The Public Repurposing Index (PRI) 2.0

Strategic Enhancement and Operational Roadmap for Regulatory-Grade AI Drug Discovery

Author
Brewtanius Ink LLC · The CollectiveOS- Immortal Tek 

Abstract

The Public Repurposing Index (PRI) 2.0 transforms a general data aggregator into a regulatory-grade hypothesis engine for rare-disease drug repurposing. Grounded in the BioLink Model for semantic interoperability, PRI consolidates curated evidence into a transparent knowledge graph compatible with federal research ecosystems. A geometric, rewired heterogeneous GNN—augmented by zero-shot ranking—prioritizes repurposing candidates and returns explainable rationales via Captum/GNNExplainer. PRI’s Context of Use (COU) is scoped to support preclinical prioritization and non-binding patient stratification proposals; outputs are accompanied by a credibility plan aligned to the FDA’s model-credibility framework, a digital thread from source data to receipt, calibration and uncertainty reporting, and continuous monitoring. A pilot on GSD1b (ORPHA:365) uses empagliflozin as a retrospective positive control and preregisters prospective hypotheses for expert adjudication. Emphasizing openness, metabolic specialization, and a public-utility funding posture, PRI 2.0 delivers auditable AI: not only ranked candidates, but a defensible evidence trail regulators and clinicians can evaluate.

Executive Summary

Problem. More than 30 million Americans live with one of 7,000+ rare diseases; roughly 95% lack an approved therapy. Conventional R&D economics remain prohibitive for ultra-rare indications.

Pivot. PRI 2.0 moves from “aggregating data” to harmonizing evidence and from opaque scoring to explainable, auditable AI aligned with current FDA and EMA expectations.

COU. COU-A: AI-assisted hypothesis generation and non-binding patient stratification proposals for research planning (risk = medium). Outputs are not standalone clinical claims.

Core design. A BioLink-conformant knowledge graph (KG); a heterogeneous + transformer GNN with rewiring to mitigate oversquashing; Captum-grade XAI that returns clinician-readable subgraphs; and Proof-Vault receipts that preserve a “digital thread” from source to decision.

Pilot. GSD1b (ORPHA:365): retrospective validation that recovers empagliflozin under a time cutoff, plus preregistered prospective candidates with expert adjudication.

Operations & sustainability. Public-utility posture, metabolic specialization, and partnerships with FDA ARC, NIH RDCRN, and ARPA-H. PRI focuses on reproducibility, transparency, and interoperability—maximizing trust and grant viability.

Outcome. A platform that accelerates plausible, defensible hypotheses for rare diseases, reduces validation cost, and delivers documentation that withstands scientific and regulatory scrutiny.

1. Introduction

Repurposing leverages the sunk safety and manufacturing knowledge of approved drugs to address the long tail of rare diseases where bespoke R&D is economically infeasible. Yet the last wave of AI-for-drug-discovery often prioritized scale over semantic rigor, and speed over transparency. The 2024–2025 regulatory turn—FDA’s model-credibility expectations and EMA’s risk-based lifecycle guidance—raises the bar. PRI 2.0 responds by unifying semantic data governance, geometric learning, explainability, and a proof-oriented operational model.

2. Context of Use (COU) & Model-Credibility Plan

2.1 Context of Use (COU-A)

Purpose. AI-assisted, evidence-linked hypotheses for drug–disease associations in rare metabolic and genetic disorders. Outputs support preclinical prioritization and non-binding patient stratification proposals; they are not prescriptive clinical decisions.

Risk. Medium (research-planning impact, not direct clinical action).

2.2 Credibility Plan (aligned to FDA model-credibility)

1. Question of interest. Which approved drugs plausibly modulate disease-relevant pathways to justify preclinical work and cohort hypotheses?

2. Risk assessment. Medium; model informs research proposals and evidence generation plans.

3. Data fitness & governance. BioLink mapping; identifier normalization (ORPHA↔MONDO; DrugBank↔ChEMBL; UniProt↔Ensembl); versioned ETL; license checks; KG integrity tests.

4. Verification. Schema conformance; unit/integration tests for ETL and KG loaders; deterministic snapshot hashes.

5. Validation.

   • Retrospective: top-k recall on known positives (e.g., GSD1b→empagliflozin under a historical cutoff), AUC-PR.

   • Prospective: expert adjudication of preregistered novel pairs with literature context and XAI subgraphs.

6. Explainability. Captum/GNNExplainer: edge/node importances compressed to a readable subgraph with confidence notes.

7. Calibration & uncertainty. Temperature scaling; Expected Calibration Error; uncertainty intervals on scores.

8. Monitoring. Version registry; drift checks; re-validation triggers after material KG or code changes.

9. Acceptance criteria. Positive-control recovery in top-10; ≥70% clinician-rated plausibility for prospective sets; XAI present for all recommendations.

