{ "version": "2.0.0", "generated": "2026-05-12", "count": 28, "disclaimer": "These benchmarks reference datasets that are NOT publicly accessible. Listed for industry-relevance reference only. Compiled from public statements, publications, and press releases. Access is gated by collaboration, data-sharing agreements, or remains closed entirely.", "private_benchmarks": [ { "id": "pfizer-phase2-internal", "name": "Pfizer Phase II Trial Benchmark Dataset", "owner": "Pfizer", "owner_type": "pharma", "stage": "phase-ii", "modality": "cross-modality", "task": "Phase II efficacy and endpoint prediction from historical proprietary trial data", "public_proxy": "hint-trialbench", "access_mechanism": "closed", "evidence_source": "Pfizer AI/ML publications 2024-2025 reference internal trial benchmarks for endpoint modeling; data not released.", "estimated_size": "~1500 trials, 40,000 patients", "why_it_matters": "Real-world trial outcomes beat public synthetic sets; models trained here inform go/no-go on live programs.", "date_sourced": "2026-05-12" }, { "id": "novartis-nibr-therapeutic-data", "name": "NIBR Therapeutics Data (NTD)", "owner": "Novartis NIBR", "owner_type": "pharma", "stage": "lead-id-admet", "modality": "small-molecule", "task": "Internal compound bioactivity + phenotypic + ADMET cross-referenced against clinical outcomes", "public_proxy": "chembl", "access_mechanism": "collaboration", "evidence_source": "Described partially in Nat Rev Drug Discov 2023 (Schneider et al.) and NIBR public talks; ~4M compound entries referenced.", "estimated_size": "~4M compounds, 20M assays", "why_it_matters": "Largest single-sponsor ADMET dataset in industry; several academic collaborations have gated access.", "date_sourced": "2026-05-12" }, { "id": "astrazeneca-cas-dmpk", "name": "AstraZeneca CAS-backed DMPK Benchmarks", "owner": "AstraZeneca", "owner_type": "pharma", "stage": "lead-id-admet", "modality": "small-molecule", "task": "Compound DMPK prediction (clearance, PPB, Vd, half-life) on internal AZ library", "public_proxy": "tdc-admet", "access_mechanism": "closed", "evidence_source": "AZ MolecularAI papers (2021-2024) cite private DMPK evaluation sets; AZ-Karolinska joint releases expose subsets.", "estimated_size": "~200k measured DMPK endpoints", "why_it_matters": "AZ uses these benchmarks internally to gate compound progression.", "date_sourced": "2026-05-12" }, { "id": "merck-admet-demystify", "name": "Merck Internal ADMET Benchmark (Demystifying ADMET)", "owner": "Merck & Co.", "owner_type": "pharma", "stage": "lead-id-admet", "modality": "small-molecule", "task": "ADMET property prediction across proprietary MSD compound library", "public_proxy": "polaris-admet", "access_mechanism": "closed", "evidence_source": "Demystifying ADMET-AI publications (J Chem Inf Model, 2023; 2025) reference internal benchmark held out from model training.", "estimated_size": "~150k compounds across 17 endpoints", "why_it_matters": "Drives compound-prioritization decisions; model performance numbers published but data not.", "date_sourced": "2026-05-12" }, { "id": "genentech-gred-sar", "name": "Genentech gRED Structure-Activity Dataset", "owner": "Genentech gRED", "owner_type": "pharma", "stage": "hit-id", "modality": "small-molecule", "task": "Structure-activity modeling on active research programs", "public_proxy": "chembl", "access_mechanism": "closed", "evidence_source": "Genentech ML team presentations (NeurIPS AI4Science, ML4MolEng 2024).", "estimated_size": "undisclosed; referenced as 'millions of assay points'", "why_it_matters": "Used to benchmark LLMs and foundation models before internal deployment.", "date_sourced": "2026-05-12" }, { "id": "recursion-full-phenomics", "name": "Recursion Full Phenomics Dataset", "owner": "Recursion Pharmaceuticals", "owner_type": "pharma", "stage": "hit-id", "modality": "small-molecule", "task": "Morphological phenotypic profiling at petabyte scale", "public_proxy": "rxrx3-phenomics-bench", "access_mechanism": "conditional-access", "evidence_source": "Recursion SEC filings, Phenom-1 paper; public RxRx1/3 subsets are a fraction of the full 50+ PB dataset.", "estimated_size": "~50 PB images, 18M compound wells", "why_it_matters": "The actual signal powering Phenom-1, Phenom-2 foundation models.", "date_sourced": "2026-05-12" }, { "id": "insilico-longevity-full", "name": "Insilico