[Forge] Attach DepMap dependency scores to therapeutic-target hypotheses done

Goal

For every hypothesis tagged therapeutic_target=True, fetch DepMap CRISPR
Chronos dependency scores via the depmap skill and persist a per-lineage
summary so the Domain-Expert persona can cite real essentiality data when
arguing druggability.

Why this matters

DepMap dependency scores are the gold-standard signal for "is knocking this
gene down lethal in cancer cell lines?". A therapeutic-target hypothesis whose
gene scores -1.0 across 1100 cell lines is fundamentally different from one
that scores 0.0 — but today the Domain Expert has no database-backed way to
distinguish them, so the Synthesizer's target_validity score drifts.

Acceptance Criteria

mean_chronos, median_chronos, n_cell_lines, percent_essential,
depmap_release, fetched_at). payload_json->>'therapeutic_target'='true' OR a
target_class IN ('kinase','GPCR','enzyme','transporter') heuristic
fires, and populates the table per lineage (lung, brain, blood, ...). skill (or DepMap public API at https://depmap.org/portal/api) and
returns dataclass results. mean Chronos per lineage when the hypothesis names a target gene. threshold marked at -0.5. checks can detect a skewed run.

Approach

Use the in-repo depmap skill if installed; otherwise hit the public

download manifest and cache the parquet under data/depmap/.

Hypothesis-level summary is computed once per release; a nightly job

re-runs whenever a new DepMap release ships (quarterly cadence).

Persona injection mirrors the GTEx pattern in

scidex/senate/personas/domain_expert.py.

Dependencies

• Quest q-555b6bea3848.
• DepMap skill or public download API.

Work Log

2026-04-27 — Implementation (task:bbdeb2ac-028b-45a6-9646-f16232658c3e)

All acceptance criteria implemented:

Migration: migrations/add_hypothesis_depmap_dependency.py — creates hypothesis_depmap_dependency table with all required columns plus UNIQUE
constraint on (hypothesis_id, lineage, depmap_release) and indexes on hypothesis_id and depmap_release. Applied and verified on live DB.

Client module: scidex/forge/depmap_client.py — follows the census_expression.py pattern with parquet cache under data/depmap/.
Primary path: DepMap portal API (/api/gene_dependency). Secondary path:
downloads Model.csv (~546KB) from DepMap 24Q2 figshare (article 25880521)
for cell-line lineage mapping. Falls back gracefully when network unavailable.
Uses DEPMAP_RELEASE = "24Q2" (latest accessible release; update to 24Q4
when figshare article becomes publicly available).

Backfill script: scripts/backfill_hypothesis_depmap.py — follows backfill_hypothesis_census.py pattern. Selects hypotheses where hypothesis_type IN ('therapeutic', 'therapeutic_genetics', 'pharmacological',
'combination_target', 'pathway_target') OR druggability_score IS NOT NULL,
with target_gene IS NOT NULL. De-duplicates API calls per gene. Upserts with ON CONFLICT DO UPDATE. Supports --dry-run, --stale-only, --ids, --limit.

API endpoint: GET /api/hypotheses/{hypothesis_id}/depmap — returns JSON {hypothesis_id, target_gene, depmap_essential_threshold, lineage_summaries[]}.

Bar chart in hypothesis page: _build_depmap_tab_html() helper generates
SVG horizontal bar chart with essentiality threshold at -0.5 (yellow dashed)
and -1.0 (red label). New "DepMap" tab added to /hypothesis/{id} tab strip.
Tab panel shows bar chart + data table with Chronos scores, %essential, and
verdict labels (Strong/Likely/—).

Domain Expert injection: _build_depmap_dependency_block() in agent.py
queries hypothesis_depmap_dependency by target_gene (joining through hypotheses) for pre-computed data, falls back to live dependency_for_gene()
API. Injected into domain expert expert_prompt before the feasibility
assessment round.

Note on data availability: DepMap portal API returns 403 in this environment
(likely IP-restricted). Backfill requires outbound network access. Run manually
after network access is confirmed:

python3 scripts/backfill_hypothesis_depmap.py --stale-only --limit 200