3. Macro Landscape & Positioning (2024–2025)

Regulatory watershed. The era of “black-box curiosity” is over; model credibility must match COU. PRI treats the platform as a scientific instrument, supplying evidence trails and explainability.

Funding shift. Venture capital contracted in 2024 (Biotech Winter); public program investment (e.g., ARPA-H) expanded, favoring open, governable infrastructure.

Collaborative competition.

• EveryCure validates the non-profit, KG-centric model at general scale. PRI differentiates through metabolic specialization and radical transparency.

• Commercial PaaS (e.g., disease-vertical players) demonstrate the efficacy of niche depth. PRI focuses on metabolic pathways to achieve a “niche monopoly” in that domain.

PRI posture. A public-utility, BioLink-native, metabolic specialist node that interlocks with FDA ARC, NIH RDCRN, and Translator/Monarch ecosystems.

4. System Overview & Digital Thread

Flow. Source → ETL (BioLink) → KG (Neo4j) → Feature Builder → Hetero+Transformer GNN with rewiring → Score + Uncertainty → XAI Subgraph → Envelope + Proof → User/Partner.

Digital thread. At each step, PRI records a hash (ETL spec, KG snapshot, model weights, XAI JSON) and writes an append-only Proof-Vault receipt, binding input lineage to outputs and policy context (COU, risk).

5. Data Governance: BioLink Semantic Layer

Rationale. BioLink supplies shared, computable semantics (classes such as ChemicalEntity/Drug, Disease, Gene, PhenotypicFeature; predicates such as treats, interacts_with, GeneToDiseaseAssociation). This prevents orphaned identifiers and inconsistent relations, the main sources of “garbage in, garbage out.”

ETL principles.

• Identifier policy. ORPHA↔MONDO; DrugBank↔ChEMBL; UniProt↔Ensembl (Biolink Model Toolkit assisted).

• Curation & quality. Deduplicate; merge synonyms; annotate evidence level; track licenses.

• Conformance. Validate class and predicate usage; forbid dangling references.

• Artifacts. Versioned ETL specs and KG snapshot hashes lodged in Proof-Vault.

Graph store. Neo4j (or Neptune) to support multi-hop, lineage-aware queries and downstream feature-building.

6. Modeling: Geometric/Transformer GNN with Rewiring

Challenge. Standard GNNs oversquash long-range signals and oversmooth deep layers—fatal for mechanistic biological cascades.

Architecture.

• Heterogeneous GNN (distinct parameters by node/edge type) preserves BioLink semantics.

• Graph Transformer attention enables global information flow.

• Rewiring (virtual edges) connects distant but functionally related nodes to mitigate oversquashing.

• Zero-shot ranking (TxGNN-style) leverages metric learning on a large medical KG, enabling generalization to sparsely labeled rare diseases.

• Calibration/uncertainty reports temper predictions for decision support.

Output. Ranked (drug,disease) pairs with uncertainty and an attached explanation subgraph.

7. Explainable AI (XAI): Clinician-Readable Rationales

Goal. Convert predictions into mechanisms clinicians can evaluate.

Methods.

• Captum (Integrated Gradients/Saliency) and GNNExplainer identify influential nodes/edges.

• Subgraph compression. The importance map collapses to a small, connected rationale (Drug→Protein→Pathway→Disease) with weights, confidence, and references.

Deliverable. Explanation Cards per candidate: score/uncertainty, rationale subgraph, literature context, and caveats.

8. Evaluation Plan & Acceptance Criteria

Retrospective evaluation.

• Task. Recover known positives under a historical data cutoff (e.g., GSD1b→empagliflozin).

• Metrics. Recall@k (10/25), AUC-PR, ECE (calibration quality).

• Success. Positive control appears in top-10 with coherent XAI.

Prospective evaluation.

• Task. Pre-register 3–5 novel candidates for metabolic indications with XAI.

• Adjudication. KOL triad (computational biology, rare-disease clinician, patient advocate) rates plausibility.

• Success. ≥70% “plausible” ratings; consistent subgraph narratives aligned with known biology.

Monitoring. Trigger re-validation on KG or code changes; track drift; maintain a model registry.

9. Pilot Protocol: GSD1b (ORPHA:365)

Rationale. Clear metabolic mechanism; empagliflozin is a strong positive control; registry data (e.g., U-IMD) enables external checks.

Design.

• Data freeze. Assemble BioLink KG from Orphanet/DrugBank/ChEMBL/Reactome up to a pre-discovery cutoff.

• Modeling. Train hetero + transformer GNN with rewiring; apply zero-shot ranking.

• Outputs. Top-k list with XAI cards; flagged references.

• Endpoints.

