Wednesday: Advanced Prediction and Design
Historical day 3 · compare complex predictors, then generate and select constrained backbones
Overview
Wednesday preserves the transition in the original workshop from predicting a structure to designing one. First, you compare next-generation complex predictors on monomers, multimers, ions, and ligands. Then you use RFdiffusion to sample backbones under length, motif, interface, partial-diffusion, fold, or symmetry constraints.
The self-paced emphasis is comparison and selection. A folder containing generated PDBs is not yet a result. Your evidence should show which constraints you used, how outputs differed, what failed, and why a subset deserves sequence design and validation.
- Compare AlphaFold3 and Chai-1 outputs using structure, confidence, entity placement, and provenance rather than a single score.
- Translate a design intent into an RFdiffusion contig, chain break, hotspot list, output prefix, and number of designs.
- Distinguish unconditional generation, motif scaffolding, partial diffusion, interface design, fold conditioning, and symmetry constraints.
- Inspect topology, diversity, motif retention, target contact, hotspots, clashes, and intended symmetry.
- Select or reject generated backbones with an explicit evidence-based rationale.
Wednesday contract
| Core path | Full bootcamp | |
|---|---|---|
| Active effort | 3–5 hours | 6–10 hours |
| Compute time | Add 1–4 hours of hosted/local jobs plus queues | Add 4–12+ hours for all RFdiffusion modes and larger galleries |
| Compute/accounts | AlphaFold Server Google sign-in, Chai-Lab account, and/or a working RFdiffusion GPU environment | Same, with enough allocation/storage for several seeds and parameter sets |
| Output | One predictor-comparison table and one backbone gallery with a selection memo | Core evidence plus all original activity modes and parameter/potential experiments |
| Finished when | You can reproduce a run and explain why at least one candidate advanced or failed | You can also compare how different constraints changed the sampled distribution |
Hosted access and queues can change. If an account or GPU is unavailable, analyze a matched set of saved outputs and record their source.
Choose a Wednesday route
Core path
- In AlphaFold3 vs. Chai-1, compare both predictors on one shared example. Prefer the 8AJY multimer + ions if your accounts allow it because it exposes more than monomer fold agreement; otherwise use 1QYS.
- In RFdiffusion, run one unconditional design as a smoke test, then run one hotspot-guided binder example for a chosen capstone target.
- Inspect all outputs, not only the top filename. Create a small gallery and advance, reject, or rerun each candidate using named criteria.
Full-bootcamp path
Complete all AlphaFold3/Chai examples (monomer, multimer/ions, and protein–ligand), then work through the RFdiffusion modes: unconditional, motif scaffolding, partial diffusion, hotspot binder design, fold conditioning, and cyclic/dihedral/tetrahedral symmetry. Add the independent project’s potential-optimization comparison.
Reference path
- Need a multimodal complex comparison? Use AlphaFold3 vs. Chai-1.
- Need a backbone-generation recipe? Jump to the relevant numbered mode in RFdiffusion.
- Already have backbones? Continue with your sequence designer and return to the capstone validation stages.
Watch → Read → Do → Check
| Module | Active time | Watch | Read | Do | Check / artifact |
|---|---|---|---|---|---|
| 1. AlphaFold3 vs. Chai-1 | 60–120 min plus queue | Recommended for the original conceptual comparison | Account/setup instructions, shared inputs, and analysis prompts | Submit one matched example to both tools or inspect matched saved outputs | Table records tool/version/date, entities, confidence, structural agreement, entity/interface placement, limitations, and intended use |
| 2. RFdiffusion | 2–4 hr core plus generation | Optional workshop context | Contigs, chain breaks, hotspots, conditioning modes, output inspection, and troubleshooting | Generate one smoke-test set and one hotspot-guided set; keep exact scripts/configs | Contact-sheet gallery and selection memo compare topology, diversity, target/hotspot contact, clashes, and next validation step |
A minimum backbone selection rubric
Use the same criteria for every candidate so that selection does not become “this one looks nice.”
| Criterion | Advance when… | Reject or rerun when… |
|---|---|---|
| Run integrity | Input, checkpoint/version, config, seed, log, and output are present | Provenance is missing or the run contains an unresolved error |
| Constraint satisfaction | Length, chains, motif, hotspot, fold, or symmetry match the design intent | The requested motif/interface/symmetry is absent or misnumbered |
| Backbone plausibility | The chain is continuous and compact with no obvious severe clashes or impossible geometry | The backbone is extended, broken, entangled, or penetrates the target |
| Diversity | The set samples distinct poses/topologies worth validating | Near-duplicates dominate; revise seeds, ranges, noise, or potentials |
| Interface hypothesis | The binder contacts the intended surface and hotspots without relying on a single visual angle | Contact is absent, misplaced, or dominated by an obvious clash |
| Next-stage value | There is a clear reason to spend sequence-design and prediction compute | No candidate improves the evidence or tests a useful alternative |
Backbone appearance is a screen, not evidence of binding. Sequence design, complex prediction, interface confidence, physical filtering, and ultimately experiment remain downstream.
Wednesday evidence checkpoints
Next step
If you are starting the capstone now, carry the Wednesday configuration and gallery forward instead of beginning a disconnected project. Thursday supplies compute intuition, docking cautions, and a planning checkpoint before the larger campaign.