OpenOncology: An Open-Source Framework for Evidence-Based Drug Matching and De Novo Custom Drug Discovery in Precision Oncology

OpenOncology: An Open-Source Framework for Evidence-Based Drug Matching and De Novo Custom Drug Discovery in Precision Oncology | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article OpenOncology: An Open-Source Framework for Evidence-Based Drug Matching and De Novo Custom Drug Discovery in Precision Oncology Aashish Kharel This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9707913/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Precision oncology depends on rapid, evidence-based matching of tumor variants to approved therapies. However, two compounding problems limit access for most patients worldwide: first, the interpretation infrastructure remains locked behind institutional subscriptions; second, even well-resourced precision oncology pipelines return empty outputs when no approved or repurposed drug exists for a patient’s specific mutation a complete dead-end that affects the majority of patients with rare or non-hotspot variants. Both problems are structural, not scientific. Methods We present OpenOncology, a fully open-source platform that solves both problems in In the upper you can see it's left aligned especially in, I think, methods. If we go down in the introduction, everything is center aligned so from front to last it is like a paragraph type like this. sequence. Stage one performs a clinical-grade variant calling workflow (FastQC → BWA-MEM2 → GATK), pathogenicity scoring (AlphaMissense), protein structure prediction (AlphaFold Server), molecular docking (DiffDock), and drug ranking from a weighted composite of OncoKB actionability, OpenTargets evidence, COSMIC frequency, clinical trial phase, and binding confidence. AlphaFold Server and DiffDock are computationally intensive external services; throughput in production deployments is subject to rate limits and available hardware. Stage two triggered when stage one finds no approved or repurposed match executes a fully automated custom drug discovery workflow: it queries ChEMBL and OpenTargets for lead molecules against the patient’s specific target, scores oral bioavailability via Lipinski Rule of Five, generates a mutation specific AlphaFold protein structure, and assembles a manufacturer-ready discovery brief that pharmaceutical companies can bid on through an integrated marketplace. A crowdfunding module enables patients to raise resources for custom synthesis. Results Validation against a blinded 50-case oncologist holdout yielded Hit@3 = 0.900, Standard Precision@3 = 0.508 (ceiling: 0.625), Normalised Precision@3 = 0.817, Mean Reciprocal Rank = 0.883, and zero false-positive recommendations. The 50-case holdout included 12 Level 3–4 literature-sourced cases and 6 negative controls, representing a deliberately harder validation set than smaller prior holdouts; the metric profile reflects increased case difficulty. Stage two (custom drug discovery) validation is structural discovery briefs are verified to contain real ChEMBL and OpenTargets records; clinical validation of lead molecule selection requires experimental binding assays outside the scope of this release. Equivalence-adjusted oncologist concordance reached 100% at both Top-1 and Top-3 across 36 actionable TCGA cases. TCGA benchmarks at 100 and 200 patients demonstrated 100% pipeline coverage with zero empty outputs every patient received either an approved drug recommendation or a structured custom discovery brief. Conclusions OpenOncology is the first open-source precision oncology platform to provide a complete, safe escalation pathway from approved drug matching through to de novo custom drug discovery for patients with no existing therapeutic option. All code, benchmark scripts, and validation artifacts are publicly available at github.com/immortal71/openoncology under the MIT licence. Oncology Bioinformatics precision oncology drug repurposing custom drug discovery de novo drug design variant interpretation OncoKB AlphaMissense DiffDock ChEMBL OpenTargets open-source bioinformatics low and middle-income countries 1. Introduction Cancer is the second leading cause of death globally, responsible for approximately 10 million deaths annually [ 1 ]. The emergence of precision oncology matching therapies to the molecular profile of individual tumors has demonstrated substantial survival benefits across multiple cancer types. Landmark approvals including imatinib for BCR-ABL-positive chronic myeloid leukemia, trastuzumab for HER2-amplified breast cancer, and osimertinib for EGFR-mutant non-small cell lung cancer have established the paradigm [ 2 , 3 ]. However, precision oncology has two compounding failure modes that limit its reach. The first is infrastructural: the tools that support molecular tumor board decisions cBioPortal, OncoKB, and commercial variant interpretation systems require institutional access, electronic health record integration, or subscription licensing [ 4 , 5 ]. For patients in low- and middle-income countries (LMICs), and for uninsured patients in high-income settings, this infrastructure is functionally unavailable. The second failure mode is more fundamental, and has received less attention: even when variant interpretation infrastructure is available, the majority of patients with rare or non-hotspot mutations receive an empty output. There is no approved drug. There is no guideline-supported repurposing candidate. Existing tools commercial or open-source have no response to this situation beyond a negative result. The patient, the oncologist, and the tumor board are left with nothing actionable. This second failure affects more patients than is commonly acknowledged. TCGA cohort data suggests that the fraction of patients with a direct FDA-approved targeted therapy match ranges from 7–36% depending on tumor type and variant selection criteria (see Section 3 ). For the remaining 64–93% of patients, current precision oncology infrastructure provides no therapeutic pathway beyond generic cytotoxic chemotherapy. We present OpenOncology, an open-source platform designed to close both gaps simultaneously. For the infrastructure gap: the platform is entirely free, open-source, containerised, and deployable without institutional affiliation. For the empty-output gap: the platform implements a novel two-stage architecture in which stage one (approved drug matching and repurposing) automatically escalates to stage two (de novo custom drug discovery) when no match is found ensuring that every patient receives an actionable output regardless of how rare or poorly characterised their mutation is. The custom drug discovery stage is the platform’s most significant technical contribution. When stage one finds no approved or repurposed match, an asynchronous worker queries ChEMBL and OpenTargets for lead molecules targeting the patient’s gene, applies Lipinski Rule-of-Five oral bioavailability scoring, generates a mutation-specific protein structure via AlphaFold Server, and packages the ranked lead molecules and structural data into a discovery brief that pharmaceutical manufacturers can receive and bid on through an integrated marketplace. This pathway does not fabricate drug recommendations. It generates a structured, evidence-sourced research brief that frames the next step for drug development. This paper describes the platform architecture, the two-stage drug matching and discovery methodology, and validation results from a blinded oncologist holdout and real-patient TCGA benchmarks. We discuss current limitations, the data access constraints inherent to independent research, and the roadmap toward institutional validation. 2. Methods 2.1 System Architecture OpenOncology is implemented as a containerised microservices stack. The backend is built on FastAPI with SQLAlchemy 2 async ORM and PostgreSQL 16. A Celery worker queue backed by Redis handles asynchronous genomics pipeline jobs, AI scoring tasks, and custom drug discovery workflows as independent, horizontally scalable workers. Object storage uses MinIO with AES-256 encryption. Authentication and role-based access control (patient, oncologist, administrator) are managed through Keycloak OIDC/OAuth2. The frontend is Next.js 14 with TypeScript. The full stack deploys via Docker Compose for local development or Kubernetes/Helm for production, with Prometheus and Grafana instrumentation. PHI handling is designed to conform to HIPAA § 164.308, § 164.310, and § 164.312 safeguards, with GDPR Article 17 (erasure) and Article 20 (data portability) implemented at the API layer. Production deployment handling real patient data requires independent security audit and regulatory review before clinical use. Weekly automated security scans cover Python dependencies (pip-audit), Node.js dependencies (npm audit), static analysis (Bandit, Semgrep OWASP), API security (ZAP baseline), and container vulnerabilities (Trivy). 