  • Primary: Retrospective recovery of empagliflozin in top-10.

  • Secondary: Prospective hypotheses rated ≥70% plausible by KOL panel.

Artifacts. ETL spec & hash; KG snapshot hash; training config; model card; XAI JSON; Proof-Vault receipts. All released in a reproducible bundle.

10. Business & Operations: Public-Utility Model

Posture. Operate as open semantic infrastructure; prioritize trust, interoperability, and auditability.

Tiers.

• Open. Free KG queries & base predictions/XAI for academics and non-profits.

• Partner. Paid “Deep Dives” (expanded XAI & semantic curation) and compute-heavy geometric analyses for patient groups/biotechs.

• Grants. ARPA-H/NIH funding for platform sustainability; FDA ARC engagement for evolving COU alignment.

Differentiation. Radical transparency (open weights/artifacts); metabolic-first specialization; regulatory-aligned documentation.

11. Partnerships & Funding Roadmap

• FDA ARC / Rare Disease Innovation Hub. Engage early on COU and credibility artifacts; request technical feedback on acceptance thresholds and evidence presentation.

• NIH RDCRN DMCC. Offer BioLink harmonization and KG services to reduce data silos; co-develop federation pathways with Translator/Monarch compatibility.

• ARPA-H (e.g., THRIVE). Propose PRI as an evidence generator for genetic medicine adjuvants and prioritization, anchored by the GSD1b pilot.

• Academic consortia. Collaborate with inborn-errors-of-metabolism centers to co-author methods and pilot results.

12. Reproducibility, Disclosures & Conclusion

Reproducibility.

• Release a model card, dataset datasheets, ETL spec, KG snapshot hash, training config, and XAI JSON examples.

• Provide a Docker compose demo (KG + minimal UI) and a Proof-Vault ledger with SHA-256 per artifact.

• Host artifacts under a permissive license (code: Apache-2.0; docs: CC BY-4.0; data per source licenses).

Limitations. PRI is a hypothesis-generation tool; not prescriptive clinical guidance. KG bias and incompleteness can affect outputs. We mitigate via BioLink harmonization, explainability, calibration, expert adjudication, and continuous monitoring.

Disclosures. Conflicts of interest will be documented. Funding, acknowledgments, and advisory roles are disclosed in the artifact bundle and releases.

Conclusion. PRI 2.0 reframes repurposing as auditable AI for rare diseases—transparent, semantic, explainable, and aligned with the regulatory moment. Its value lies not only in producing candidates but in the evidence trail that links sources, semantics, models, and rationales into a defensible whole. With the GSD1b pilot and targeted partnerships, PRI is positioned to become a durable node in the national rare-disease knowledge infrastructure.

Appendix A — Submission-Grade Checklists

COU & Credibility One-Pager

• COU stated (scope, risk, decision impact).

• Credibility activities mapped to metrics (data fitness; V&V; XAI; calibration; monitoring).

• Acceptance criteria & re-validation policy.

Digital Thread Diagram Content

• Boxes: Source → ETL → KG → Features → Model → XAI → Envelope → Proof.

• Labels: version, hash, timestamp, license, policy profile.

• Output: subgraph (weighted), explanation card, Proof-Vault receipt.

Pilot SOP (GSD1b)

• Retrospective timeline and data freeze.

• ETL/KG build specification; pre-registered metrics.

• Reviewer instructions; reporting template; artifact checklist.

Appendix B — Methods (Expanded)

Data sources & mapping. Orphanet, DrugBank, ChEMBL, Reactome, UniProt/Ensembl, and selected literature corpora, harmonized via the BioLink Model Toolkit. All ETL pipelines produce versioned specs and snapshot hashes.

Knowledge graph. Neo4j (or Neptune) stores BioLink classes/predicates; graph integrity checks enforce schema; lineage tables track provenance.

Model. Hetero GNN with transformer attention; rewiring adds functional “shortcuts” to mitigate oversquashing; zero-shot metric learning module provides cross-disease generalization; calibration yields trustworthy scores.

Explainability. Captum (Integrated Gradients/Saliency) and GNNExplainer yields edge/node importances; subgraphs are compressed to human-legible rationales with references.

Proof-oriented operations. Each run appends a signed receipt containing envelope context, artifact hashes, and outputs—supporting a reproducible “digital thread.”

Appendix C — Advisory & Governance

Advisory Board targets.

• Geom. deep learning (modeling and evaluation).

• Rare-disease clinical (inborn errors of metabolism).

• Patient advocacy (community perspective and artifact usability).

Governance. QC → GATA → PRIME rails; policy packs defining acceptable risk and override procedures; model registry and operational thresholds for re-validation.

License
Text: CC BY-4.0. Code/ETL tools: Apache-2.0. Data: subject to source licenses and terms.