Capstone Evidence Guide & Rubric

How to know when your binder-design project is complete—and defensible

Purpose

The capstone is not complete merely because every command ran. A strong submission makes a reproducible, evidence-based argument for why one design should move forward and why the alternatives should not.

Use this page at three points:

  1. Before starting, to plan what you will save.
  2. At each milestone, to check whether your evidence is sufficient.
  3. Before finishing, to self-assess the complete project.

Download the evidence template

Required Evidence Portfolio

Your final format can be a report, notebook, repository, slide deck, or video, but it should contain all seven artifacts below.

1. Target and epitope brief

  • Target name, biological role, structure accession, chain, and residue range
  • The binding surface or epitope you intend to target
  • Three to six hotspot residues and the evidence used to choose them
  • Important caveats such as missing density, glycosylation, alternate conformations, cofactors, or chain-numbering differences

Completion evidence: one annotated structure image plus a concise rationale.

2. Reproducible run configuration

  • Tool release or commit, environment, model checkpoint, and hardware
  • Exact input-file identifiers and checksums when practical
  • Commands, configuration files, random seeds, and output paths
  • Any deviations from the course’s known-good stack

Completion evidence: another learner could reconstruct the run without guessing hidden settings.

3. Backbone-generation evidence

  • Number and type of designs generated
  • Representative gallery showing both successful and failed outputs
  • Topology, target contact, hotspot contact, clashes, and diversity observations
  • A documented first-pass filter

Completion evidence: a gallery or table that explains why candidates were kept or removed.

4. Sequence-design evidence

  • Sequence-design method and settings
  • Number of sequences generated per backbone
  • Sequence recovery, diversity, composition, and any fixed residues
  • Checks for obvious liabilities relevant to your target or intended use

Completion evidence: a compact candidate table linking sequences back to their parent backbones.

5. Structural validation

  • Complex-prediction method and settings
  • Per-chain confidence, interface confidence, and PAE interpretation
  • Interface geometry, hotspot recovery, clashes, and model agreement
  • At least one independent check beyond a single confidence score

Completion evidence: figures and metrics that distinguish local fold confidence from interface confidence.

6. Failure analysis and iteration

  • At least one failed or misleading result
  • Your diagnosis of the likely cause
  • The change you made in response
  • Whether the next run supported your diagnosis

Completion evidence: a short before/after comparison. A scientifically useful failure is stronger evidence than an unexplained success.

7. Final selection memo

  • The candidate you would advance and the alternatives you rejected
  • The criteria and trade-offs behind that decision
  • Remaining uncertainty and the next computational or experimental test
  • A clear statement that computational confidence is not experimental validation

Completion evidence: a one-page decision memo or equivalent final section.

Twenty-Point Self-Assessment Rubric

Score each category from 1–4. Use the lowest description that still fits the work consistently.

A. Reproducibility

  • 4 — Reconstructable: inputs, versions, commands, seeds, outputs, and deviations are complete and internally consistent.
  • 3 — Mostly reproducible: the main run can be repeated, but one or two minor details require inference.
  • 2 — Partially documented: settings are summarized, but exact inputs or commands are missing.
  • 1 — Not reproducible: results cannot be connected reliably to a specific configuration.

B. Structural and computational evidence

  • 4 — Converging evidence: multiple appropriate metrics and visual checks support the conclusions, with limitations interpreted correctly.
  • 3 — Adequate evidence: core confidence and interface checks are present, but validation is narrow.
  • 2 — Score-driven: conclusions rely primarily on one summary score or an attractive structure image.
  • 1 — Unsupported: outputs are shown without meaningful validation.

C. Scientific judgment

  • 4 — Defensible decisions: selection criteria are explicit, trade-offs are weighed, failures are investigated, and uncertainty changes the conclusion appropriately.
  • 3 — Reasonable decisions: choices are explained and major limitations are acknowledged.
  • 2 — Descriptive decisions: observations are listed, but the reasoning connecting them to selection is weak.
  • 1 — Completion-based: a candidate is selected mainly because the pipeline produced it.

D. Iteration and learning

  • 4 — Hypothesis-driven iteration: a failure or comparison motivates a specific change that is then evaluated.
  • 3 — Documented iteration: multiple settings are compared and their effects are discussed.
  • 2 — Repetition without diagnosis: several runs are shown, but changes are not tied to a clear hypothesis.
  • 1 — Single pass: no meaningful iteration or failure analysis is included.

E. Communication

  • 4 — Clear evidence trail: figures, tables, labels, captions, and narrative allow a reader to follow the complete argument efficiently.
  • 3 — Understandable: the story is coherent with only minor gaps or presentation problems.
  • 2 — Fragmented: key outputs exist, but their relationships or significance are hard to follow.
  • 1 — Incomplete: essential artifacts or explanations are absent.

Interpreting Your Score

  • 17–20: Portfolio ready. The work presents a defensible computational argument and clearly identifies what remains unproven.
  • 13–16: Complete with revisions. Strengthen the lowest-scoring category before presenting or sharing the work.
  • 9–12: Pipeline run, not yet a scientific case. Add missing evidence, comparison, and decision logic.
  • 5–8: Reconstruct the evidence trail. Confirm inputs and outputs before drawing conclusions.
ImportantCompletion threshold

For course completion, aim for at least 13/20, with no category below 2. The score is a self-assessment aid—not evidence that a binder will work experimentally.

Suggested Review Sequence

  1. Review your target brief before spending substantial compute.
  2. Review reproducibility and backbone evidence before sequence design.
  3. Review sequence and validation evidence before selecting a finalist.
  4. Complete the rubric and revise the lowest category.
  5. Compare your reasoning with the worked selection example.

← Capstone overview

Worked example →