[
  {
    "dataset_id": "DS-2026-MAPPING-PDXE",
    "dataset_name": "PDXE NIBR MONDO-labelled bulk RNA-seq TPM",
    "version": "v1 / 2026-09-11;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Research benchmarking of methods that match biological models and patient-derived tumour expression profiles using measured gene expression and MONDO disease labels. This register entry describes the benchmark data export, not an approved clinical decision system; no training/validation/test partition is defined by this export.",
    "source": "Novartis Institutes for BioMedical Research (NIBR), PDXE; Gao et al., Nature Medicine (2015), DOI 10.1038/nm.3954. Reference: https://figshare.com/articles/dataset/pdxe_sample_annotations_txt/13331072",
    "source_and_acquisition": "The benchmark preparation pipeline uses NIBR PDXE sample annotations (https://ndownloader.figshare.com/files/25680443) and mouse/human isoform TPM measurements (https://ndownloader.figshare.com/files/53789948), distributed through Figshare. Preparation comprises selection of human transcripts, transcript-to-Ensembl-gene mapping and aggregation, alignment with sample annotations, MONDO disease labels and tissue/histology metadata for the benchmark. MONDO ID/name annotations are added from disease descriptions; they are not taken from the authors’ files.",
    "license": "The PDXE data deposit by Joshua Korn identifies CC BY 4.0. ",
    "ai_augmentation": null,
    "provenance": "First benchmark release v1 (2026-09-11). Experimental expression and study metadata originate from the authors/resources named in source. Our team harmonizes metadata and adds MONDO disease annotations as described in processing_steps. From our annotated matrix, 446 samples are assessed, 357 retained and 89 excluded for missing/invalid MONDO IDs.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified on the v1 export: all 357 retained rows have a syntactically valid MONDO ID; all genes and their order are unchanged; sample identifiers are unique and retained in source order; expression is finite and nonnegative; the output contains only the benchmark metadata allowlist and no auxiliary AnnData slots. Gene audit: 315 artificial columns, all zero; panel membership and ordering match the inspected reference exactly.",
    "processing_steps": [
      "Authors’ data: The source data used were NIBR PDXE mouse/human isoform TPM and sample annotation table, including Primary_Site and Histological_subtype. MONDO and tissue-ontology columns are not author-supplied fields.",
      "Expression processing: Human ENST transcripts were selected. Transcript-version suffixes were removed. Transcripts were mapped to Ensembl genes. Unmapped transcripts were discarded. Transcript-level TPM values were summed for each gene. The matrix was transposed so that rows represented samples and columns represented genes. Sample annotations were joined to the expression matrix. TPM values were stored in CSR format.",
      "Metadata harmonization: The tissue column was created from the authors’ Primary_Site field. The condition column was created from Histological_subtype. The original site and histology fields were preserved.",
      "Gene-ID harmonization: Version suffixes were removed from ENSG IDs. Identifiers were resolved, with priority given to Ensembl, NCBI, symbol and synonym mappings. Unresolved IDs were removed.",
      "Gene-panel selection: Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded.",
      "Gene completion: Missing panel genes were appended with all-zero expression and artificial_gene=True. Genes already present at this stage were marked artificial_gene=False. This value did not imply nonzero expression. The final matrix was sorted by Ensembl ID.",
      "PDXE-specific rescaling: After gene standardization, each sample was rescaled over the selected panel using normalize_total(target_sum=1e6).",
      "Disease annotation: Disease descriptions were mapped to MONDO by our team. The ontology_disease_matched_id and ontology_disease_matched_name columns were added."
