Tuesday: Structure Prediction

Historical day 2 · inspect structures, interpret confidence, and compare prediction strategies

Overview

Tuesday preserves the structure-prediction day of the original workshop. The self-paced goal is not to memorize model architecture or obtain a plausible-looking PDB. It is to turn a prediction into a defensible interpretation: what is locally confident, what is uncertain relative to what, how the output compares with experiment or another model, and what decision the evidence supports.

By the end of Tuesday, you’ll be able to:
  • Inspect chains, residues, distances, interfaces, and confidence coloring in PyMOL.
  • Explain the folding problem, Anfinsen’s hypothesis, Levinthal’s paradox, and the role of modern prediction metrics.
  • Explain at a useful level how MSA information, the Evoformer, the structure module, and recycling contribute to AlphaFold2.
  • Interpret pLDDT and PAE without confusing local confidence with a globally determined arrangement.
  • Compare AlphaFold2 and ESMFold by evidence, speed, MSA dependence, and fitness for a stated question.

Tuesday contract

Core path Full bootcamp
Active effort 5–7 hours 8–11 hours
Compute time Add 15–90 minutes per prediction plus any hosted/HPC queue Add several prediction variants and optional multimer runs
Compute Laptop + PyMOL; hosted ColabFold or local 16+ GB GPU for prediction Local/hosted GPU, multiple model settings, and saved confidence outputs
Output Structure figure, metric interpretation, AF2 confidence analysis, and comparison memo Core evidence plus architecture notes and parameter/model comparisons
Finished when You can state what you trust, what you do not trust, and what evidence you would seek next You can also explain how inputs/settings plausibly changed the observed output

Recordings are the original live explanations. Unless a page says otherwise, they are optional companions to the written core and hands-on work.

Choose a Tuesday route

Core path

Complete all four modules in order, but take the short route through AlphaFold2:

  1. Make one annotated PyMOL view.
  2. Build a metric-selection note from the structure-prediction lesson.
  3. Predict or obtain GFP outputs, interpret pLDDT and PAE, and make one trust/uncertainty decision.
  4. Compare the same target with ESMFold and AlphaFold2, then write a model-choice memo.

Full-bootcamp path

Add the live recordings and slides, the AlphaFold2 Architecture Deep Dive, model/recycle/MSA experiments, the optional multimer, and the optional ESMFold speed/memory/determinism sections.

Reference path

Watch → Read → Do → Check

Module Active time Watch Read Do Check / artifact
1. PyMOL and VS Code 60–90 min Optional original workshop recording Navigation, selections, representations, sequence/structure connection, and remote workflow Inspect one PDB; select a biologically meaningful region; measure or label it; save a figure PNG, PDB ID, and command history make the view reproducible
2. Structure Prediction 60–90 min Optional original workshop recording Folding foundations, CASP, RMSD/lDDT/GDT/TM-score, and method families Compare two structures and choose the metric(s) appropriate to a stated question One-page metric note explains why a single score is insufficient
3. AlphaFold2 and OpenFold + architecture deep dive 60–90 min core; +90–120 min deep dive Optional recording/slides Core workflow, MSA intuition, pLDDT, PAE, and practical limits; attention and geometry are optional enrichment Predict GFP locally or in ColabFold, inspect the ranked PDB and confidence outputs, and compare with 1GFL Figure combines structure/confidence evidence; note separates observation, interpretation, uncertainty, and next step
4. ESMFold vs. AlphaFold2 2–3 hr plus inference Optional original workshop recording Language-model prediction, confidence extraction, tradeoffs, and disagreement Run or inspect both models on the same sequence; align with 1GFL; compare confidence, RMSD, and runtime Model-choice memo selects a method for two different scenarios and explains disagreement
NoteIf a prediction cannot run

Use a saved ranked PDB, scores JSON, PAE image, and log from the course or a peer. Record their provenance and complete the same analysis. The learning checkpoint is the interpretation, not access to a particular GPU or hosted quota.

A reusable interpretation pattern

For every predicted structure, write five short statements:

  1. Input: sequence(s), MSA mode/source, model/settings, and comparison structure if any.
  2. Observation: pLDDT pattern, PAE blocks/off-diagonal regions, agreement across models/seeds, and structural comparison.
  3. Interpretation: which local regions and relative arrangements are supported.
  4. Uncertainty: disorder, domain motion, interface ambiguity, sparse evolutionary information, or model disagreement.
  5. Decision: what the structure is fit for and what evidence is required before a higher-stakes use.

“The structure looks right” is not an interpretation. “The core is locally confident, but high inter-domain PAE means the relative orientation should not define an interface experiment” is.

Tuesday evidence checkpoints

Next step

Proceed when you can interpret uncertainty rather than merely generate coordinates. Wednesday extends that habit to multimodal complex prediction and then switches from predicting a sequence to generating backbones under design constraints.

← Monday

Wednesday: Advanced Prediction and Design →