Longevity Benchmark (Full Dataset)", "owner": "Insilico Medicine", "owner_type": "pharma", "stage": "virtual-cell", "modality": "cross-modality", "task": "Aging prediction, lifespan extension, geroprotector scoring on proprietary longevity cohorts", "public_proxy": "longevity-bench-insilico", "access_mechanism": "conditional-access", "evidence_source": "Insilico Longevity Benchmark portal exposes public subsets; full dataset references internal clinical collaborations.", "estimated_size": "~1M individuals across NHANES + internal cohorts, 500k methylation samples", "why_it_matters": "World's largest curated longevity-relevant dataset; powers Insilico's aging drug programs.", "date_sourced": "2026-05-12" }, { "id": "moderna-mrna-internal", "name": "Moderna mRNA Design Internal Benchmark", "owner": "Moderna", "owner_type": "pharma", "stage": "hit-id", "modality": "rna-therapeutic", "task": "Codon optimization, UTR design, translation efficiency prediction", "public_proxy": "mrna-design-bench", "access_mechanism": "closed", "evidence_source": "Moderna publications reference proprietary expression datasets for mRNA design validation.", "estimated_size": "~100k constructs with HEK293 and primary-cell expression readout", "why_it_matters": "Critical for mRNA therapeutic design; informs every Moderna pipeline program.", "date_sourced": "2026-05-12" }, { "id": "pfizer-mrna-internal", "name": "Pfizer mRNA/LNP Internal Benchmark", "owner": "Pfizer", "owner_type": "pharma", "stage": "hit-id", "modality": "rna-therapeutic", "task": "LNP formulation + mRNA design optimization", "public_proxy": "mrna-design-bench", "access_mechanism": "closed", "evidence_source": "Pfizer-BioNTech publications cite proprietary LNP-mRNA benchmark sets.", "estimated_size": "undisclosed", "why_it_matters": "Underpins Comirnaty and next-gen mRNA vaccines.", "date_sourced": "2026-05-12" }, { "id": "deep-genomics-rna", "name": "Deep Genomics RNA Therapeutics Benchmark", "owner": "Deep Genomics", "owner_type": "biotech", "stage": "hit-id", "modality": "rna-therapeutic", "task": "Splicing prediction, antisense oligonucleotide design", "public_proxy": "mrna-design-bench", "access_mechanism": "collaboration", "evidence_source": "Deep Genomics AI platform papers and press releases; BigRNA model training set is proprietary.", "estimated_size": "undisclosed", "why_it_matters": "Powers BigRNA foundation model.", "date_sourced": "2026-05-12" }, { "id": "isomorphic-internal", "name": "Isomorphic Labs Internal Structure/Docking Benchmark", "owner": "Isomorphic Labs (Alphabet)", "owner_type": "biotech", "stage": "hit-id", "modality": "cross-modality", "task": "Structure-based drug design evaluation", "public_proxy": "plinder-v2", "access_mechanism": "closed", "evidence_source": "Isomorphic Labs public statements and AlphaFold 3 paper supplementary reference proprietary evaluation held out from training.", "estimated_size": "undisclosed", "why_it_matters": "Ground truth for AlphaFold 3+ and future Isomorphic models.", "date_sourced": "2026-05-12" }, { "id": "meta-fair-protein-internal", "name": "Meta FAIR Protein Design Internal Eval", "owner": "Meta FAIR / EvolutionaryScale", "owner_type": "biotech", "stage": "hit-id", "modality": "biologic", "task": "ESM / ESM3 evaluation on internal proprietary fitness experiments", "public_proxy": "proteingym", "access_mechanism": "closed", "evidence_source": "ESM3 paper supplementary mentions internal held-out wet-lab set.", "estimated_size": "undisclosed", "why_it_matters": "Validates claimed generalization for Meta/EvolutionaryScale's protein foundation models.", "date_sourced": "2026-05-12" }, { "id": "ibm-rxn-retrosyn-internal", "name": "IBM RXN Internal Retrosynthesis Benchmark", "owner": "IBM Research", "owner_type": "tech", "stage": "hit-id", "modality": "small-molecule", "task": "Retrosynthesis route planning on proprietary rule sets", "public_proxy": "uspto-retrosyn", "access_mechanism": "collaboration", "evidence_source": "IBM RXN papers reference internal datasets beyond USPTO public set.", "estimated_size": "~5-10M proprietary reactions beyond USPTO", "why_it_matters": "Deeper retrosynthesis benchmark for industrial routes.", "date_sourced": "2026-05-12" }, { "id": "fda-cdrh-ai-validation", "name": "FDA CDRH Internal AI Validation Sets", "owner": "FDA Center for Devices and Radiological