    ],
    "bias_assessment": "MONDO disease labels were added by our team based on the metadata columns provided by the original authors.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-11"
  },
  {
    "dataset_id": "DS-2026-MAPPING-METAPRISM",
    "dataset_name": "METAPRISM MONDO-labelled bulk RNA-seq TPM",
    "version": "v1 / 2026-09-11; ",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Research benchmarking of methods that match biological models and patient-derived tumour expression profiles using measured gene expression and MONDO disease labels added by our team. This register entry describes the benchmark data export, not an approved clinical decision system; no training/validation/test partition is defined by this export.",
    "source": "Gustave Roussy, META-PRISM study of refractory/metastatic cancers. Reference: https://github.com/gustaveroussy/MetaPRISM_Public/tree/master/data",
    "source_and_acquisition": "The benchmark preparation pipeline uses the META-PRISM authors’ TPM table Data_Table_2.rna_gene_tpm.tsv and supplementary clinical Tables 1 and 2, available through https://github.com/gustaveroussy/MetaPRISM_Public/tree/master/data and the authors’ download service https://nextcloud.gustaveroussy.fr/s/dbCwPWKZMj6BmDp. Preparation comprises selection of tumour RNA records, matching expression to clinical annotations by sample/subject identifiers, and selection of expression, MONDO disease labels and tissue/histology metadata for the benchmark. MONDO ID/name annotations are added by our team from disease descriptions; they are not taken from the authors’ files.",
    "license": "The processed expression tables used in this dataset were made publicly available by the META-PRISM authors.",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "First benchmark release v1 (2026-09-11). Experimental expression and study metadata originate from the authors/resources named in source. Our team harmonizes metadata and adds MONDO disease annotations as described in processing_steps. 932 samples are assessed, 932 retained and 0 excluded for missing/invalid MONDO IDs. META-PRISM mutation data were also used in the benchmark and are documented separately under DS-2026-MAPPING-MUTATED-GENES, DS-2026-MAPPING-ASSAYED-SAMPLES and DS-2026-MAPPING-ASSAYED-GENES.",
    "legal_basis_and_rights":"n/a",
    "quality_criteria": "Verified on the v1 export: all 932 retained rows have a syntactically valid MONDO ID; all 19260 genes and their order are unchanged; sample identifiers are unique and retained in source order; expression is finite and nonnegative; the output contains only the benchmark metadata allowlist and no auxiliary AnnData slots. Gene audit: 79 artificial columns, all zero; panel membership and ordering match the inspected reference exactly.",
    "processing_steps": [
      "Authors’ data: The source data used were META-PRISM gene TPM table and supplementary clinical tables, including Primary_Tumor_Site, Histology, Biopsy_Site. MONDO and tissue-ontology columns are not author-supplied fields.",
      "Expression/sample processing: RNA_T clinical records were selected. Supplementary tables were joined by Subject_Id. The TPM table was transposed, and expression samples were matched to clinical metadata. The documented preparation was run with keep_primary=False. After metadata matching, 932 of 947 expression samples were retained.",
      "Metadata harmonization: The tissue column was created from Biopsy_Site. The condition column was created from Histology. The selected original tumour-site, histology and biopsy-site fields were preserved.",
      "Gene-ID harmonization: Version suffixes were removed from ENSG IDs. Identifiers were resolved, with priority given to Ensembl, NCBI, symbol and synonym mappings. Unresolved IDs were removed.",
      "Gene-panel selection: Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded.",
      "Gene completion: Missing panel genes were appended with all-zero expression and artificial_gene=True. Genes already present at this stage were marked artificial_gene=False. This value did not imply nonzero expression. The final matrix was sorted by Ensembl ID.",
      "Disease annotation: Disease descriptions were mapped to MONDO by our team. The ontology_disease_matched_id and ontology_disease_matched_name columns were added."
