports biology learning because structure can suggest questions without proving function or clinical meaning.
Try this
Evidence question: Which role produces, validates, annotates or applies the model? Answer in three parts: identify the visible field or file, explain how it bears on the question, and state the boundary of that evidence. Preserve the database’s own object names and accession rather than flattening everything into the word ‘data’.
Try this: Choose one role and design a confidence-aware handover for the next specialist. Turn the result into a compact claim-evidence-reasoning note, then ask a partner to retrace it. Keep the identifier, retrieval date and source page in the note so the route can be repeated.
Evidence checkpoint
Record the evidence object for structure literacy opens many pathways, its identifier or version context, and one limit that must remain visible.
Contents · Previous chapter · Continue at the Science Learning Hub
Confidence makes the model more useful, not less exciting
AlphaFold DB gives students immediate access to rich structural hypotheses. The most powerful reading habit is to keep sequence, coordinates, pLDDT, PAE and experimental comparison on separate lines.
A predicted structure can inspire questions about domains, motifs and experiments. Its confidence information tells you where to zoom in, where to pause and which claims still need evidence.
| Reading object | What it can support | Question to keep visible |
|---|---|---|
| UniProt accession | Identity of the modelled sequence | Does it match the experimental isoform? |
| pLDDT | Local per-residue confidence | Which regions are low confidence? |
| PAE | Confidence in relative residue placement | Are domain orientations dependable? |
| Experimental structure | Observed coordinates under stated conditions | Are construct and state comparable? |
Sources and next reading
Substantive resource claims in this article were checked against the primary pages below. A database record is a structured route into evidence; it does not by itself prove a new biological conclusion. Recheck the live record, evidence labels and release context before using it in assessed work or research.
Primary sources
- AlphaFold Protein Structure Database home — accessed 24 October 2026.
- AlphaFold DB FAQ — accessed 24 October 2026.
- AlphaFold DB downloads and disclaimer — accessed 24 October 2026.
- AlphaFold DB API — accessed 24 October 2026.
- EMBL-EBI AlphaFold training — accessed 24 October 2026.