Health", "owner_type": "regulatory", "stage": "post-market-rwe", "modality": "cross-modality", "task": "Regulatory validation of medical AI/ML devices", "public_proxy": "faers-bench", "access_mechanism": "closed", "evidence_source": "FDA CDRH AI/ML Action Plan (2021-2025) mentions internal validation datasets for 510(k) and De Novo pathway.", "estimated_size": "undisclosed", "why_it_matters": "Regulatory gold standard for AI-device approval.", "date_sourced": "2026-05-12" }, { "id": "open-targets-pharma-partners", "name": "Open Targets Pharma Partner Extensions", "owner": "Open Targets consortium", "owner_type": "consortium", "stage": "target-id", "modality": "cross-modality", "task": "Target-disease association with proprietary pharma datasets layered on", "public_proxy": "open-targets", "access_mechanism": "data-sharing-agreement", "evidence_source": "Open Targets consortium papers (GSK, Sanofi, BMS, Pfizer) reference member-only data extensions.", "estimated_size": "undisclosed", "why_it_matters": "Adds internal essentiality/chemogenomics signal to the public target-disease graph.", "date_sourced": "2026-05-12" }, { "id": "bms-internal-sar", "name": "Bristol-Myers Squibb Internal SAR Benchmark", "owner": "Bristol-Myers Squibb", "owner_type": "pharma", "stage": "hit-id", "modality": "small-molecule", "task": "Structure-activity modeling on internal BMS research programs", "public_proxy": "chembl", "access_mechanism": "closed", "evidence_source": "BMS AI papers (Schuffenhauer et al.) cite proprietary SAR sets.", "estimated_size": "~1M assay points", "why_it_matters": "Industry benchmark for generative chemistry deployment.", "date_sourced": "2026-05-12" }, { "id": "valence-internal-expand", "name": "Valence Labs Internal ADMET Extensions", "owner": "Valence Labs (Recursion)", "owner_type": "biotech", "stage": "lead-id-admet", "modality": "small-molecule", "task": "ADMET prediction on Recursion-generated compounds beyond Polaris-public releases", "public_proxy": "polaris-admet", "access_mechanism": "conditional-access", "evidence_source": "Valence/Recursion publications + Polaris Hub description of partial data release.", "estimated_size": "~150k compounds", "why_it_matters": "Full set trains Recursion's ADMET foundation models.", "date_sourced": "2026-05-12" }, { "id": "sanofi-internal-pk", "name": "Sanofi Internal PK/PD Benchmark", "owner": "Sanofi", "owner_type": "pharma", "stage": "ind-enabling", "modality": "small-molecule", "task": "PK projection, interspecies scaling", "public_proxy": "pkpd-obach", "access_mechanism": "closed", "evidence_source": "Sanofi AI collaborations with Exscientia reference internal PK benchmarks.", "estimated_size": "undisclosed", "why_it_matters": "Drives AI-enabled PK modeling decisions.", "date_sourced": "2026-05-12" }, { "id": "gilead-hcv-antiviral", "name": "Gilead Internal Antiviral Benchmark", "owner": "Gilead Sciences", "owner_type": "pharma", "stage": "hit-id", "modality": "small-molecule", "task": "Antiviral potency modeling (HCV, HIV, SARS-CoV-2) on proprietary compound archive", "public_proxy": "asap-antiviral-2025", "access_mechanism": "closed", "evidence_source": "Gilead research publications reference internal antiviral screening collections.", "estimated_size": "~500k compounds antiviral screened", "why_it_matters": "Real-world antiviral MoA coverage not seen in public sets.", "date_sourced": "2026-05-12" }, { "id": "genomic-health-rwe", "name": "Flatiron Health Real-World Oncology Benchmark", "owner": "Flatiron Health (Roche)", "owner_type": "pharma", "stage": "post-market-rwe", "modality": "cross-modality", "task": "Real-world outcomes and patient journey modeling", "public_proxy": "mimic-benchmark", "access_mechanism": "data-sharing-agreement", "evidence_source": "Flatiron publications describe de-identified oncology EHR + outcomes as a proprietary benchmark for AI models.", "estimated_size": "~4M oncology patients", "why_it_matters": "Gold standard for RWE AI; widely licensed to pharma.", "date_sourced": "2026-05-12" }, { "id": "tempus-onco-ai", "name": "Tempus Oncology AI Benchmark", "owner": "Tempus Labs", "owner_type": "biotech", "stage": "target-id", "modality": "cross-modality", "task": "Multi-omic oncology prediction on Tempus's clinical-molecular database", "public_proxy": "cptac-proteogenomic", "access_mechanism": "data-sharing-agreement", "evidence_source": "Tempus