    ],
    "bias_assessment": "MONDO disease labels were added by our team based on the metadata columns provided by the original authors.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-11"
  },
  {
    "dataset_id": "DS-2026-MAPPING-TCGA",
    "dataset_name": "TCGA MONDO-labelled bulk RNA-seq TPM",
    "version": "v1 / 2026-09-11; ",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Research benchmarking of methods that match biological models and patient-derived tumour expression profiles using measured gene expression and MONDO disease labels added by our team. This register entry describes the benchmark data export, not an approved clinical decision system; no training/validation/test partition is defined by this export.",
    "source": "The Cancer Genome Atlas (TCGA), NCI/NHGRI; Genomic Data Commons (GDC). Reference: https://portal.gdc.cancer.gov/",
    "source_and_acquisition": "The benchmark preparation pipeline uses TCGA unstranded TPM gene-expression quantification and associated sample, case and diagnostic metadata from the NCI Genomic Data Commons (https://portal.gdc.cancer.gov/). Preparation comprises assembling the sample-by-gene expression matrix, joining annotations by GDC file and case/patient identifiers, and selecting expression, MONDO disease labels and tissue/disease metadata for the benchmark. MONDO ID/name annotations are added by our team from disease descriptions; they are not taken from the authors’ files.",
    "license": "GDC open-access",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "First benchmark release v1 (2026-09-11). Experimental expression and study metadata originate from the authors/resources named in source. Our team harmonizes metadata and adds MONDO disease annotations as described in processing_steps. From our annotated matrix, 11504 samples are assessed, 11425 retained and 79 excluded for missing/invalid MONDO IDs. TCGA mutation data were also used in the benchmark and are documented separately under DS-2026-MAPPING-MUTATED-GENES, DS-2026-MAPPING-ASSAYED-SAMPLES and DS-2026-MAPPING-ASSAYED-GENES.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified on the v1 export: all 11425 retained rows have a syntactically valid MONDO ID; all 19260 genes and their order are unchanged; sample identifiers are unique and retained in source order; expression is finite and nonnegative; the output contains only the benchmark metadata allowlist and no auxiliary AnnData slots. Gene audit: 29 artificial columns, all zero; panel membership and ordering match the inspected reference exactly.",
    "processing_steps": [
      "Original data: The source data used were GDC TCGA unstranded gene TPM and associated sample/case/diagnosis metadata, including primary_site and disease_type. MONDO and tissue-ontology columns are not original GDC fields.",
      "Data integration: The assembled GDC TPM matrix was read. File metadata were joined by file_id, and clinical and diagnostic data were joined by case or patient identifiers. The entity_submitter_id field was used as the sample identifier.",
      "Metadata harmonization: The tissue column was created from primary_site. The condition column was created from Tissue Type (tumour/normal). Selected GDC disease and site descriptors were retained.",
      "Gene-ID harmonization: Version suffixes were removed from ENSG IDs. Identifiers were resolved, with priority given to Ensembl, NCBI, symbol and synonym mappings. Unresolved IDs were removed.",
      "Gene-panel selection: Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded.",
      "Gene completion: Missing panel genes were appended with all-zero expression and artificial_gene=True. Genes already present at this stage were marked artificial_gene=False. This value did not imply nonzero expression. The final matrix was sorted by Ensembl ID.",
      "Disease annotation: Disease descriptions were mapped to MONDO by our team. The ontology_disease_matched_id and ontology_disease_matched_name columns were added."
    ],
    "bias_assessment": "MONDO disease labels were added by our team based on the metadata columns provided by the original authors.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-11"
  },
  {
    "dataset_id": "DS-2026-MAPPING-DEPMAP",
    "dataset_name": "DepMap 24Q4 MONDO-labelled cell-line RNA-seq TPM",
    "version": "v1 / 2026-09-11;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Research benchmarking of methods that match biological models and patient-derived tumour expression profiles using measured gene expression and MONDO disease labels added by our team. This register entry describes the benchmark data export, not an approved clinical decision system; no training/validation/test partition is defined by this export.",
    "source": "Broad Institute, DepMap 24Q4 Public release; DOI 10.25452/figshare.plus.27993248. Reference: https://plus.figshare.com/articles/dataset/DepMap_24Q4_Public/27993248",