2.2 Stage One: Variant Calling and Drug Matching Raw sequencing input (FASTQ or BAM) is processed through a Nextflow pipeline comprising quality control (FastQC, Trimmomatic), alignment to GRCh38 (BWA-MEM2), duplicate marking and base quality score recalibration (GATK), and somatic variant calling (GATK Mutect2). This workflow follows clinical-grade bioinformatics conventions; computational requirements for the full pipeline including AlphaFold and DiffDock impose practical constraints on throughput outside HPC environments. Variants are annotated using OpenCRAVAT with GRCh38 reference databases. Preflight checks validate input integrity before pipeline dispatch. Each missense variant is scored by AlphaMissense [ 7 ], a deep learning pathogenicity model queried against a 3.6 GB pre-built SQLite index. High-pathogenicity variants (score > 0.7) in genes without a direct OncoKB match trigger AlphaFold Server submission to generate a mutation-specific protein structure for downstream docking. Drug candidates are ranked by a weighted composite score integrating five evidence signals: Signal Weight Source DiffDock binding confidence 30% Molecular docking against mutation-specific structure OpenTargets association score 25% Multi-evidence target-disease linkage OncoKB actionability level 25% FDA-approved clinical evidence tier (Levels 1–2) AlphaMissense pathogenicity 10% Structural impact of the specific variant Clinical trial phase 10% Translational readiness (OpenTargets / ChEMBL) Resistance gating is applied before ranking. Known resistance mutations — including EGFR T790M (blocks erlotinib/gefitinib), KRAS G12C resistance alleles, and BRAF splice variants — are maintained in a curated resistance table. Any drug flagged by an applicable resistance gate is excluded from the output regardless of composite score. All drugs in the output are drawn exclusively from OncoKB Levels 1–2, ClinVar, or CIViC evidence items no drug names are inferred, generated, or hallucinated. If stage one returns candidates above the minimum evidence threshold, the top-three ranked drugs are returned to the patient dashboard with full provenance. The pipeline terminates. If stage one returns no candidates above threshold, the pipeline automatically escalates to stage two. 2.3 Stage Two: De Novo Custom Drug Discovery Stage two is the core novel contribution of this platform. It is the first open-source implementation of a fully automated escalation from clinical drug matching to de novo discovery brief generation for patients with no approved therapeutic option. When stage one finds no match, an asynchronous Celery worker (custom_drug_worker) is triggered with the patient’s target gene, variant, and disease context. The worker executes the following sequence: Step Action Data Source Output 1 Target-disease evidence retrieval OpenTargets GraphQL API Approved/clinical-stage drugs for the gene, association scores, mechanistic context 2 Lead molecule retrieval ChEMBL REST API Compounds with SMILES, molecular weight, LogP, PSA, HBD/HBA counts 3 Oral bioavailability scoring Lipinski Rule-of-Five algorithm Ro5 compliance score per lead molecule 4 Composite lead ranking Phase + binding evidence + Ro5 Ranked lead list, scores clamped to [0,1] 5 Mutation-specific structure AlphaFold Server 3D protein structure (.cif → .pdb) for the specific variant 6 Discovery brief assembly All above sources Structured brief saved to DrugRequest job in PostgreSQL 7 Marketplace dispatch OpenOncology pharma marketplace Manufacturers notified; bidding opens on custom synthesis The discovery brief contains: the target gene and disease context, ranked lead molecules with full molecular property data (MW, LogP, PSA, HBD/HBA), oral-exposure scores, clinical phase of each compound, scaffold and fragment notes for medicinal chemistry handoff, and the mutation-specific AlphaFold structure for three-dimensional docking context. Evidence sources and matched query terms from OpenTargets and ChEMBL are surfaced directly in the brief so oncologists and manufacturers can audit the prioritisation logic. Critically, the discovery brief is not a drug recommendation it is a structured research brief framing the next steps for drug development. Every lead molecule in the brief is a real, named compound with a real ChEMBL or OpenTargets record. No compound names or properties are generated or hallucinated. If ChEMBL and OpenTargets return no cancer-matched lead for the target and disease context, the brief explicitly reports this as an integration gap rather than substituting placeholder content. The pharma marketplace allows accredited pharmaceutical manufacturers to review the discovery brief and submit competitive bids for custom synthesis of lead molecules or analogs; the marketplace module is implemented at the platform level, and full production pharma onboarding requires independent regulatory and compliance review. An integrated crowdfunding module with Stripe Elements enables patients to raise resources toward synthesis costs, with milestone notifications at 25%, 50%, 75%, and 100% of target. 2.4 Validation Design Three independent validation approaches assessed the stage one drug ranking algorithm. Stage two validation is necessarily limited to structural and coverage analysis given the absence of ground-truth custom synthesis outcomes at this stage of platform maturity. All validation cases are literature derived or drawn from public databases (OncoKB, cBioPortal/TCGA); no prospective clinical cohort was used. Blinded oncologist holdout (n = 50) 50 clinically annotated cases were assembled from published molecular tumor board case reports and OncoKB clinical examples, expanded from an initial 24-case set by adding 30 literature-sourced cases from JCO Precision Oncology, Annals of Oncology, and Nature Medicine. The expanded set includes 12 Level 1–2 literature cases, 12 Level 3–4 cases, and 6 negative controls (expect_empty=True). Cases requiring live OncoKB credentials were flagged as network_dependent and excluded from offline runs. The pipeline was run without access to gold-standard labels and evaluated on Hit@3, Standard Precision@3, Normalised Precision@3, Mean Reciprocal Rank, and false positive rate. TCGA 100-patient benchmark : 100 real de-identified patient profiles were retrieved from cBioPortal TCGA cohorts. Pipeline outputs were classified by tier: Tier 1 (FDA-approved direct match), Tier 2 (repurposing candidate), or Tier 3/4 (custom discovery escalation). Stage two escalation was triggered and brief assembly was verified for all Tier 3/4 cases. TCGA 200-patient benchmark 200-patient set constructed with deliberate overrepresentation of rare variants and variants with no approved therapy, to stress-test escalation logic and confirm zero empty outputs. All validation scripts are in the repository and reproducible from public data. 2.5 Oncologist Concordance Benchmark Concordance analysis against multi-cohort TCGA clinical records (n = 1,713 label cases) assessed alignment between OpenOncology stage one recommendations and real oncologist-selected therapies. Metrics: exact match and equivalence-adjusted match at Top-1 and Top-3, plus Jaccard similarity. The 97.9% no-prediction rate reflects TCGA’s unselected population composition most patients carry non-actionable variants or received cytotoxic chemotherapy and is expected rather than pathological. 3. Results 3.1 Blinded Oncologist Holdout (n = 50) Metric Result Interpretation Hit@3 0.900 Gold drug in top-3 for 90% of cases Standard Precision@3 0.508 (ceiling: 0.625) Within the expected range for a harder mixed-difficulty holdout Normalised Precision@3 0.817 Strong performance adjusting for single-drug gold standards Mean Reciprocal Rank 0.883 Gold drug consistently near top of ranked list False positive rate 0.000 Zero non-approved or fabricated drugs in any top-3 output Hit@3 = 0.900 indicates the gold-standard drug appeared in the top-3 for 45 of 50 cases. False positive rate of zero confirms the evidence-only constraint is effective: the pipeline cannot produce a drug recommendation without a corresponding OncoKB, ClinVar, or CIViC record. The expanded 50-case holdout included 12 Level 3–4 literature cases and 6 negative controls, representing a deliberately harder validation set than smaller prior holdouts; the metric reduction relative to earlier 24-case results reflects increased case difficulty rather than pipeline regression. Standard Precision@3 of 0.508 against a ceiling of 0.625 indicates strong performance given the mixed-difficulty case composition. 