prism

PRISM — Predicate-defined, Reproducible, Identifier-bridged, Set-intersection Mining

PRISM answers "what entities are common to property set A and property set B?" on linked open data without relying on recall. A remembered list of candidates is a hypothesis, not a set. PRISM replaces hand-seeded lists with reproducible predicate chains, then triangulates weak axes across multiple evidence sources so the intersection does not silently collapse.

Before you begin

• Call TogoMCP_Usage_Guide (or TogoMCP-Test:TogoMCP_Usage_Guide) first, and run its GATE 0 classification.
• Read references/sparql-templates.md — parameterized queries for every phase (hierarchy expansion, GO→UniProt, PubTator co-occurrence, ClinVar gene-anchored significance, ChEMBL targets, Reactome membership, TogoID bridging).
• Read references/worked-example.md for a full end-to-end run with a filled-in provenance ledger.

When to use vs. not

Use PRISM when the question is an intersection of properties ("both A and B", "common to", "druggable target for disease", "shared genes", "overlap between"). Do not invoke the full machinery for a single-fact lookup ("what is the UniProt ID of CETP?") or a single-axis enumeration ("list genes in pathway X") — those are ordinary TogoMCP queries. The tell for PRISM is two or more constraints of different kinds that must be satisfied simultaneously.

---

The PRISM mnemonic → five phases

P — Pose as sets & Predicate-define each axis   (kill recall)
R — Reach the full hierarchy                     (transitive closure)
I — Integrate multi-source evidence              (triangulate weak axes)
S — Stitch by stable ID & intersect              (set algebra)
M — Measure actionability & audit provenance     (stratify + ledger)

Do the phases in order. The two most-skipped, most-damaging-to-skip steps are R (hierarchy expansion) and I (triangulation). Skipping R silently under-covers a set; skipping I lets a single biased source collapse the intersection.