3. AlphaFold2 and OpenFold

AlphaFold2 predicts a protein’s three-dimensional structure from its amino-acid sequence. In this lesson, the goal is not simply to produce a PDB file: it is to decide which parts of a prediction are credible and explain the evidence behind that decision.

NoteLesson contract
Time 60–90 minutes core; 30–45 minutes optional experiments
Prerequisites Tuesday lessons 1–2; FASTA familiarity; working ColabFold notebook or LocalColabFold installation
Watch / Read / Do Watch the workshop recording or read this page; complete the prediction and interpretation tasks
Outcomes Explain why MSAs help; interpret pLDDT, pTM, and PAE; compare a prediction with an experimental structure
Artifact GFP prediction folder plus a completed evidence table and short trust statement
Complete when You can identify confident and uncertain regions, report the comparison to 1GFL, and justify whether the model is useful for experimental planning
Next Continue to ESMFold vs. AlphaFold2 after saving your artifact
TipChoose your route
  • Watch + do: Watch the workshop recording, pausing at the prompts below, then complete the core lab.
  • Read + do: Read the short explanations on this page and complete the core lab.
  • Full bootcamp: Complete the core lab, parameter experiments, and the optional architecture deep dive.

Watch: Workshop Session

The recording is an alternative explanation of the core ideas, not an additional requirement. Pause before each confidence plot is explained and make your own interpretation first.

🎥 Live workshop recording — AlphaFold2 Structure Prediction
đź“Š View slide deck

Loop 1: Inspect Before You Predict

Explain

AlphaFold2’s result is a claim about structure accompanied by estimates of uncertainty. Its main outputs are:

  1. 3D coordinates: predicted atomic positions, usually saved as a PDB or mmCIF file.
  2. pLDDT: confidence in the local environment around each residue, from 0 to 100.
  3. pTM: confidence in the overall fold and relative placement of domains, from 0 to 1.
  4. PAE: predicted error in the relative position of every pair of residues.

Predict

Before reading the check below, write down which metric you would use for each question:

  • Is this loop locally well modeled?
  • Are two domains positioned reliably relative to one another?
  • Is the overall fold likely to be correct?

Do

Open one AlphaFold Database entry for a protein you know at alphafold.ebi.ac.uk. Inspect both the structure coloring and the PAE plot. Record one confident region and one uncertainty you would investigate.

Check

  • Local region: use pLDDT.
  • Relative domain placement: use PAE.
  • Overall fold: use pTM, supported by pLDDT and PAE rather than interpreted alone.
ImportantConfidence is not correctness

A confident prediction can still miss a ligand-induced state, alternative conformation, oligomeric assembly, or biological context. A low-confidence region may be intrinsically disordered or flexible rather than simply “wrong.” Treat the metrics as evidence, not a verdict.

Loop 2: Read Confidence as Evidence

Explain: pLDDT

pLDDT Interpretation Sensible response
>90 Very high local confidence Usually suitable for residue-level interpretation
70–90 Confident Generally reliable local fold
50–70 Low confidence Treat local geometry cautiously
<50 Very low confidence Consider disorder, flexibility, missing context, or failure

Low pLDDT often occurs in flexible loops and intrinsically disordered regions. It is information, not automatically a failed prediction.

Explain: pTM, ipTM, and PAE

  • pTM estimates confidence in the global topology. Values above 0.8 are often strong; 0.5–0.8 requires scrutiny; below 0.5 is weak.
  • ipTM estimates confidence in a multimer interface and should be considered with interface PAE and biological evidence.
  • PAE answers a pairwise question: if the model is aligned on residue i, how uncertain is the position of residue j? Low values mean confident relative placement.

On a PAE heatmap, low-error diagonal blocks often correspond to well-resolved domains. High error between blocks means the domains may each be credible while their relative orientation is not.

Predict

A two-domain protein has high pLDDT throughout but high PAE between the two domain blocks. Which conclusion is best?

  1. Both domains are certainly wrong.
  2. Each domain may be locally credible, but their relative orientation is uncertain.
  3. High pLDDT guarantees the complete model is correct.

Do

Sketch a two-block PAE matrix and label the within-domain and between-domain regions. This takes one minute and makes later plots much easier to read.

Check

The best conclusion is 2. pLDDT evaluates local environments; PAE exposes uncertainty in relative placement.

Loop 3: Explain Where the Signal Comes From

Explain

AlphaFold2 uses three main inputs:

  1. The query sequence to predict.
  2. A multiple sequence alignment (MSA) of related sequences.
  3. Optional structural templates from known homologs.

An MSA reveals co-evolution. Residues that contact one another in 3D can undergo compensating mutations: when one side of an interaction changes, a change at the partner position may preserve the interaction. Across many diverse homologs, these correlated patterns constrain possible folds.

flowchart LR
    Seq[Query sequence] --> Search[Database search]
    Search --> MSA[Multiple sequence alignment]
    MSA --> Coev[Co-evolutionary signal]
    Coev --> Model[AlphaFold2]
    Seq --> Model
    Templates[Optional templates] --> Model
    Model --> Structure[Coordinates + confidence]

MSA depth helps, but diversity matters more than a raw sequence count. The effective number of sequences, Neff, discounts redundant near-copies.

Neff Typical expectation
>1000 High-confidence predictions are likely
100–1000 Good predictions for many proteins
30–100 Some regions may be unreliable
<30 Significant uncertainty; compare alternative methods

Proteins with few natural homologs—including many designed proteins—therefore present a special challenge.