3.2 TCGA Benchmarks and Stage Two Coverage Cohort Tier 1 (FDA-approved) Tier 2 (Repurposing) Stage 2 escalation Empty outputs 100-patient 36 (36%) 64 (64%) 0 (all stage 1 matched) 0 (100% covered) 200-patient 15 (7.5%) 0 185 (92.5%) 0 (100% covered) Both benchmarks achieved 100% coverage. In the 200-patient cohort deliberately constructed to include rare variants and non-hotspot mutations 92.5% of patients had no approved drug match and were escalated to stage two. In every case, the custom drug discovery worker successfully assembled a discovery brief. Zero patients received an empty output. This confirms that the escalation pathway functions as designed under adversarial conditions. The lower Tier 1 rate in the 200-patient cohort (7.5% vs 36%) validates the benchmark design: the expanded set correctly represents the real distribution of precision oncology cases, where direct FDA-approved matches are the exception rather than the rule. Stage two escalation is not a fallback — it is the primary pathway for the majority of precision oncology patients. 3.3 Oncologist Concordance Match type Top-1 Top-3 Jaccard Exact match (strict) 27.78% (10/36) 50.0% (18/36) 0.1887 Equivalence-adjusted 100% (36/36) 100% (36/36) 0.5804 Equivalence-adjusted concordance of 100% at both Top-1 and Top-3 indicates that for every actionable case, OpenOncology and the oncologist selected drugs from the same therapeutic class or mechanism. The gap between exact and equivalence-adjusted rates reflects cases where multiple agents within a class are clinically acceptable a pattern consistent with real tumor board practice where drug selection within a class depends on availability, toxicity profile, and patient factors outside the scope of variant-based ranking. 3.4 Stage Two: Discovery Brief Quality Across all 185 stage-two escalations in the 200-patient benchmark, the custom_drug_worker successfully retrieved ChEMBL lead molecules and OpenTargets association data for the target gene in every case. Lead molecule counts per brief ranged from 3 to 47 depending on target druggability. Ro5 oral-exposure scoring was applied to all leads, with a median of 68% of leads achieving full Ro5 compliance across the benchmark cohort. AlphaFold structure generation was triggered for all 185 cases and completed successfully for 178 (96.2%); the seven failures were attributable to AlphaFold Server rate limiting and were logged as pending rather than errors. Discovery brief quality was evaluated structurally — verifying presence of all required fields (target context, ranked leads, molecular properties, Ro5 scores, evidence sources, matched terms) — rather than clinically, as ground-truth custom synthesis outcomes are not available. No placeholder content, fabricated compound names, or missing evidence attributions were observed in any brief. 4. Discussion 4.1 The Empty-Output Problem and Why It Matters The most important finding from the TCGA benchmarks is not the stage one validation metrics — those confirm that the drug ranking algorithm works correctly for known actionable variants. The most important finding is that 92.5% of patients in the adversarial 200-patient cohort had no approved drug match, and that every single one of them received a structured, evidence-sourced discovery brief rather than an empty output. This matters because the existing precision oncology ecosystem has optimised almost entirely for the 7–36% of patients with a direct approved drug match. For the majority of patients — those with rare variants, variants in understudied genes, or variants in well-studied genes without approved therapies — current tools have no answer. They return nothing. The patient and oncologist are no better off for having done the molecular analysis. OpenOncology takes the position that returning nothing is not acceptable when an actionable next step exists. For every patient with a specific mutation and a named target gene, there is a body of ChEMBL and OpenTargets evidence that can be structured into a discovery brief. That brief may not result in a treatment it requires drug development timelines and resources far beyond the scope of this platform but it provides a concrete next step, a named set of lead compounds, and a structured interface for pharmaceutical manufacturers to engage with the patient’s case. That is categorically different from an empty output. 4.2 Comparison with Existing Open-Source Tools OpenOncology occupies a distinct niche relative to every existing precision oncology platform, open-source or commercial. cBioPortal [ 4 ] provides comprehensive multi-omics visualisation and cohort analysis but performs no drug ranking and has no patient-facing recommendation output. OncoKB [ 5 ] provides the most rigorous clinical evidence curation for known actionable variants but requires institutional registration, performs no variant calling or structural analysis, and has no escalation pathway for unannotated variants. Commercial platforms including Foundation Medicine CDx and Tempus xT provide end-to-end clinical reporting but require proprietary laboratory infrastructure, carry significant per-test costs, and publish no open validation methodology. The critical distinction is the custom drug discovery pathway. No existing open-source precision oncology tool and to the authors’ knowledge, no commercial tool provides an automated escalation from negative drug matching to a manufacturer-ready de novo discovery brief. The closest analogues are drug repurposing databases such as DGIdb [ 12 ] and PharmGKB [13], which provide static gene drug interaction tables but no dynamic escalation, no lead molecule scoring, no structural generation, and no marketplace integration. It is important to distinguish this custom discovery pathway from AI-based de novo molecular generation (generative chemistry). OpenOncology does not generate new molecular structures. It retrieves and ranks existing compounds from ChEMBL and OpenTargets against a specific target. This is a deliberate design choice: generative chemistry outputs require extensive experimental validation before any confidence in their properties is justified, and presenting a generative compound to a patient as a drug candidate would be scientifically irresponsible. Every lead molecule in an OpenOncology discovery brief is a real, named compound with a real experimental record. 4.3 Limitations The 50-case blinded holdout, while expanded from an initial 24-case set by incorporating 30 literature-sourced cases from peer-reviewed molecular tumor board reports, remains a moderate-sized validation set. The deliberate inclusion of Level 3–4 and negative control cases increases difficulty and ecological validity, but prospective validation on a larger, independently adjudicated cohort is needed before clinical deployment of stage one recommendations. Stage two validation is structural rather than clinical. We can verify that discovery briefs are correctly assembled and contain real compounds with real evidence, but we cannot yet verify that the ranked lead molecules are the most appropriate starting points for drug development against the specific mutation. That validation requires experimental binding assays and medicinal chemistry expertise that are outside the scope of an independent open-source project. The platform currently lacks access to institutional genomics datasets (dbGaP, ICGC, EHR-linked cohorts). This limits variant diversity in the validation data and may reduce performance on rare variants or underrepresented ancestries. The 97.9% no-prediction rate in the concordance benchmark reflects TCGA composition rather than pipeline failure, but the actionable validation set is consequently limited to 36 cases. OpenOncology has not been submitted for regulatory clearance as a clinical decision support tool. The platform is designed to surface evidence for oncologist review, not to replace it. No treatment decision should be made on the basis of OpenOncology output alone. Discovery briefs from stage two represent early-stage research leads not clinical candidates and any compound emerging from this pathway requires full preclinical and clinical development before patient use. 4.4 Future Directions The highest-priority development goal is institutional partnership for EHR-linked validation. A single collaborating cancer center would enable prospective validation of stage one recommendations against real tumor board decisions, and would provide access to the diverse variant profiles needed to evaluate stage two brief quality against experimental outcomes. The multi-omics roadmap(Phase-6) extends the platform to RNA-seq expression data and methylation profiling, enabling tumour microenvironment-aware drug ranking. Federated learning architecture is planned to allow institutional partners to contribute to model improvement without sharing raw patient data. On the custom discovery side, integration of AlphaFold-based binding pocket prediction and AutoDock Vina re-scoring for ChEMBL leads would strengthen the structural basis of stage two rankings. Collaboration with medicinal chemists to evaluate the scaffold/fragment notes in real discovery briefs would provide the first experimental signal on stage two clinical utility. 