Capstone Project
Computational binder design as a reproducible evidence trail
Capstone mission
Given a target protein and an intended binding surface, develop and defend a computational binder-design hypothesis. The deliverable is not a claim that a binder works. It is a reproducible portfolio showing how you selected a target and epitope, generated and sequence-designed candidates, validated complexes, learned from failure, and chose the next candidate or experiment.
This capstone preserves the final project from the original in-person bootcamp. The live groups presented for about 20 minutes; self-paced learners can submit an equivalent written report, narrated slide deck, notebook, repository, or video.
- Evidence guide and 20-point rubric — the seven required artifacts, completion criteria, and self-assessment standard
- Worked selection example — an illustrative PD-L1 evidence trail that advances a lower-scoring candidate for better reasons
- Download the evidence template — a ready-to-fill Markdown portfolio outline
Capstone contract
| Core capstone | Full campaign | |
|---|---|---|
| Active effort | 10–25 hours | 20–40+ hours |
| Compute time | Add 5–30 hours of generation, sequence design, validation, and queues | Add 20–100+ hours for larger sweeps, seeds, methods, and repeated validation |
| Scope | One target, one defined epitope, a teaching-scale campaign, several sequence variants, and at least two finalists | Broader hotspot/parameter/potential comparison with more candidates and independent checks |
| Output | Seven-artifact evidence portfolio and completed rubric | Core portfolio plus deeper ablations, expanded validation, and presentation-ready synthesis |
| Finished when | All seven artifacts exist; rubric is at least 13/20 with no category below 2 | The evidence trail is reconstructable and further compute is driven by a new hypothesis, not missing documentation |
Compute ranges are planning estimates, not guarantees. Target length, campaign size, GPU, queue policy, model, and failed jobs dominate wall-clock time. Start every stage with the smallest smoke test that could expose a bad path, chain, residue number, environment, or output assumption.
Two valid participation modes
Generation mode
Run the pipeline yourself and preserve exact inputs, versions, commands/configurations, seeds, logs, raw outputs, and hardware context.
Analysis mode
If GPU, account, model, license, or quota access is unavailable, use supplied or previously generated outputs. Preserve their source and reconstruct as much provenance as possible. You still complete target analysis, filtering, confidence/interface interpretation, failure analysis, and a conditional selection memo. Clearly distinguish evidence you generated from evidence you inherited.
Both modes use the same rubric. Analysis mode is not permission to hide missing provenance; it is a scientifically honest fallback when infrastructure is outside the learner’s control.
Watch → Read → Do → Check
- Watch (optional): the two original workshop recordings below provide the broad tool landscape and a practical workflow walkthrough.
- Read: this overview, your selected target page, the rubric, and the relevant tool lessons.
- Do: produce the seven staged artifacts below, preserving failed and rejected outputs as well as finalists.
- Check: score the portfolio with the rubric, compare the decision logic with the worked example, and revise the weakest category.
Original workshop recordings
Choose a target
Choose a target whose biological surface you can explain and whose structure you can prepare confidently. Do not choose solely because a name is familiar.
| Target | Starting structure | Design context | First question to resolve |
|---|---|---|---|
| PD-L1 | 4ZQK | Immune-checkpoint interface | Which PD-1-facing residues define a compact, accessible epitope? |
| IFNAR2 | 3SE3 | Cytokine-receptor surface | Which chain, state, and native interface support the intended competition hypothesis? |
| IL-7Rα | 3DI2 | Interleukin-receptor surface | How do chain numbering and the native ligand interface map to hotspots? |
| Bet v 1 | 4A88 | Allergen surface | Which exposed patch supports the intended blocking or recognition hypothesis? |
| TrkA receptor | 1WWW | NGF-receptor complex | Use the TrkA–NGF complex; 1WWC is an NT-3–TrkC structure, not an alternative TrkA target |
| TEM-1 β-lactamase | 1FQG | Enzyme surface/active-site context | Should a binder occlude a functional site or target a distinct exposed surface? |
| GM2 activator protein | 1G13 | Lipid-transport protein | Does the apo structure represent the pocket state relevant to the design mission? |
| β-Glucosidase | 2JIE | Enzyme/ligand complex | How should the trapped intermediate and active-site geometry influence surface choice? |
Each target page provides structure context, candidate hotspots, preparation guidance, and an interactive view. Verify chain IDs and residue numbers in the exact file you will use; numbering from a paper, UniProt, a biological assembly, and a deposited PDB may differ.