Predict

If 1,000 MSA rows are almost identical, will they provide the same information as 1,000 diverse homologs? State why in one sentence.

Do

When you run GFP below, open the MSA coverage plot. Note whether coverage and depth remain consistent across the full sequence or drop in particular regions.

Check

Near-identical rows provide less independent evolutionary evidence, so their effective depth is lower than that of a diverse alignment.

Check Your Understanding
Why is the MSA so important for AlphaFold2's accuracy?
It provides template structures for the model to copy.
It reveals co-evolutionary patterns that imply 3D contacts.
It only increases the size of the training data.
NoteAlphaFold2, OpenFold, and ColabFold
  • AlphaFold2 is DeepMind’s original JAX implementation with released inference code and weights.
  • OpenFold is a trainable PyTorch reproduction designed for community research and extension.
  • ColabFold combines AlphaFold2 models with fast MMseqs2-based MSA generation, making inference more accessible.

For the neural-network mechanics, continue to the optional architecture deep dive. It covers Evoformer representations, outer product mean, triangle updates, invariant point attention, recycling, settings, and extensions.

Loop 4: Generate and Evaluate a GFP Prediction

Explain

Green Fluorescent Protein (GFP) is a useful practice target because it is a compact, well-characterized protein with an experimental reference structure. Your job is to produce a prediction and then build an evidence-based trust statement.

Predict

Before running the model, write two expectations:

  • Will GFP appear as one coherent domain or several independently moving domains in PAE?
  • Which parts of the sequence, if any, do you expect to have lower pLDDT?

Keep these predictions. Comparing expectations with results is part of the exercise.

Do: Run ColabFold

Download 1GFL.fasta, or save the following as gfp.fasta:

>GFP
MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTFSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK

Choose one route:

LocalColabFold or HPC

colabfold_batch gfp.fasta gfp_output/

Hosted notebook

Open the ColabFold notebook, paste the sequence, and run all cells.

Expected runtime is approximately 5–15 minutes, depending on available hardware and queue time. Key outputs normally include:

  • a ranked PDB structure;
  • a JSON file of confidence scores;
  • an MSA coverage plot;
  • a PAE heatmap.
TipIf compute is unavailable

Use an instructor-provided or previously generated GFP output folder and complete the analysis. Interpreting evidence is the learning objective; a temporary GPU or queue failure should not stop the lesson.

Do: Analyze Local Confidence

Load the top-ranked structure in PyMOL. Adjust the filename to match your output:

load gfp_output/gfp_relaxed_rank_001_alphafold2_ptm_model_1_seed_000.pdb, af2_gfp
spectrum b, blue_white_red, minimum=50, maximum=100

Then answer:

  • Which regions have high or low pLDDT?
  • Does a low-confidence segment look like a terminal tail, loop, or larger structural region?
  • Does the confidence pattern match your prediction?

Do: Compare with Experiment

fetch 1GFL
align af2_gfp, 1GFL

Record the reported RMSD and inspect the overlay. Do not use RMSD alone: note whether deviations are global or concentrated in a few flexible regions.

Do: Interpret PAE

Open the PAE image and answer:

  • Is the plot dominated by one low-error block or several blocks?
  • Are there high-PAE regions?
  • What does the pattern imply about GFP’s domain organization and relative-coordinate confidence?

Check: Build the Artifact

Fill in the evidence table and save it with your prediction outputs.

Evidence Your GFP result Interpretation
Average pLDDT
pTM score
RMSD to 1GFL
Prediction time
MSA depth / coverage
Regions with pLDDT <70
PAE pattern

Finish with a three-sentence trust statement:

  1. Which parts of the model would you trust, and for what type of decision?
  2. Which uncertainty most limits its use?
  3. What additional evidence would you seek before a high-stakes experiment?
ImportantDefinition of completion

This lesson is complete when your folder contains the prediction (or provided outputs), evidence table, and trust statement. A rendered structure alone is not sufficient.

Optional: Test One Assumption

Choose one experiment; running all three is enrichment rather than core work.

Fewer recycles

colabfold_batch --num-recycle 1 gfp.fasta gfp_1recycle/

Single-sequence mode

colabfold_batch --msa-mode single_sequence gfp.fasta gfp_single_seq/

All five model parameter sets

colabfold_batch --num-models 5 gfp.fasta gfp_all_models/

Before each run, predict which metric will change most. Afterward, compare confidence, structural agreement, and runtime. Record whether the evidence supports your prediction.

Optional Multimer Prediction

In a FASTA entry, separate chains with a colon:

>homodimer
SEQUENCEOFCHAINA:SEQUENCEOFCHAINB

Run colabfold_batch homodimer.fasta homodimer_output/, then interpret ipTM and the off-diagonal regions of the PAE plot. A plausible-looking interface without supporting ipTM, PAE, or biological evidence should not be treated as validated.

Decide and Continue

Before moving on, be able to answer these questions without reopening the page:

  1. Why does evolutionary diversity matter more than a raw MSA row count?
  2. What different uncertainty does PAE reveal compared with pLDDT?
  3. When could a low-confidence region be biologically meaningful?
  4. What evidence would make you distrust an otherwise attractive structure?

For implementation detail, read the AlphaFold2 architecture deep dive. Otherwise, save your artifact and continue to ESMFold vs. AlphaFold2.