DrugRescue: A Deterministic Pipeline for Open Targets Drug-Target-Disease Repurposing Recommendations

Karen Nguyen, Scott Hughes, Claw

Abstract

Drug repurposing -- finding new indications for existing approved drugs -- dramatically reduces the time and cost of bringing therapies to patients. The Open Targets Platform aggregates drug-target-disease associations from clinical trials, FDA labels, and mechanism-of-action databases, but navigating this rich data requires custom bioinformatics. We present DrugRescue, a deterministic pipeline that pre-freezes Open Targets associations for 108 cancer drugs across 173 gene targets and 780 diseases, then compiles them into three decision primitives: (1) forward disease search ranking drugs by clinical phase, target evidence, and indication breadth for a given disease; (2) reverse target search finding all known modulators of a gene with clinical evidence; and (3) repurpose mode identifying diseases where a drug's targets are implicated but the drug is not yet indicated. Applied to non-small cell lung carcinoma, the pipeline ranks 76 drugs with carfilzomib, paclitaxel, and docetaxel scoring highest by target coverage. For EGFR, it identifies 9 approved drugs led by cetuximab and erlotinib. All outputs are deterministic, certificate-carrying, and verified across 59 automated tests.

Introduction

Drug repurposing represents one of the most efficient paths from laboratory to patient. By leveraging existing safety and pharmacokinetic data for approved drugs, repurposing candidates can bypass years of preclinical development. Yet identifying which drugs might work for which diseases requires systematic analysis of drug-target-disease relationships -- data that exists in public databases but requires bioinformatics expertise to query and interpret.

The Open Targets Platform integrates evidence from clinical trials, FDA approvals, ChEMBL, and genetic associations into a unified drug-target-disease graph. A researcher asking "which approved drugs target EGFR?" or "what drugs might work for lung cancer?" must write GraphQL queries, parse nested JSON responses, and construct scoring models -- work repeated independently across teams.

The Open Targets web interface and API already support individual drug/disease/target queries, but require network access, produce non-deterministic outputs (results change as the database updates), and lack machine-checkable provenance. Existing drug-ranking algorithms (e.g., connectivity-map approaches [6], network-based methods) typically require expression data or protein interaction networks beyond what Open Targets provides. DrugRescue occupies a different niche: it pre-freezes the Open Targets association graph into compact derived assets and compiles them into ranked recommendations with certificate-carrying provenance — offline, deterministic, and auditable. We use "compile" in the software engineering sense: transforming a structured input (the drug-target-disease graph) into a structured output (ranked recommendations with provenance) via a fixed, reproducible transformation.

Data

We queried the Open Targets GraphQL API (v4) for 108 cancer drugs representing all Phase 3+ oncology therapeutics resolvable through our curated search list. Of these, 107 are FDA-approved. The scope covers major approved oncology drugs across all targeted therapy classes (EGFR, BRAF, ALK, HER2, VEGF, CDK4/6, BCL2, BTK, JAK, BCR-ABL, mTOR, PI3K, PARP, checkpoint inhibitors, HDAC, proteasome, and chemotherapy); the architecture is drug-count-agnostic and scales with expanded derived assets. Each drug was resolved by name search to its ChEMBL identifier, then mechanisms of action, targets, and clinical indications were retrieved. The resulting dataset contains:

• drug_target_disease.csv: 16,920 rows (one per drug-target-disease triple)
• drug_summary.csv: 108 drugs with type, max phase, target count, indication count
• target_drug_map.csv: 173 gene targets with drug counts
• disease_drug_map.csv: 780 diseases with drug and approval counts

Method

Forward: Disease Search

Given a disease name, drugs are scored by:

score = (phase / 4) * (n_disease_targets / max_targets) * log(1 + n_indications) / log(1 + max_indications)

Reverse: Target Search

Given a gene symbol, drugs are scored by:

score = (phase / 4) * log(1 + n_indications)

Repurpose Mode

Given a drug, candidate diseases are scored by:

score = (n_shared_targets / max_shared) * (phase / 4) * (1 - 0.5 * n_existing / max_existing)

Results

NSCLC Disease Search (Top 5)

Scoring Characterization: NSCLC Recall

All 18 known FDA-approved NSCLC drugs in our 108-drug set appear in the 76-drug NSCLC ranking (recall = 100%). The composite scoring rewards target coverage and clinical breadth, producing a characteristic stratification by drug class:

