Worked Capstone Selection Example
An illustrative evidence trail—not a validated binder
The candidate names and numerical results on this page are fictional. They demonstrate the expected reasoning and documentation pattern; they are not experimentally or computationally validated PD-L1 binders.
The design question
Mission: design a small binder that approaches the PD-1-facing surface of PD-L1 while contacting a compact set of native-interface hotspots.
Chosen structure: PD-L1 chain A from 4ZQK.
Initial hotspots: A56, A58, A113, A122, and A123.
Why these residues: they form two nearby contact clusters on the observed PD-1 interface. Using both clusters encourages the generated binder to cover the intended face rather than attach to an unrelated exposed patch.
Run record
The project notebook records:
- Exact structure download URL and preparation commands
- Hash of the cleaned target file
- RFdiffusion release or commit and checkpoint
- Full contig and hotspot configuration
- Number of designs, seeds, GPU model, and run date
- LigandMPNN and complex-prediction versions and commands
The original files are stored under immutable run directories such as runs/2025-07-18_hotspot_v1/; summary figures are generated from those directories rather than edited manually.
First-pass backbone review
Thirty hotspot-conditioned backbones were generated. The first filter removed:
- 8 designs that did not contact both hotspot clusters
- 5 with severe target or self clashes
- 4 with extended or implausible binder topology
- 3 near-duplicates that added little structural diversity
Ten backbones advanced to sequence design. The report includes a montage of all 30, not only the successful ten, so the selection rate and common failure modes remain visible.
Candidate comparison
Three candidates survived sequence design and complex prediction.
Candidate PDL1-A
- Strong binder fold confidence
- Compact interface and good hotspot coverage
- Low predicted error within each chain
- High uncertainty in the relative chain orientation
Interpretation: attractive monomer and interface geometry, but the inter-chain uncertainty prevents a confident advance decision.
Candidate PDL1-B
- Slightly lower binder confidence than A
- Consistent interface orientation across multiple predictions
- Contacts four of five intended hotspots
- One buried unsatisfied polar group near the interface
Interpretation: the most reproducible interface, but the polar defect needs redesign or explicit energetic evaluation.
Candidate PDL1-C
- Highest single summary score
- Contacts only one hotspot cluster
- Several models rotate the binder toward a different PD-L1 surface
- Sequence contains a hydrophobic surface patch away from the interface
Interpretation: the headline score is misleading. The geometry does not support the original competitive-binding mission.
Selection memo
Advance PDL1-B conditionally. It is not the highest-scoring candidate, but it has the strongest agreement across independent complex predictions and best preserves the intended interface geometry. Before additional expensive computation, redesign the buried polar position, generate a small sequence neighborhood, and test whether the interface orientation remains stable.
Hold PDL1-A as a backup. Its fold and contacts are promising, but chain-orientation uncertainty requires more sampling.
Reject PDL1-C. The high summary score does not compensate for missing the design mission and exhibiting a potential solubility liability.
What makes this a strong submission
The example does not claim that PDL1-B binds. It does five more defensible things:
- Connects every conclusion to a saved run and identifiable input.
- Shows failures and attrition, not only finalists.
- Separates fold confidence from interface confidence.
- Chooses a candidate using the biological mission rather than one number.
- Ends with the uncertainty and next test that could change the decision.
Your turn
Use the same pattern for your target:
- State the design mission.
- Define selection criteria before looking at final scores.
- Preserve rejected designs and explain why they failed.
- Compare at least two plausible finalists.
- Make a conditional decision and name the next test.