5. Conclusions We have presented OpenOncology, an open-source precision oncology platform with two distinct contributions. The first a validated, evidence-only drug ranking algorithm achieving Hit@3 = 0.900 and zero false positives on a blinded 50-case holdout that includes deliberately hard Level 3–4 literature cases and negative controls demonstrates that rigorous clinical drug matching can be performed without institutional infrastructure or cost barriers.The second a fully automated custom drug discovery escalation pathway triggered when no approved therapy matches addresses a problem that existing precision oncology tools do not attempt to solve: the majority of patients whose mutations have no approved drug. By ensuring that every patient receives an actionable output either a ranked drug recommendation or a manufacturer-ready discovery brief OpenOncology reframes what a precision oncology platform should be responsible for. The question is not only which approved drug matches your mutation. For most patients, the more important question is: given that no approved drug matches, what happens next? OpenOncology is the first open-source platform to provide a structured, evidence-sourced, non-empty answer to that question. All code, data, validation artifacts, and benchmark scripts are available at github.com/immortal71/openoncology under the MIT licence. Declarations Competing Interests The author declares no competing interests. OpenOncology is developed as a non-commercial open-source project with no financial relationship with any pharmaceutical company, diagnostic laboratory, or clinical software vendor. Acknowledgements The author thanks the contributors to the OpenOncology repository and the open-source communities maintaining OncoKB, cBioPortal, OpenTargets, ChEMBL, DiffDock, AlphaMissense, and AlphaFold. This work was conducted without external funding. Data Availability All benchmark scripts and validation artifacts are publicly available at github.com/immortal71/openoncology . Benchmark JSON files (real_patient_benchmark_100.json, real_patient_benchmark_200.json) and concordance labels (scripts/concordance_labels.json) are included in the repository and derived from publicly accessible cBioPortal/TCGA data. The 50-case holdout artifactsare available at validation_results/holdout_50_results.txt and validation_results/holdout_50_metrics.json. No proprietary or restricted data were used. References World Health Organization (2024) Cancer. Global Health Estimates 2024. WHO, Geneva Druker BJ et al (2001) Efficacy and safety of a specific inhibitor of the BCR-ABL tyrosine kinase in chronic myeloid leukemia. N Engl J Med 344(14):1031–1037 Soria JC et al (2018) Osimertinib in untreated EGFR-mutated advanced non-small-cell lung cancer. N Engl J Med 378(2):113–125 Cerami E et al (2012) The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov 2(5):401–404 Chakravarty D et al (2017) OncoKB: A precision oncology knowledge base. JCO Precis Oncol. ;1:PO.17.00011. Cheng J et al (2023) Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381(6664):eadg7492 Jumper J et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature 596(7873):583–589 Corso G et al (2023) DiffDock: Diffusion steps, twists, and turns for molecular docking. ICLR. arXiv:2210.01776 Ochoa D et al (2023) The next-generation Open Targets Platform: reimagined, redesigned, rebuilt. Nucleic Acids Res 51(D1):D1353–D1359 Mendez D et al (2019) ChEMBL: towards direct deposition of bioassay data. Nucleic Acids Res 47(D1):D930–D940 Freshour SL et al (2021) Integration of the Drug-Gene Interaction Database (DGIdb 4.0) with open crowdsource efforts. Nucleic Acids Res 49(D1):D1144–D1151 Whirl-Carrillo M et al (2021) An Evidence-Based Framework for Evaluating Pharmacogenomics Knowledge for Personalized Medicine. Clin Pharmacol Ther 110(3):563–572 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9707913","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":640657814,"identity":"b568b3fc-e933-4120-be1a-eec1b55734de","order_by":0,"name":"Aashish Kharel","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABA0lEQVRIie2QMUsDMRTHcwjXRXCNi36FlINXhOP6QVxeOMgt9Rt0iEu7+FUEJ+dAQJdq1kAd7pa4ONStwyG+6yIU7lo3wfyGEML/x3v/MBaJ/E1ShswI0V2br5zO5NYcr2CqOkUfVtiPYndPg8rZ8jnUdfuWTUYvTY2nrrhfWpoyz6/7FL6aTYRcBLi6qzKBfF0+riQpT+pG9zlmlnKpbS6MSjmKdQmGlETbXuXSvQeOLSkujLaIryW4ZlgRHoFTaxBe0dcZU4A/MGXsP4DLhc2EDye0YYngaQoOdLlwVTjftnb84FSy+dTFFFzV1Jt53l9/H7lL4rHxjulvwpFIJPI/+Aa6DGspkBnilgAAAABJRU5ErkJggg==","orcid":"","institution":"Independent Researcher","correspondingAuthor":true,"prefix":"","firstName":"Aashish","middleName":"","lastName":"Kharel","suffix":""}],"badges":[],"createdAt":"2026-05-13 21:27:21","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9707913/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9707913/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109759613,"identity":"68d70727-b467-43f3-91fd-56828a1cb5b1","added_by":"auto","created_at":"2026-05-22 07:27:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":183802,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9707913/v1/7b8000dd-82c0-4d0c-a4a9-66749fb68312.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eOpenOncology: An Open-Source Framework for Evidence-Based Drug Matching and De Novo Custom Drug Discovery in Precision Oncology\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCancer is the second leading cause of death globally, responsible for approximately 10\u0026nbsp;million deaths annually [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The emergence of precision oncology matching therapies to the molecular profile of individual tumors has demonstrated substantial survival benefits across multiple cancer types. Landmark approvals including imatinib for BCR-ABL-positive chronic myeloid leukemia, trastuzumab for HER2-amplified breast cancer, and osimertinib for EGFR-mutant non-small cell lung cancer have established the paradigm [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHowever, precision oncology has two compounding failure modes that limit its reach. The first is infrastructural: the tools that support molecular tumor board decisions cBioPortal, OncoKB, and commercial variant interpretation systems require institutional access, electronic health record integration, or subscription licensing [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. For patients in low- and middle-income countries (LMICs), and for uninsured patients in high-income settings, this infrastructure is functionally unavailable.\u003c/p\u003e \u003cp\u003eThe second failure mode is more fundamental, and has received less attention: even when variant interpretation infrastructure is available, the majority of patients with rare or non-hotspot mutations receive an empty output. There is no approved drug. There is no guideline-supported repurposing candidate. Existing tools commercial or open-source have no response to this situation beyond a negative result. The patient, the oncologist, and the tumor board are left with nothing actionable.\u003c/p\u003e \u003cp\u003eThis second failure affects more patients than is commonly acknowledged. TCGA cohort data suggests that the fraction of patients with a direct FDA-approved targeted therapy match ranges from 7\u0026ndash;36% depending on tumor type and variant selection criteria (see Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e3\u003c/span\u003e). For the remaining 64\u0026ndash;93% of patients, current precision oncology infrastructure provides no therapeutic pathway beyond generic cytotoxic chemotherapy.\u003c/p\u003e \u003cp\u003eWe present OpenOncology, an open-source platform designed to close both gaps simultaneously. For the infrastructure gap: the platform is entirely free, open-source, containerised, and deployable without institutional affiliation. For the empty-output gap: the platform implements a novel two-stage architecture in which stage one (approved drug matching and repurposing) automatically escalates to stage two (de novo custom drug discovery) when no match is found ensuring that every patient receives an actionable output regardless of how rare or poorly characterised their mutation is.\u003c/p\u003e \u003cp\u003eThe custom drug discovery stage is the platform\u0026rsquo;s most significant technical contribution. When stage one finds no approved or repurposed match, an asynchronous worker queries ChEMBL and OpenTargets for lead molecules targeting the patient\u0026rsquo;s gene, applies Lipinski Rule-of-Five oral bioavailability scoring, generates a mutation-specific protein structure via AlphaFold Server, and packages the ranked lead molecules and structural data into a discovery brief that pharmaceutical manufacturers can receive and bid on through an integrated marketplace. This pathway does not fabricate drug recommendations. It generates a structured, evidence-sourced research brief that frames the next step for drug development.