RareCure: An Open-Source AI Pipeline for Context-Adaptive Treatment Discovery in Rare Solid Tumors

Background: Patients with rare solid tumors — particularly soft tissue sarcomas, encompassingover 50 histological subtypes affecting approximately 13,000 Americans annually — face acritical treatment discovery gap. Fewer than 5% of sarcoma subtypes have dedicated clinicaltrials, and five-year overall survival for metastatic disease remains below 20%.Objective: This work introduces RareCure, an open-source AI pipeline independently designedby the author to automatically generate evidence-ranked therapeutic option dossiers for raresolid tumors from genomic or clinical data.Methods: RareCure integrates six computational modules: (1) somatic variant processing withtiered annotation, (2) neoantigen prediction with dual HLA resolution, (3) drug-gene interactionmatching across four curated databases with harmonized evidence scoring, (4) clinical trialeligibility screening with ontology-aware query expansion, (5) retrieval-augmented evidence generationfrom indexed biomedical literature, and (6) a context-adaptive orchestration agent thatdynamically adjusts ranking weights using large language model reasoning with deterministicweight clamping for reproducibility. The pipeline was validated retrospectively on 260 soft tissuesarcoma patients from The Cancer Genome Atlas (TCGA-SARC) using a cloud-based LLM forresearch analysis of de-identified public data. The architecture supports fully offline operationvia locally hosted open-source models for institutional deployment scenarios involving protectedhealth information.Results: RareCure identified at least one Tier 1 or Tier 2 drug match (FDA-approved orgenomically matched) in 30.0% of 260 patients (78/260; 95% CI: 24.5–36.0%). Biomarker-drivenmatching from patient-specific somatic mutations was achieved in 78.8% of patients (205/260;95% CI: 73.4–83.6%). Total compute cost for the 260-patient cohort was $303.74 ($1.17 perpatient). Deterministic weight clamping was triggered in 0.0% of standard evaluation runs;adversarial prompt testing confirmed that the clamping mechanism correctly intercepted andcorrected extreme weight distributions.Conclusions: RareCure demonstrates that end-to-end treatment discovery for rare solid tumorscan be automated, identifying clinically actionable options in nearly one-third of sarcomapatients at approximately 4,000-fold lower cost than commercial genomic interpretation services.This work contributes to the national effort to bring precision oncology to underserved cancerpopulations. Available at https://github.com/DanielMartin-Arogyasami/RareCure underthe MIT license.