    "source_and_acquisition": "The benchmark preparation pipeline uses the Broad Institute DepMap 24Q4 Public release (https://plus.figshare.com/articles/dataset/DepMap_24Q4_Public/27993248), specifically OmicsExpressionProteinCodingGenesTPMLogp1.csv and associated profile/sample annotations, available through the DepMap download catalogue (https://depmap.org/portal/api/download/files). Preparation comprises converting log2(TPM+1) expression to TPM, linking cell-line annotations, MONDO disease labels and tissue/disease metadata for the benchmark. MONDO ID/name annotations are added by our team from disease descriptions; they are not taken from the authors’ files.",
    "license": "DepMap 24Q4 Public, DOI 10.25452/figshare.plus.27993248, is listed as CC BY 4.0 on its release deposit.",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "First benchmark release v1 (2026-09-11). Experimental expression and study metadata originate from the authors/resources named in source. Our team harmonizes metadata and adds MONDO disease annotations as described in processing_steps. From our annotated matrix, 1673 samples are assessed, 1508 retained and 165 excluded for missing/invalid MONDO IDs; all 19260 genes are retained. DepMap mutation data were also used in the benchmark and are documented separately under DS-2026-MAPPING-MUTATED-GENES, DS-2026-MAPPING-ASSAYED-SAMPLES and DS-2026-MAPPING-ASSAYED-GENES.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified on the v1 export: all 1508 retained rows have a syntactically valid MONDO ID; all 19260 genes and their order are unchanged; sample identifiers are unique and retained in source order; expression is finite and nonnegative; the output contains only the benchmark metadata allowlist and no auxiliary AnnData slots. The retained DepMap set contains 6 samples with two syntactically valid MONDO IDs absent from the pinned benchmark graph (MONDO:0019210 and MONDO:0850271). Gene audit: 285 artificial columns, all zero.",
    "processing_steps": [
      "Authors’ data: The source data used were DepMap 24Q4 OmicsExpressionProteinCodingGenesTPMLogp1.csv and release sample/profile metadata. MONDO and tissue-ontology columns are not author-supplied fields.",
      "Expression processing: Expression values were converted from log2(TPM+1) to TPM. Expression values were stored in CSR format with Ensembl gene identifiers.",
      "Metadata integration: Cell-characteristics annotations were linked by CVCL_id. The tissue column was created from cancer lineage, and the condition column was created from cancer primary disease.",
      "Gene-ID harmonization: Version suffixes were removed from ENSG IDs. Identifiers were resolved, with priority given to Ensembl, NCBI, symbol and synonym mappings. Unresolved IDs were removed.",
      "Gene-panel selection: Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded.",
      "Gene completion: Missing panel genes were appended with all-zero expression and artificial_gene=True. Genes already present at this stage were marked artificial_gene=False. This value did not imply nonzero expression. The final matrix was sorted by Ensembl ID.",
      "Disease annotation: Disease descriptions were mapped to MONDO by our team. The ontology_disease_matched_id and ontology_disease_matched_name columns were added."
    ],
    "bias_assessment": "MONDO disease labels were added by our team based on the metadata columns provided by the original authors.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-11"
  },
  {
    "dataset_id": "DS-2026-MAPPING-GSE317901",
    "dataset_name": "GEO GSE317901 MONDO-labelled bulk RNA-seq TPM",
    "version": "v1 / 2026-09-11;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Research benchmarking of methods that match biological models and patient-derived tumour expression profiles using measured gene expression and MONDO disease labels added by our team. This register entry describes the benchmark data export, not an approved clinical decision system; no training/validation/test partition is defined by this export.",
    "source": "NCBI GEO GSE317901, APHRODYTHE head-and-neck squamous-cell carcinoma patient/PDX study; Queen’s University Belfast and collaborators. Reference: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE317901",
    "source_and_acquisition": "The benchmark preparation pipeline uses author-supplied TPM expression and sample annotations from NCBI GEO series GSE317901 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE317901), the APHRODYTHE patient/PDX study. The authors describe sequencing-data processing in the GEO sample records, including https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSM9481476. Benchmark preparation comprises linking expression with sample annotations and selecting expression, MONDO disease labels and available tissue/sample descriptions. MONDO ID/name annotations are added by our team from disease descriptions; they are not taken from the authors’ files.",