\u003c/p\u003e \u003cp\u003eThis paper describes the platform architecture, the two-stage drug matching and discovery methodology, and validation results from a blinded oncologist holdout and real-patient TCGA benchmarks. We discuss current limitations, the data access constraints inherent to independent research, and the roadmap toward institutional validation.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 System Architecture\u003c/h2\u003e \u003cp\u003eOpenOncology is implemented as a containerised microservices stack. The backend is built on FastAPI with SQLAlchemy 2 async ORM and PostgreSQL 16. A Celery worker queue backed by Redis handles asynchronous genomics pipeline jobs, AI scoring tasks, and custom drug discovery workflows as independent, horizontally scalable workers. Object storage uses MinIO with AES-256 encryption. Authentication and role-based access control (patient, oncologist, administrator) are managed through Keycloak OIDC/OAuth2. The frontend is Next.js 14 with TypeScript. The full stack deploys via Docker Compose for local development or Kubernetes/Helm for production, with Prometheus and Grafana instrumentation.\u003c/p\u003e \u003cp\u003ePHI handling is designed to conform to HIPAA \u0026sect;\u0026nbsp;164.308, \u0026sect;\u0026nbsp;164.310, and \u0026sect;\u0026nbsp;164.312 safeguards, with GDPR Article 17 (erasure) and Article 20 (data portability) implemented at the API layer. Production deployment handling real patient data requires independent security audit and regulatory review before clinical use. Weekly automated security scans cover Python dependencies (pip-audit), Node.js dependencies (npm audit), static analysis (Bandit, Semgrep OWASP), API security (ZAP baseline), and container vulnerabilities (Trivy).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Stage One: Variant Calling and Drug Matching\u003c/h2\u003e \u003cp\u003eRaw sequencing input (FASTQ or BAM) is processed through a Nextflow pipeline comprising quality control (FastQC, Trimmomatic), alignment to GRCh38 (BWA-MEM2), duplicate marking and base quality score recalibration (GATK), and somatic variant calling (GATK Mutect2). This workflow follows clinical-grade bioinformatics conventions; computational requirements for the full pipeline including AlphaFold and DiffDock impose practical constraints on throughput outside HPC environments. Variants are annotated using OpenCRAVAT with GRCh38 reference databases. Preflight checks validate input integrity before pipeline dispatch.\u003c/p\u003e \u003cp\u003eEach missense variant is scored by AlphaMissense [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], a deep learning pathogenicity model queried against a 3.6 GB pre-built SQLite index. High-pathogenicity variants (score\u0026thinsp;\u0026gt;\u0026thinsp;0.7) in genes without a direct OncoKB match trigger AlphaFold Server submission to generate a mutation-specific protein structure for downstream docking.\u003c/p\u003e \u003cp\u003eDrug candidates are ranked by a weighted composite score integrating five evidence signals:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSignal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWeight\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSource\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiffDock binding confidence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMolecular docking against mutation-specific structure\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOpenTargets association score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti-evidence target-disease linkage\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOncoKB actionability level\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFDA-approved clinical evidence tier (Levels 1\u0026ndash;2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlphaMissense pathogenicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStructural impact of the specific variant\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical trial phase\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTranslational readiness (OpenTargets / ChEMBL)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eResistance gating is applied before ranking. Known resistance mutations \u0026mdash; including EGFR T790M (blocks erlotinib/gefitinib), KRAS G12C resistance alleles, and BRAF splice variants \u0026mdash; are maintained in a curated resistance table. Any drug flagged by an applicable resistance gate is excluded from the output regardless of composite score.\u003c/p\u003e \u003cp\u003eAll drugs in the output are drawn exclusively from OncoKB Levels 1\u0026ndash;2, ClinVar, or CIViC evidence items no drug names are inferred, generated, or hallucinated. If stage one returns candidates above the minimum evidence threshold, the top-three ranked drugs are returned to the patient dashboard with full provenance. The pipeline terminates. If stage one returns no candidates above threshold, the pipeline automatically escalates to stage two.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Stage Two: De Novo Custom Drug Discovery\u003c/h2\u003e \u003cp\u003eStage two is the core novel contribution of this platform. It is the first open-source implementation of a fully automated escalation from clinical drug matching to de novo discovery brief generation for patients with no approved therapeutic option.\u003c/p\u003e \u003cp\u003eWhen stage one finds no match, an asynchronous Celery worker (custom_drug_worker) is triggered with the patient\u0026rsquo;s target gene, variant, and disease context. The worker executes the following sequence:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStep\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAction\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eData Source\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOutput\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTarget-disease evidence retrieval\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOpenTargets GraphQL API\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eApproved/clinical-stage drugs for the gene, association scores, mechanistic context\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLead molecule retrieval\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eChEMBL REST API\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCompounds with SMILES, molecular weight, LogP, PSA, HBD/HBA counts\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOral bioavailability scoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLipinski Rule-of-Five algorithm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRo5 compliance score per lead molecule\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eComposite lead ranking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePhase\u0026thinsp;+\u0026thinsp;binding evidence\u0026thinsp;+\u0026thinsp;Ro5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRanked lead list, scores clamped to [0,1]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMutation-specific structure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAlphaFold Server\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3D protein structure (.cif \u0026rarr; .pdb) for the specific variant\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDiscovery brief assembly\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAll above sources\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStructured brief saved to DrugRequest job in PostgreSQL\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMarketplace dispatch\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOpenOncology pharma marketplace\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eManufacturers notified; bidding opens on custom synthesis\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe discovery brief contains: the target gene and disease context, ranked lead molecules with full molecular property data (MW, LogP, PSA, HBD/HBA), oral-exposure scores, clinical phase of each compound, scaffold and fragment notes for medicinal chemistry handoff, and the mutation-specific AlphaFold structure for three-dimensional docking context. Evidence sources and matched query terms from OpenTargets and ChEMBL are surfaced directly in the brief so oncologists and manufacturers can audit the prioritisation logic.\u003c/p\u003e \u003cp\u003eCritically, the discovery brief is not a drug recommendation it is a structured research brief framing the next steps for drug development. Every lead molecule in the brief is a real, named compound with a real ChEMBL or OpenTargets record. No compound names or properties are generated or hallucinated. If ChEMBL and OpenTargets return no cancer-matched lead for the target and disease context, the brief explicitly reports this as an integration gap rather than substituting placeholder content.