publications (Nat Med 2023-2025) reference internal benchmarks.", "estimated_size": "~200k sequenced patients + outcomes", "why_it_matters": "Feeds Tempus AI oncology product.", "date_sourced": "2026-05-12" }, { "id": "ginkgo-biologics", "name": "Ginkgo Bioworks Biologics Design Benchmark", "owner": "Ginkgo Bioworks", "owner_type": "biotech", "stage": "hit-id", "modality": "biologic", "task": "Protein engineering and enzyme design across Ginkgo's foundry", "public_proxy": "protein-design-bench-2026", "access_mechanism": "collaboration", "evidence_source": "Ginkgo AI publications and press releases reference foundry-scale internal evaluation.", "estimated_size": "~millions of enzyme variants with activity measurements", "why_it_matters": "Ground truth for industrial enzyme design.", "date_sourced": "2026-05-12" }, { "id": "roche-pred-admet", "name": "Roche pRED ADMET Benchmark", "owner": "Roche pRED", "owner_type": "pharma", "stage": "lead-id-admet", "modality": "small-molecule", "task": "ADMET modeling on Roche's internal library", "public_proxy": "tdc-admet", "access_mechanism": "closed", "evidence_source": "Roche pRED publications on ML-ADMET reference proprietary datasets.", "estimated_size": "~300k compound-endpoint pairs", "why_it_matters": "Underpins Roche's internal ML-ADMET deployment.", "date_sourced": "2026-05-12" }, { "id": "chugai-ab", "name": "Chugai Antibody Engineering Benchmark", "owner": "Chugai Pharmaceutical (Roche)", "owner_type": "pharma", "stage": "hit-id", "modality": "biologic", "task": "Bispecific and switch antibody optimization", "public_proxy": "therapeutic-antibody-bench-2026", "access_mechanism": "collaboration", "evidence_source": "Chugai publications reference internal switch-antibody dataset.", "estimated_size": "undisclosed", "why_it_matters": "Unique bispecific / switch format benchmark.", "date_sourced": "2026-05-12" }, { "id": "exscientia-design", "name": "Exscientia Precision Medicine Benchmark", "owner": "Exscientia", "owner_type": "biotech", "stage": "hit-id", "modality": "small-molecule", "task": "Generative design + phenotypic outcome modeling on proprietary drug programs", "public_proxy": "chembl", "access_mechanism": "closed", "evidence_source": "Exscientia publications and 10-K references for Centaur AI outputs.", "estimated_size": "undisclosed", "why_it_matters": "Precision medicine + generative chem coupled benchmark.", "date_sourced": "2026-05-12" }, { "id": "xaira-foundation-internal", "name": "Xaira Therapeutics Foundation Benchmark", "owner": "Xaira Therapeutics", "owner_type": "biotech", "stage": "virtual-cell", "modality": "cross-modality", "task": "Virtual-cell + target-based design on proprietary multi-omic panels", "public_proxy": "virtualcellbench-2026", "access_mechanism": "closed", "evidence_source": "Xaira launch materials (Apr 2024) and subsequent publications mention proprietary benchmarks for foundation models.", "estimated_size": "undisclosed; built on Illumina-scale partnerships", "why_it_matters": "New well-funded entrant; benchmarks shape entire platform.", "date_sourced": "2026-05-12" }, { "id": "open-problems-sponsor-private", "name": "Open Problems Sponsor-Private Challenge Data", "owner": "Open Problems consortium (sponsors: 10x Genomics, Chan Zuckerberg Biohub)", "owner_type": "consortium", "stage": "virtual-cell", "modality": "cross-modality", "task": "Sponsor-provided test sets for Kaggle-hosted competitions", "public_proxy": "openproblems-perturbation", "access_mechanism": "conditional-access", "evidence_source": "Kaggle Open Problems competition rules; test labels held out until competition close.", "estimated_size": "varies per competition (~100k cells each)", "why_it_matters": "Partially-private model for prospective benchmark evaluation.", "date_sourced": "2026-05-12" }, { "id": "takeda-pipeline-ai", "name": "Takeda Internal AI Pipeline Benchmark", "owner": "Takeda", "owner_type": "pharma", "stage": "lead-id-admet", "modality": "small-molecule", "task": "ADMET + PK on Takeda internal compound archive", "public_proxy": "tdc-admet", "access_mechanism": "collaboration", "evidence_source": "Takeda AI platform publications reference internal evaluation set.", "estimated_size": "undisclosed", "why_it_matters": "Japan-based pharma with distinct chemistry class distribution.", "date_sourced": "2026-05-12" } ] }