OCDocker.Workbench.AblationProtocolSimilarity module

Expanded feature-policy similarity for OCScore ablation protocols.

Each executed study uses feature_policy_metadata.json from one replica when available. Catalog-only policies fall back to bundled YAML policies resolved with apply_feature_policy against the same workspace candidate discovery path used by OCDocker.Workbench.AblationDesign.

The read-only payload powers GET /api/ablation-protocol-similarity and the Protocol similarity zone in the Ablation dashboard tab.

OCDocker.Workbench.AblationProtocolSimilarity.build_ablation_protocol_similarity_analysis(root, *, reference_policy=None, metric=None, include_catalog_only=False, max_depth=6)[source]

Build protocol similarity analysis for one Workbench root.

Each selected policy is expanded with the same candidate feature universe, compared with Jaccard similarity, clustered for visualization, and annotated with optional outcome metrics and reference diffs.

Parameters:

  • root (str or pathlib.Path) – Workbench root to scan.

  • reference_policy (str or None) – Reference policy name for diffs. Defaults to full_ocscore when present.

  • metric (str or None) – Optional comparison metric used for cluster mean overlays.

  • include_catalog_only (bool) – When True, always include the full bundled and workspace policy catalog. When False, restrict to policies with executed workspace folders or ablation result manifests. If none are detected, the full catalog is still returned so feature-level similarity remains usable.

  • max_depth (int) – Maximum scan depth for result manifests when resolving metric overlays.

Returns:

Similarity analysis payload for API and dashboard consumers.

Return type:

WorkbenchAblationProtocolSimilarity

Ablation protocol similarity

This helper compares OCScore ablation protocols by their expanded feature sets rather than by raw YAML include/exclude tokens. Each bundled or workspace policy is resolved with OCDocker.OCScore.Utils.FeaturePolicy.apply_feature_policy() against a shared candidate feature universe discovered from replica feature_policy_metadata.json or, as a fallback, from raw_prepare tables via OCDocker.Workbench.AblationDesign.

Wildcard patterns such as ligand_* therefore expand to every matching column before Jaccard similarity, family rollups, reference diffs, and hierarchical clustering are computed. Clustering uses feature similarity only; an optional comparison metric overlays mean outcomes per cluster without changing the cluster assignment.

The read-only payload is exposed by the local API at GET /api/ablation-protocol-similarity and rendered in the Protocol similarity zone of the Ablation dashboard tab (heatmap, family grid, cluster summaries, and reference diffs). Outcome-oriented ablation tables remain in OCDocker.Workbench.Ablation.

curl 'http://127.0.0.1:8765/api/ablation-protocol-similarity?metric=auc:max&reference=full_ocscore&include_catalog_only=true'

Query parameters:

  • metric — optional comparison metric for cluster mean overlays (same syntax as OCDocker.Workbench.Ablation.parse_ablation_metric()).

  • reference — reference policy for add/remove diffs (defaults to full_ocscore when available).

  • include_catalog_only — defaults to true; when false, restrict to executed workspace folders (API/CLI). The dashboard always loads the full catalog and filters client-side.