\u003c/p\u003e \u003cp\u003eThe pharma marketplace allows accredited pharmaceutical manufacturers to review the discovery brief and submit competitive bids for custom synthesis of lead molecules or analogs; the marketplace module is implemented at the platform level, and full production pharma onboarding requires independent regulatory and compliance review. An integrated crowdfunding module with Stripe Elements enables patients to raise resources toward synthesis costs, with milestone notifications at 25%, 50%, 75%, and 100% of target.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Validation Design\u003c/h2\u003e \u003cp\u003eThree independent validation approaches assessed the stage one drug ranking algorithm. Stage two validation is necessarily limited to structural and coverage analysis given the absence of ground-truth custom synthesis outcomes at this stage of platform maturity. All validation cases are literature derived or drawn from public databases (OncoKB, cBioPortal/TCGA); no prospective clinical cohort was used.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eBlinded oncologist holdout (n\u0026thinsp;=\u0026thinsp;50)\u003c/strong\u003e \u003cp\u003e50 clinically annotated cases were assembled from published molecular tumor board case reports and OncoKB clinical examples, expanded from an initial 24-case set by adding 30 literature-sourced cases from JCO Precision Oncology, Annals of Oncology, and Nature Medicine. The expanded set includes 12 Level 1\u0026ndash;2 literature cases, 12 Level 3\u0026ndash;4 cases, and 6 negative controls (expect_empty=True). Cases requiring live OncoKB credentials were flagged as network_dependent and excluded from offline runs. The pipeline was run without access to gold-standard labels and evaluated on Hit@3, Standard Precision@3, Normalised Precision@3, Mean Reciprocal Rank, and false positive rate.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTCGA 100-patient benchmark\u003c/b\u003e: 100 real de-identified patient profiles were retrieved from cBioPortal TCGA cohorts. Pipeline outputs were classified by tier: Tier 1 (FDA-approved direct match), Tier 2 (repurposing candidate), or Tier 3/4 (custom discovery escalation). Stage two escalation was triggered and brief assembly was verified for all Tier 3/4 cases.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eTCGA 200-patient benchmark\u003c/strong\u003e \u003cp\u003e200-patient set constructed with deliberate overrepresentation of rare variants and variants with no approved therapy, to stress-test escalation logic and confirm zero empty outputs. All validation scripts are in the repository and reproducible from public data.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Oncologist Concordance Benchmark\u003c/h2\u003e \u003cp\u003eConcordance analysis against multi-cohort TCGA clinical records (n\u0026thinsp;=\u0026thinsp;1,713 label cases) assessed alignment between OpenOncology stage one recommendations and real oncologist-selected therapies. Metrics: exact match and equivalence-adjusted match at Top-1 and Top-3, plus Jaccard similarity. The 97.9% no-prediction rate reflects TCGA\u0026rsquo;s unselected population composition most patients carry non-actionable variants or received cytotoxic chemotherapy and is expected rather than pathological.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Blinded Oncologist Holdout (n\u0026thinsp;=\u0026thinsp;50)\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResult\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInterpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHit@3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGold drug in top-3 for 90% of cases\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStandard Precision@3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.508 (ceiling: 0.625)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eWithin the expected range for a harder mixed-difficulty holdout\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormalised Precision@3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.817\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStrong performance adjusting for single-drug gold standards\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Reciprocal Rank\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGold drug consistently near top of ranked list\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse positive rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eZero non-approved or fabricated drugs in any top-3 output\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eHit@3\u0026thinsp;=\u0026thinsp;0.900 indicates the gold-standard drug appeared in the top-3 for 45 of 50 cases. False positive rate of zero confirms the evidence-only constraint is effective: the pipeline cannot produce a drug recommendation without a corresponding OncoKB, ClinVar, or CIViC record. The expanded 50-case holdout included 12 Level 3\u0026ndash;4 literature cases and 6 negative controls, representing a deliberately harder validation set than smaller prior holdouts; the metric reduction relative to earlier 24-case results reflects increased case difficulty rather than pipeline regression. Standard Precision@3 of 0.508 against a ceiling of 0.625 indicates strong performance given the mixed-difficulty case composition.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 TCGA Benchmarks and Stage Two Coverage\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCohort\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTier 1 (FDA-approved)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTier 2 (Repurposing)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStage 2 escalation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEmpty outputs\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e100-patient\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (36%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e64 (64%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (all stage 1 matched)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0 (100% covered)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e200-patient\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 (7.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e185 (92.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0 (100% covered)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eBoth benchmarks achieved 100% coverage. In the 200-patient cohort deliberately constructed to include rare variants and non-hotspot mutations 92.5% of patients had no approved drug match and were escalated to stage two. In every case, the custom drug discovery worker successfully assembled a discovery brief. Zero patients received an empty output. This confirms that the escalation pathway functions as designed under adversarial conditions.\u003c/p\u003e \u003cp\u003eThe lower Tier 1 rate in the 200-patient cohort (7.5% vs 36%) validates the benchmark design: the expanded set correctly represents the real distribution of precision oncology cases, where direct FDA-approved matches are the exception rather than the rule. Stage two escalation is not a fallback \u0026mdash; it is the primary pathway for the majority of precision oncology patients.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Oncologist Concordance\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabe\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMatch type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTop-1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTop-3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eJaccard\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExact match (strict)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e27.78% (10/36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50.0% (18/36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1887\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEquivalence-adjusted\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e100% (36/36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100% (36/36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5804\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eEquivalence-adjusted concordance of 100% at both Top-1 and Top-3 indicates that for every actionable case, OpenOncology and the oncologist selected drugs from the same therapeutic class or mechanism. The gap between exact and equivalence-adjusted rates reflects cases where multiple agents within a class are clinically acceptable a pattern consistent with real tumor board practice where drug selection within a class depends on availability, toxicity profile, and patient factors outside the scope of variant-based ranking.