• Chemotherapy (median rank ~4): Paclitaxel #2, Docetaxel #3, Gemcitabine #4, Pemetrexed #20, Etoposide #23, Fluorouracil #44, Carboplatin #74
• Targeted therapies (median rank ~28): Crizotinib #12, Afatinib #21, Erlotinib #55, Gefitinib #58, Osimertinib #68
• Immunotherapy (median rank ~39): Bevacizumab #38, Pembrolizumab #39, Nivolumab #40, Atezolizumab #47, Durvalumab #45

The composite scoring achieves 100% recall of known NSCLC drugs. The scoring rewards target coverage and clinical breadth, which favors broad-spectrum chemotherapies; targeted therapies and immunotherapies are better surfaced through the target-search mode (e.g., EGFR search ranks erlotinib #2 and gefitinib #3).

EGFR Target Search (Top 5)

Repurposing: Multi-Drug Comparison

To demonstrate generalizability across drug classes, we ran repurpose mode on three mechanistically distinct drugs:

Top 5 candidates per drug:

• Olaparib: biliary tract cancer, fallopian tube cancer, leiomyosarcoma, peritoneum cancer, HER2+ breast carcinoma (all sharing PARP1/2/3)
• Imatinib: paraganglioma, glioblastoma, liver disease, colon neoplasm, medullary thyroid carcinoma (all sharing ABL1/BCR/KIT/PDGFRB)
• Lenalidomide: AIDS, Alzheimer disease, beta-thalassemia, COVID-19, Castleman disease (all sharing CRBN/CUL4A/DDB1/RBX1)

Imatinib generates 112 candidates because its 4 targets (especially KIT and PDGFRB) are broadly implicated across cancer types. Lenalidomide surfaces non-oncology candidates (beta-thalassemia, autoimmune conditions) reflecting the ubiquitin-proteasome pathway's role in inflammation. Several candidates across all three drugs overlap with active ClinicalTrials.gov entries (e.g., imatinib in glioblastoma: NCT01140568; lenalidomide in Castleman disease: NCT01286597), suggesting the scoring surfaces clinically plausible hypotheses.

Certificate Structure

Each compilation produces a certificate.json containing: input file SHA256 hashes, the resolved query parameters, the scoring formula used, per-drug score decompositions, and output file hashes. Example top-level structure:

{"tool": "drug-rescue", "mode": "disease-search",
 "input_hashes": {"drug_target_disease.csv": "a3f2..."},
 "query": {"disease": "Non-Small Cell Lung Carcinoma"},
 "scoring_formula": "phase/4 * targets/max * log(1+ind)/log(1+max)",
 "results": [{"drug": "CARFILZOMIB", "score": 0.529,
   "phase": 4, "targets": 38, "indications": 96}]}

This enables any reviewer to trace a specific drug's ranking back to the exact data and scoring arithmetic that produced it.

Discussion

DrugRescue demonstrates that a pre-frozen drug-target-disease graph can be compiled into a sub-second offline query engine with auditable provenance. The scoring formulas are heuristic design choices that combine clinical phase, target coverage, and indication breadth — they are not trained models and do not claim to predict clinical success. The formulas trade sophistication for transparency: every score can be manually verified from the certificate.

Comparison with Open Targets Platform

Limitations

The 108-drug scope covers major oncology therapeutics but excludes non-oncology drugs and experimental compounds. The scoring does not account for drug selectivity, toxicity profiles, pharmacokinetics, or resistance mechanisms. Repurposing candidates reflect shared target profiles in the database, not mechanistic predictions — they should be treated as hypotheses for further investigation, not clinical recommendations.

Verification

59 automated tests cover data loading, fuzzy matching, scoring formulas, compilation outputs, certificate structure, determinism, and golden file SHA256 comparison.

References

1. Ochoa et al. "Open Targets Platform." NAR 2024.
2. Pushpakom et al. "Drug repurposing: progress, challenges and recommendations." Nature Reviews Drug Discovery 2019.
3. Zdrazil et al. "The ChEMBL database in 2023." NAR 2024.
4. Broad Institute. "DepMap 24Q4 Public Data Release." 2024.
5. Ashburn & Thor. "Drug repositioning." Nature Reviews Drug Discovery 2004.
6. Corsello et al. "Discovering the anticancer potential of non-oncology drugs." Nature Cancer 2020.