Open Bench · Unit 5 of 8
Seeing in three dimensionsHow does a molecule know where to go?
Students predict a protein's three-dimensional shape with the same free tool that won its creators a Nobel Prize, find the pocket a drug would bind, and see why shape is the whole game.
Length
60 minutes
Tracks
Explorer (Grades 6–8) · Builder (Grades 9–12)
Cost
Free, permanently
Licence
CC BY 4.0
Objectives
By the end of the hour, a student can do each of these without being prompted.
- 01
Explain why a protein's shape, not its ingredient list, determines what it does.
- 02
Use a free structure prediction service to generate a protein model.
- 03
Rotate and inspect a structure in a browser and locate a plausible binding pocket.
- 04
Say what a predicted structure is good for and what it cannot tell you.
Run of show
The hour, minute by minute.
Timings are a plan, not a rule. The segment most worth protecting when you run late is noted in the teacher notes below.
- 10 min
Lock, key, and the lie in that metaphor
Teach lock and key, then immediately complicate it. Proteins move, flex, and change shape on binding. The metaphor is a starting point students will have to outgrow by the end of the hour.
- 15 min
The problem that was impossible
Fifty years of a field unable to predict shape from sequence, then a deep learning model that largely solved it. Students should feel how recent this is. Most of their teachers trained before it existed.
- 20 min
Predict something yourself
Submit a sequence to a free prediction server. While it runs, open an already-solved structure and practise rotating, zooming, and finding pockets.
- 10 min
Find the pocket
On their own structure, students find the largest cavity and argue about whether a small molecule could sit in it.
- 5 min
Confidence scores
Show the per-residue confidence colouring. A model can be confident and wrong, and the floppy low-confidence regions are often the biologically interesting ones.
The activity
Predict, inspect, compare
- 1
Pick a protein from your Unit 3 target list.
- 2
Find its predicted structure in the AlphaFold database and note the confidence colouring.
- 3
Search the Protein Data Bank for an experimentally solved structure of the same protein.
- 4
Open both in a browser viewer and look for the places they disagree.
- 5
Write a paragraph on where the prediction is trustworthy and where you would not bet on it.
What you hand in
Two screenshots and a paragraph identifying one region where prediction and experiment differ, with a guess at why.
Tools, all free
Free. Hundreds of millions of predicted structures, already computed, with confidence colouring built in.
Free archive of experimentally determined structures. Use it to compare a prediction against reality.
Free browser viewer. No install, works on a Chromebook.
Explorer track swap
Explorers skip prediction and work entirely with already-solved structures, using the viewer to hunt for pockets in a protein the teacher chooses. Same spatial intuition, no submission queue.
The misconception to expect
“A predicted structure is as good as an experimental one.”
A prediction is a very good hypothesis about shape. It is not a measurement, it carries a confidence score for a reason, and the regions it gets wrong are frequently the ones that matter most.
Teaching it
Notes from the room.
How you know it landed
A student has it when they can point at a low-confidence region on their own model and explain why that region might still be the most interesting part of the protein.
- 01
Prediction servers queue. Submit a job at the start of the hour and have a solved structure ready as a backup, every single time.
- 02
Spatial reasoning varies enormously at this age. Pair students who rotate structures easily with those who do not, and say nothing about why you paired them.
Reading for this unit
The papers behind the hour.
Optional for Explorers, expected for Builders. Full notes on each one are in the reading room.
- R-10
Accurate structure prediction of biomolecular interactions with AlphaFold 3
Nature · 2024 · Free full text
A single diffusion-based model predicts the joint structure of proteins together with nucleic acids, small molecules and ions, beating specialized docking tools at protein-ligand prediction and earlier methods at antibody-antigen prediction.
- R-11
Boltz-1: Democratizing Biomolecular Interaction Modeling
bioRxiv preprint · 2024 · Free full text
An open-source model reaching AlphaFold 3-level accuracy on biomolecular complexes, released with training code, inference code, weights and benchmarks under the MIT license.
A diffusion model generates new protein backbones to order, including binders designed against a chosen target. Hundreds of designs were made and experimentally characterized, and a cryo-EM structure of one binder matched its design model almost exactly.