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Stage Two: Discovery Brief Quality\u003c/h2\u003e \u003cp\u003eAcross all 185 stage-two escalations in the 200-patient benchmark, the custom_drug_worker successfully retrieved ChEMBL lead molecules and OpenTargets association data for the target gene in every case. Lead molecule counts per brief ranged from 3 to 47 depending on target druggability. Ro5 oral-exposure scoring was applied to all leads, with a median of 68% of leads achieving full Ro5 compliance across the benchmark cohort. AlphaFold structure generation was triggered for all 185 cases and completed successfully for 178 (96.2%); the seven failures were attributable to AlphaFold Server rate limiting and were logged as pending rather than errors.\u003c/p\u003e \u003cp\u003eDiscovery brief quality was evaluated structurally \u0026mdash; verifying presence of all required fields (target context, ranked leads, molecular properties, Ro5 scores, evidence sources, matched terms) \u0026mdash; rather than clinically, as ground-truth custom synthesis outcomes are not available. No placeholder content, fabricated compound names, or missing evidence attributions were observed in any brief.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.1 The Empty-Output Problem and Why It Matters\u003c/h2\u003e \u003cp\u003eThe most important finding from the TCGA benchmarks is not the stage one validation metrics \u0026mdash; those confirm that the drug ranking algorithm works correctly for known actionable variants. The most important finding is that 92.5% of patients in the adversarial 200-patient cohort had no approved drug match, and that every single one of them received a structured, evidence-sourced discovery brief rather than an empty output.\u003c/p\u003e \u003cp\u003eThis matters because the existing precision oncology ecosystem has optimised almost entirely for the 7\u0026ndash;36% of patients with a direct approved drug match. For the majority of patients \u0026mdash; those with rare variants, variants in understudied genes, or variants in well-studied genes without approved therapies \u0026mdash; current tools have no answer. They return nothing. The patient and oncologist are no better off for having done the molecular analysis.\u003c/p\u003e \u003cp\u003eOpenOncology takes the position that returning nothing is not acceptable when an actionable next step exists. For every patient with a specific mutation and a named target gene, there is a body of ChEMBL and OpenTargets evidence that can be structured into a discovery brief. That brief may not result in a treatment it requires drug development timelines and resources far beyond the scope of this platform but it provides a concrete next step, a named set of lead compounds, and a structured interface for pharmaceutical manufacturers to engage with the patient\u0026rsquo;s case. That is categorically different from an empty output.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Comparison with Existing Open-Source Tools\u003c/h2\u003e \u003cp\u003eOpenOncology occupies a distinct niche relative to every existing precision oncology platform, open-source or commercial.\u003c/p\u003e \u003cp\u003ecBioPortal [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] provides comprehensive multi-omics visualisation and cohort analysis but performs no drug ranking and has no patient-facing recommendation output. OncoKB [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] provides the most rigorous clinical evidence curation for known actionable variants but requires institutional registration, performs no variant calling or structural analysis, and has no escalation pathway for unannotated variants. Commercial platforms including Foundation Medicine CDx and Tempus xT provide end-to-end clinical reporting but require proprietary laboratory infrastructure, carry significant per-test costs, and publish no open validation methodology.\u003c/p\u003e \u003cp\u003eThe critical distinction is the custom drug discovery pathway. No existing open-source precision oncology tool and to the authors\u0026rsquo; knowledge, no commercial tool provides an automated escalation from negative drug matching to a manufacturer-ready de novo discovery brief. The closest analogues are drug repurposing databases such as DGIdb [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] and PharmGKB [13], which provide static gene drug interaction tables but no dynamic escalation, no lead molecule scoring, no structural generation, and no marketplace integration.\u003c/p\u003e \u003cp\u003eIt is important to distinguish this custom discovery pathway from AI-based de novo molecular generation (generative chemistry). OpenOncology does not generate new molecular structures. It retrieves and ranks existing compounds from ChEMBL and OpenTargets against a specific target. This is a deliberate design choice: generative chemistry outputs require extensive experimental validation before any confidence in their properties is justified, and presenting a generative compound to a patient as a drug candidate would be scientifically irresponsible. Every lead molecule in an OpenOncology discovery brief is a real, named compound with a real experimental record.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Limitations\u003c/h2\u003e \u003cp\u003eThe 50-case blinded holdout, while expanded from an initial 24-case set by incorporating 30 literature-sourced cases from peer-reviewed molecular tumor board reports, remains a moderate-sized validation set. The deliberate inclusion of Level 3\u0026ndash;4 and negative control cases increases difficulty and ecological validity, but prospective validation on a larger, independently adjudicated cohort is needed before clinical deployment of stage one recommendations.\u003c/p\u003e \u003cp\u003eStage two validation is structural rather than clinical. We can verify that discovery briefs are correctly assembled and contain real compounds with real evidence, but we cannot yet verify that the ranked lead molecules are the most appropriate starting points for drug development against the specific mutation. That validation requires experimental binding assays and medicinal chemistry expertise that are outside the scope of an independent open-source project.\u003c/p\u003e \u003cp\u003eThe platform currently lacks access to institutional genomics datasets (dbGaP, ICGC, EHR-linked cohorts). This limits variant diversity in the validation data and may reduce performance on rare variants or underrepresented ancestries. The 97.9% no-prediction rate in the concordance benchmark reflects TCGA composition rather than pipeline failure, but the actionable validation set is consequently limited to 36 cases.\u003c/p\u003e \u003cp\u003eOpenOncology has not been submitted for regulatory clearance as a clinical decision support tool. The platform is designed to surface evidence for oncologist review, not to replace it. No treatment decision should be made on the basis of OpenOncology output alone. Discovery briefs from stage two represent early-stage research leads not clinical candidates and any compound emerging from this pathway requires full preclinical and clinical development before patient use.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Future Directions\u003c/h2\u003e \u003cp\u003eThe highest-priority development goal is institutional partnership for EHR-linked validation. A single collaborating cancer center would enable prospective validation of stage one recommendations against real tumor board decisions, and would provide access to the diverse variant profiles needed to evaluate stage two brief quality against experimental outcomes.\u003c/p\u003e \u003cp\u003eThe multi-omics roadmap(Phase-6) extends the platform to RNA-seq expression data and methylation profiling, enabling tumour microenvironment-aware drug ranking. Federated learning architecture is planned to allow institutional partners to contribute to model improvement without sharing raw patient data.\u003c/p\u003e \u003cp\u003eOn the custom discovery side, integration of AlphaFold-based binding pocket prediction and AutoDock Vina re-scoring for ChEMBL leads would strengthen the structural basis of stage two rankings. Collaboration with medicinal chemists to evaluate the scaffold/fragment notes in real discovery briefs would provide the first experimental signal on stage two clinical utility.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eWe have presented OpenOncology, an open-source precision oncology platform with two distinct contributions. The first a validated, evidence-only drug ranking algorithm achieving Hit@3\u0026thinsp;=\u0026thinsp;0.900 and zero false positives on a blinded 50-case holdout that includes deliberately hard Level 3\u0026ndash;4 literature cases and negative controls demonstrates that rigorous clinical drug matching can be performed without institutional infrastructure or cost barriers.The second a fully automated custom drug discovery escalation pathway triggered when no approved therapy matches addresses a problem that existing precision oncology tools do not attempt to solve: the majority of patients whose mutations have no approved drug.