    "license": "NCBI places no restrictions on use/distribution of GEO data, subject to third-party rights: https://www.ncbi.nlm.nih.gov/geo/info/disclaimer.html.",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "First benchmark release v1 (2026-09-11). Experimental expression and study metadata originate from the authors/resources named in source. Our team harmonizes metadata and adds MONDO disease annotations as described in processing_steps. From our annotated matrix, 114 samples are assessed, 114 retained and 0 excluded for missing/invalid MONDO IDs; all 19260 genes are retained.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified on the v1 export: all 114 retained rows have a syntactically valid MONDO ID; all 19260 genes and their order are unchanged; sample identifiers are unique and retained in source order; expression is finite and nonnegative; the output contains only the benchmark metadata allowlist and no auxiliary AnnData slots. Gene audit: 106 artificial columns, all zero.",
    "processing_steps": [
      "Authors’ data: The source data used were GEO GSE317901 TPM measurements and patient/PDX sample descriptions. MONDO labels are annotations added by our team, not fields supplied by the study authors.",
      "Authors’ sequencing processing: The study authors used FastQC, Trim Galore and STAR for sequencing processing. For PDX samples, human and mouse alignment and XenofilteR separation were performed, followed by HTSeq-count. These steps were performed by the authors and were separate from our benchmark annotation. The methods and availability of TPM tables were documented in GEO sample GSM9481476.",
      "Dataset integration: TPM measurements were associated with sample metadata. The tissue column in the resulting dataset contained sample descriptions.",
      "Gene-ID harmonization: Version suffixes were removed from ENSG IDs. Identifiers were resolved, with priority given to Ensembl, NCBI, symbol and synonym mappings. Unresolved IDs were removed.",
      "Gene-panel selection: Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded.",
      "Gene completion: Missing panel genes were appended with all-zero expression and artificial_gene=True. Genes already present at this stage were marked artificial_gene=False. This value did not imply nonzero expression. The final matrix was sorted by Ensembl ID.",
      "Disease annotation: Disease descriptions were mapped to MONDO by our team. The ontology_disease_matched_id and ontology_disease_matched_name columns were added. All 114 input samples carried the stored exact-match method. This method described our annotation and was not supplied by GEO."
    ],
    "bias_assessment": "MONDO disease labels were added by our team based on the metadata columns provided by the original authors.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-11"
  },
  {
    "dataset_id": "DS-2026-MAPPING-MUTATED-GENES",
    "dataset_name": "Non-silent mutated genes by expression sample",
    "version": "v1 / 2026-09-15;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Represent mutation presence for sample matching and Jaccard evaluation over the shared gene panel.",
    "source": "TCGA Pan-Cancer Atlas 2018 mutation files distributed through cBioPortal Datahub (https://github.com/cBioPortal/datahub/tree/master/public); Gustave Roussy META-PRISM publication supplementary Table S6 (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/#sec63); Broad Institute DepMap Public 25Q3, OmicsSomaticMutations.csv (https://depmap.org/portal/download/all/).",
    "source_and_acquisition": "TCGA: The preparation pipeline downloaded cBioPortal Datahub study folders with Git LFS and read public/*tcga_pan_can_atlas_2018*/data_mutations.txt. Cohort mutation tables were concatenated and their variant annotations were standardized. META-PRISM: Mutation calls were read from Table S6 of the publication supplement. DepMap: Raw somatic mutation calls were read from OmicsSomaticMutations.csv for release 25Q3; download URLs were resolved through the provider endpoint https://depmap.org/portal/api/download/files. For the benchmark, variants with vep_impact HIGH or MODERATE were collapsed to gene-level mutation presence. TCGA patient-level calls and META-PRISM subject-level calls were linked to the corresponding expression samples, while DepMap models were matched by model ID. The derived exports were aligned to the expression sample identifiers.",
    "license": "TCGAl: GDC open-access. META-PRISM: (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/). DepMap: CC by 4.0",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "One row per dataset, expression sample and Ensembl gene with at least one qualifying mutation. The supplied reference export contains 345620 rows: tcga 244203, depmap 100850, metaprism 567.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified column names and string types; no nulls, empty key values or duplicate keys. All assayed samples exist in the corresponding expression H5ADs.",
    "processing_steps": [
      "Mutation filtering: The documented non-silent definition was applied by selecting variants with HIGH or MODERATE impact. Multiple qualifying variants within a gene were collapsed to binary mutation presence.",
      "Sample matching: Patient-level TCGA calls and subject-level META-PRISM calls were associated with their expression samples. DepMap models were matched to expression sample identifiers.",
      "Mutation export: Positive sample–gene entries were exported. Gene symbols were attached where available from expression metadata."
