Target prioritization matrix

Use the Target prioritization matrix skill.

Seed / scope:   oncology — DepMap-essential, druggable protein kinases (cell-cycle & mitotic checkpoint)
Modality:       small molecule (also flag PROTAC tractability)
Criteria:       target class · small-molecule tractability · DepMap essentiality · genetic constraint (LOEUF) ·
                curated safety liabilities · competition (# drugs targeting) · clinical/literature/IP activity scan
Stance:         show the full field, most-essential first, so I can weigh validation vs therapeutic-window risk
Shortlist size: 12
Output format:  markdown table + CSV

Why not just ask the model?

Ask a plain model to "shortlist druggable, essential kinase targets for oncology and score them by essentiality, safety and competition" and it names a handful of famous kinases and invents the numbers — DepMap gene-effect scores, LOEUF values, "drugs in development" counts — that read authoritative and are fabricated.

• Plain model: a few remembered kinases; made-up essentiality, constraint, safety, and competition figures.
• Amass: the real GeneCore records for the live filtered search — every DepMap gene-effect (PLK1 −2.73, WEE1 −2.48, CDK1 −2.34…), every gnomAD LOEUF, every curated safety liability (AURKA: mutagenic / cardiac / neutropenia), and the real referencesDrugCore edge count per target (AURKA 19 → WEE1/TTK 2), verbatim from the returned fields.

Why a skill, not a prompt?

Because the value is the interview and the orchestration, not one question. The skill asks modality first, then which criteria to score, then the output format — and turns those answers into GeneCore enum filters (targetClass=ENZYME, isEssential=true, isDruggable=true, tractabilityModality). It unions two query angles to break the 10-result cap (10 + 10, overlap 4 → 16 unique), reads the target-intelligence fields straight from the search response, then fans out one get_amass_genecore_record per candidate for the referencesDrugCore competition edge and one TrialCore / BiomedCore / PatentCore search each for the activity scan. Then — because it is a general workbench — it offers to freeze this exact configuration into a dedicated, one-line reusable skill, so the workbench spawns the narrow, repeatable tool.

Honest scope

Candidates are a union of relevance-ranked top-10 GeneCore searches — a broadened sample, not a census. Target-intelligence fields are verbatim from Open Targets / gnomAD / DepMap as harmonized in GeneCore; null = "no data recorded," not zero. drugs_targeting is the length of each gene's referencesDrugCore edge — 0 would mean unprecedented, not undruggable. The TrialCore / BiomedCore / PatentCore columns are capped ≤10 samples: on this mature class all 12 kinases saturate at 10+, so those columns don't differentiate — the discriminating competition signal is the DrugCore edge, and the skill says so rather than presenting three dead columns as a ranking. And DepMap essentiality is double-edged — a pan-essential kinase is a strong dependency and a narrow-therapeutic-window risk.