\u003c/p\u003e \u003cp\u003eBy ensuring that every patient receives an actionable output either a ranked drug recommendation or a manufacturer-ready discovery brief OpenOncology reframes what a precision oncology platform should be responsible for. The question is not only which approved drug matches your mutation. For most patients, the more important question is: given that no approved drug matches, what happens next? OpenOncology is the first open-source platform to provide a structured, evidence-sourced, non-empty answer to that question.\u003c/p\u003e \u003cp\u003eAll code, data, validation artifacts, and benchmark scripts are available at github.com/immortal71/openoncology under the MIT licence.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe author declares no competing interests. OpenOncology is developed as a non-commercial open-source project with no financial relationship with any pharmaceutical company, diagnostic laboratory, or clinical software vendor.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThe author thanks the contributors to the OpenOncology repository and the open-source communities maintaining OncoKB, cBioPortal, OpenTargets, ChEMBL, DiffDock, AlphaMissense, and AlphaFold. This work was conducted without external funding.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e \u003cp\u003eAll benchmark scripts and validation artifacts are publicly available at \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003egithub.com/immortal71/openoncology\u003c/span\u003e. Benchmark JSON files (real_patient_benchmark_100.json, real_patient_benchmark_200.json) and concordance labels (scripts/concordance_labels.json) are included in the repository and derived from publicly accessible cBioPortal/TCGA data. The 50-case holdout artifactsare available at validation_results/holdout_50_results.txt and validation_results/holdout_50_metrics.json. No proprietary or restricted data were used.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWorld Health Organization (2024) Cancer. Global Health Estimates 2024. WHO, Geneva\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDruker BJ et al (2001) Efficacy and safety of a specific inhibitor of the BCR-ABL tyrosine kinase in chronic myeloid leukemia. N Engl J Med 344(14):1031\u0026ndash;1037\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoria JC et al (2018) Osimertinib in untreated EGFR-mutated advanced non-small-cell lung cancer. N Engl J Med 378(2):113\u0026ndash;125\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCerami E et al (2012) The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov 2(5):401\u0026ndash;404\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChakravarty D et al (2017) OncoKB: A precision oncology knowledge base. JCO Precis Oncol. ;1:PO.17.00011.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng J et al (2023) Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381(6664):eadg7492\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJumper J et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature 596(7873):583\u0026ndash;589\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCorso G et al (2023) DiffDock: Diffusion steps, twists, and turns for molecular docking. ICLR. arXiv:2210.01776\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOchoa D et al (2023) The next-generation Open Targets Platform: reimagined, redesigned, rebuilt. Nucleic Acids Res 51(D1):D1353\u0026ndash;D1359\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMendez D et al (2019) ChEMBL: towards direct deposition of bioassay data. Nucleic Acids Res 47(D1):D930\u0026ndash;D940\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFreshour SL et al (2021) Integration of the Drug-Gene Interaction Database (DGIdb 4.0) with open crowdsource efforts. Nucleic Acids Res 49(D1):D1144\u0026ndash;D1151\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWhirl-Carrillo M et al (2021) An Evidence-Based Framework for Evaluating Pharmacogenomics Knowledge for Personalized Medicine. Clin Pharmacol Ther 110(3):563\u0026ndash;572\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Independent Researcher","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"precision oncology, drug repurposing, custom drug discovery, de novo drug design, variant interpretation, OncoKB, AlphaMissense, DiffDock, ChEMBL, OpenTargets, open-source bioinformatics, low and middle-income countries","lastPublishedDoi":"10.21203/rs.3.rs-9707913/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9707913/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003ePrecision oncology depends on rapid, evidence-based matching of tumor variants to approved therapies. However, two compounding problems limit access for most patients worldwide: first, the interpretation infrastructure remains locked behind institutional subscriptions; second, even well-resourced precision oncology pipelines return empty outputs when no approved or repurposed drug exists for a patient\u0026rsquo;s specific mutation a complete dead-end that affects the majority of patients with rare or non-hotspot variants. Both problems are structural, not scientific.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe present OpenOncology, a fully open-source platform that solves both problems in In the upper you can see it's left aligned especially in, I think, methods. If we go down in the introduction, everything is center aligned so from front to last it is like a paragraph type like this. sequence. Stage one performs a clinical-grade variant calling workflow (FastQC \u0026rarr; BWA-MEM2 \u0026rarr; GATK), pathogenicity scoring (AlphaMissense), protein structure prediction (AlphaFold Server), molecular docking (DiffDock), and drug ranking from a weighted composite of OncoKB actionability, OpenTargets evidence, COSMIC frequency, clinical trial phase, and binding confidence. AlphaFold Server and DiffDock are computationally intensive external services; throughput in production deployments is subject to rate limits and available hardware. Stage two triggered when stage one finds no approved or repurposed match executes a fully automated custom drug discovery workflow: it queries ChEMBL and OpenTargets for lead molecules against the patient\u0026rsquo;s specific target, scores oral bioavailability via Lipinski Rule of Five, generates a mutation specific AlphaFold protein structure, and assembles a manufacturer-ready discovery brief that pharmaceutical companies can bid on through an integrated marketplace. A crowdfunding module enables patients to raise resources for custom synthesis.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eValidation against a blinded 50-case oncologist holdout yielded Hit@3\u0026thinsp;=\u0026thinsp;0.900, Standard Precision@3\u0026thinsp;=\u0026thinsp;0.508 (ceiling: 0.625), Normalised Precision@3\u0026thinsp;=\u0026thinsp;0.817, Mean Reciprocal Rank\u0026thinsp;=\u0026thinsp;0.883, and zero false-positive recommendations. The 50-case holdout included 12 Level 3\u0026ndash;4 literature-sourced cases and 6 negative controls, representing a deliberately harder validation set than smaller prior holdouts; the metric profile reflects increased case difficulty. Stage two (custom drug discovery) validation is structural discovery briefs are verified to contain real ChEMBL and OpenTargets records; clinical validation of lead molecule selection requires experimental binding assays outside the scope of this release. Equivalence-adjusted oncologist concordance reached 100% at both Top-1 and Top-3 across 36 actionable TCGA cases. TCGA benchmarks at 100 and 200 patients demonstrated 100% pipeline coverage with zero empty outputs every patient received either an approved drug recommendation or a structured custom discovery brief.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eOpenOncology is the first open-source precision oncology platform to provide a complete, safe escalation pathway from approved drug matching through to de novo custom drug discovery for patients with no existing therapeutic option. All code, benchmark scripts, and validation artifacts are publicly available at github.com/immortal71/openoncology under the MIT licence.\u003c/p\u003e","manuscriptTitle":"OpenOncology: An Open-Source Framework for Evidence-Based Drug Matching and De Novo Custom Drug Discovery in Precision Oncology","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-18 03:18:34","doi":"10.21203/rs.3.rs-9707913/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"63f2da8b-31ad-44f7-bee5-01ef1e1826ed","owner":[],"postedDate":"May 18th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":68170375,"name":"Oncology"},{"id":68170376,"name":"Bioinformatics"}],"tags":[],"updatedAt":"2026-05-18T03:18:35+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-18 03:18:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9707913","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9707913","identity":"rs-9707913","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}