    ],
    "bias_assessment": "TCGA, META-PRISM and DepMap differ in biological models and mutation ascertainment.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-15"
  },
  {
    "dataset_id": "DS-2026-MAPPING-ASSAYED-SAMPLES",
    "dataset_name": "Expression samples with mutation calls",
    "version": "v1 / 2026-09-15;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Distinguish samples with mutation data but no qualifying mutation from samples without mutation calls.",
    "source": "TCGA Pan-Cancer Atlas 2018 mutation files distributed through cBioPortal Datahub (https://github.com/cBioPortal/datahub/tree/master/public); Gustave Roussy META-PRISM publication supplementary Table S6 (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/#sec63); Broad Institute DepMap Public 25Q3, OmicsSomaticMutations.csv (https://depmap.org/portal/download/all/).",
    "source_and_acquisition": "TCGA: The preparation pipeline downloaded cBioPortal Datahub study folders with Git LFS and read public/*tcga_pan_can_atlas_2018*/data_mutations.txt. Cohort mutation tables were concatenated and their variant annotations were standardized. META-PRISM: Mutation calls were read from Table S6 of the publication supplement. DepMap: Raw somatic mutation calls were read from OmicsSomaticMutations.csv for release 25Q3; download URLs were resolved through the provider endpoint https://depmap.org/portal/api/download/files. For the benchmark, variants with vep_impact HIGH or MODERATE were collapsed to gene-level mutation presence. TCGA patient-level calls and META-PRISM subject-level calls were linked to the corresponding expression samples, while DepMap models were matched by Consensus_model_id. The derived exports were aligned to the expression sample identifiers.",
    "license": "TCGAl: GDC open-access. META-PRISM: (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/). DepMap: CC by 4.0",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "One row per dataset and expression sample represented in the mutation-call matrices, including all-zero rows. The supplied reference export contains 10696 rows: tcga 8877, depmap 1465, metaprism 354.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified column names and string types; no nulls, empty key values or duplicate keys. All assayed samples exist in the corresponding expression H5ADs.",
    "processing_steps": [
      "Sample selection: All sample rows were taken from the documented mutation matrices. Samples whose variants did not pass the non-silent filter were included.",
      "Sample matching: TCGA patients and META-PRISM subjects were associated with their expression samples. DepMap models were matched to expression sample identifiers.",
      "Sample export: Unique dataset and sample_id pairs were exported for samples present in the expression datasets."
    ],
    "bias_assessment": "TCGA, META-PRISM and DepMap differ in biological models and mutation ascertainment.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-15"
  },
  {
    "dataset_id": "DS-2026-MAPPING-ASSAYED-GENES",
    "dataset_name": "Gene sets represented in the mutation matrices",
    "version": "v1 / 2026-09-15;",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Define the per-dataset gene sets and their 124-gene intersection for mutation-based evaluation.",
    "source": "TCGA Pan-Cancer Atlas 2018 mutation files distributed through cBioPortal Datahub (https://github.com/cBioPortal/datahub/tree/master/public); Gustave Roussy META-PRISM publication supplementary Table S6 (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/#sec63); Broad Institute DepMap Public 25Q3, OmicsSomaticMutations.csv (https://depmap.org/portal/download/all/).",
    "source_and_acquisition": "TCGA: The preparation pipeline downloaded cBioPortal Datahub study folders with Git LFS and read public/*tcga_pan_can_atlas_2018*/data_mutations.txt. Cohort mutation tables were concatenated and their variant annotations were standardized. META-PRISM: Mutation calls were read from Table S6 of the publication supplement. DepMap: Raw somatic mutation calls were read from OmicsSomaticMutations.csv for release 25Q3; download URLs were resolved through the provider endpoint https://depmap.org/portal/api/download/files. For the benchmark, variants with vep_impact HIGH or MODERATE were collapsed to gene-level mutation presence. TCGA patient-level calls and META-PRISM subject-level calls were linked to the corresponding expression samples, while DepMap models were matched by Consensus_model_id. The derived exports were aligned to the expression sample identifiers.",
    "license": "TCGA: GDC open-access. META-PRISM: (https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/). DepMap: CC by 4.0",
    "is_synthetic": false,
    "ai_augmentation": null,
    "provenance": "One row per dataset and Ensembl gene represented as a column in the mutation matrix. The supplied reference export contains 37298 rows: tcga 18566, depmap 18608, metaprism 124.",
    "legal_basis_and_rights": "n/a",
    "quality_criteria": "Verified column names and string types; no nulls, empty key values or duplicate keys. All assayed samples exist in the corresponding expression H5ADs.",
    "processing_steps": [
      "Gene extraction: Gene columns were extracted from the documented TCGA, DepMap and META-PRISM mutation matrices.",
      "Gene export: Unique dataset and gene_id_ensembl pairs were exported. The gene sets were not restricted to the expression gene panel.",
      "Comparison panel: The intersection of the three gene sets was used as the comparison panel. The supplied intersection contained 124 genes."