<img src="../images/course-dashboard/capstone-workflow.png" alt="Capstone workflow from target selection through backbone generation, sequence design, validation, and final selection.">
<p class="lesson-visual-caption">Treat the project as a sequence of evidence gates. Do not scale the next stage until the current artifact is reconstructable and interpretable.</p><img src="../images/course-dashboard/binder-target-interface.png" alt="Binder design target interface with hotspot residues on the target and a designed binder approaching the surface.">
<p class="lesson-visual-caption">A biologically motivated surface and a compact hotspot set provide a more testable design mission than an unrestricted target surface.</p>Project stages and required evidence
Stage 1 — Define the target and epitope
Inspect the biological assembly, chains, residue numbering, missing regions, ligands/cofactors, glycans, alternate states, and native partners. State the mechanism you hope a binder would test.
Exit artifact: annotated target image plus a concise brief containing accession, chain/range, intended surface, three to six justified hotspots, and caveats.
Stage 2 — Freeze a reproducible run configuration
Record input identifiers/checksums, preparation commands, tool releases/commits, environment, checkpoints, hardware, exact commands/configurations, seeds, output paths, and deviations from the course’s known-good stack.
Exit artifact: another learner can reconstruct the smoke test without guessing a hidden setting.
Stage 3 — Generate and filter backbones
Begin with one or two outputs. Confirm the input chain, contig/chain break, hotspots, length, output location, and structure integrity. Then scale to a teaching-size campaign. Show the full distribution—not only favorites—and apply a consistent first-pass filter for topology, target/hotspot contact, clashes, and diversity.
Exit artifact: labeled gallery/table of successful and failed backbones with keep/reject reasons and attrition counts.
Stage 4 — Design sequences
Generate several sequences per advanced backbone. Preserve the parent-backbone link and record model/settings, fixed positions, temperature/seeds, recovery, diversity, composition, and relevant sequence liabilities.
Exit artifact: compact candidate table linking every sequence to its parent backbone and settings.
Stage 5 — Validate the designed complex
Predict the binder–target complex and inspect per-chain fold confidence separately from interface confidence and relative-chain uncertainty. Evaluate PAE, interface geometry, hotspot recovery, clashes, consistency across predictions, and at least one independent check beyond a single headline score.
Exit artifact: comparable figures and metrics for at least two finalists, with explicit uncertainty.
Stage 6 — Diagnose a failure and iterate
Choose at least one failed or misleading result. Form a cause hypothesis, make one targeted change, and ask whether the next run supports that diagnosis. Changing many settings at once weakens the inference.
Exit artifact: before/after comparison connecting failure → diagnosis → change → evidence → revised conclusion.
Stage 7 — Make a conditional selection
Define selection criteria before choosing a finalist. Compare the candidate you would advance with credible alternatives. State trade-offs, remaining uncertainty, and the next computational or experimental test that could change the decision.
Exit artifact: one-page final selection memo or equivalent final section.
Completion and self-assessment
Your format may be a report, notebook, repository README, narrated deck, or video, but the seven artifacts must remain easy to find and connected to their underlying files.
Use the Capstone Evidence Guide & Rubric to score:
- reproducibility;
- structural and computational evidence;
- scientific judgment;
- iteration and learning; and
- communication.
The course completion standard is 13/20 or higher with no category below 2. A high course score means the computational case is clear and defensible. It does not predict experimental success.
If your reasoning is becoming score-driven, compare it with the worked selection example. The exemplar’s strongest candidate is not the one with the highest summary score; it is the one whose evidence best matches the biological mission while making the remaining defect explicit.
What to do next
- Download the evidence template.
- Choose a target and draft Stage 1 before running a model.
- Review the Stage 2 configuration with the rubric before scaling compute.
- Stop at each exit artifact; repair missing provenance or interpretation before continuing.
- Share the final portfolio with a peer if possible and ask them to identify one unsupported inference.