    ],
    "bias_assessment": "TCGA, META-PRISM and DepMap differ in biological models and mutation ascertainment.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-15"
  },
  {
    "dataset_id": "DS-2026-MAPPING-MONDO",
    "dataset_name": "MONDO disease ontology",
    "version": "v1 / 2026-09-14",
    "dataset_type": "benchmark",
    "related_ai_system": [
      "Ingenix Research Cluster"
    ],
    "intended_purpose": "Provide disease identifiers, names and ancestor relationships for ontology-based comparison of disease labels in the mapping benchmark.",
    "source": "Mondo Disease Ontology (MONDO), Monarch Initiative. Official ontology reference: https://purl.obolibrary.org/obo/mondo.obo.",
    "source_and_acquisition": "Only the MONDO component was retained in the benchmark export. The official MONDO ontology is available in OWL, OBO and JSON formats.",
    "license": "Creative Commons Attribution 4.0 International (CC BY 4.0). MONDO licence: https://github.com/monarch-initiative/mondo/blob/master/LICENSE.",
    "ai_augmentation": null,
    "provenance": "The export contains 15,481 MONDO nodes, 25,555 parent relations and ten hierarchy columns. The non-human animal disease branch was excluded.",
    "legal_basis_and_rights": "The export contains ontology identifiers, disease names and graph relationships, without patient or sample records. Reuse of MONDO ontology content is subject to CC BY 4.0",
    "quality_criteria": "The recorded export audit confirms an acyclic graph with one root, MONDO:0000001, and preserved reachability among retained MONDO nodes. No dangling source parent references or conflicting nonempty source MONDO mappings were recorded.",
    "processing_steps": [
      "Source preparation: MONDO terms and relationships were prepared for the benchmark.",
      "Source validation: The source graph was checked for cycles and dangling parent references. Nonempty MONDO identifier mappings were checked for conflicts.",
      "Hierarchy processing: Direct parent relations between retained MONDO nodes were preserved. Paths through non-MONDO nodes were projected onto the nearest MONDO ancestors. Nine derived relations and their source paths were recorded.",
      "Branch filtering: MONDO:0005583 and its 62 exclusive descendants were excluded. No benchmark sample carried an excluded label.",
      "Hierarchy annotation: Parent relationships, ancestor closures, names, counts and levels were recomputed. The mondo_level field was defined as the minimum distance from a root. The original scalar parent was retained when it remained a direct parent. Otherwise, the first sorted parent was selected. The direct_ancestors_ids field was retained as the authoritative multi-parent representation.",
      "Export schema: Ten columns were retained, covering MONDO identifiers, names, levels, parents and ancestors."
    ],
    "bias_assessment": "Coverage and hierarchy reflect the source ontology content and the selected MONDO subgraph.",
    "data_owner": "Alicja Augustyniak",
    "last_reviewed": "2026-09-16"
  }
]