README.md

---
configs:
- config_name: mondo_ontology
  default: true
  data_files:
  - split: train
    path: mondo_ontology/train-00000-of-00001.parquet
- config_name: mutated_genes
  data_files:
  - split: train
    path: mutations/mutated_genes.parquet
- config_name: assayed_samples
  data_files:
  - split: train
    path: mutations/assayed_samples.parquet
- config_name: assayed_genes
  data_files:
  - split: train
    path: mutations/assayed_genes.parquet
task_categories:
- other
tags:
- genomics
- bulk-rna-seq
- cancer
- gene-expression
- mutations
- mondo
- benchmark
---

# Mapping benchmark

A collection of harmonized gene-expression datasets, somatic mutation tables and a MONDO disease hierarchy for benchmarking the matching of biological models to patient tumour profiles.

The collection brings together patient samples, patient-derived xenografts (PDXs) and cancer cell lines. Expression datasets share a **19,260-gene reference panel**, and disease descriptions were mapped to MONDO by our team using metadata supplied by the original authors.

| Component | Contents | Format |
|---|---|---|
| Expression | Five datasets with TPM values, sample metadata and MONDO labels | AnnData / H5AD |
| Mutations | Gene-level mutation presence and sample/gene reference tables for TCGA, META-PRISM and DepMap | Parquet |
| Disease hierarchy | 15,481 MONDO terms and 25,555 parent relations | Parquet |

## Expression datasets

Each H5AD contains a sample-by-gene matrix. All five datasets use the same ordered gene panel.

| Dataset | Biological material | Samples | Genes | Zero-padded genes | File |
|---|---|---:|---:|---:|---|
| PDXE | Patient-derived xenografts | 357 | 19,260 | 315 | [pdxe_tpm.h5ad](pdxe_tpm.h5ad) |
| META-PRISM | Metastatic patient tumours | 932 | 19,260 | 79 | [metaprism_tpm.h5ad](metaprism_tpm.h5ad) |
| TCGA | Patient tissue samples | 11,425 | 19,260 | 29 | [tcga_tpm.h5ad](tcga_tpm.h5ad) |
| DepMap 24Q4 | Cancer cell lines | 1,508 | 19,260 | 285 | [depmap_tpm.h5ad](depmap_tpm.h5ad) |
| GSE317901 | Head-and-neck cancer patient/PDX samples | 114 | 19,260 | 106 | [gse317901_tpm.h5ad](gse317901_tpm.h5ad) |

### What is stored in each H5AD?

| Field | Description |
|---|---|
| `X` | Gene-expression values in TPM, stored as a sparse matrix |
| `obs_names` | Sample identifiers used to join expression and mutation data |
| `obs["ontology_disease_matched_id"]` | MONDO disease identifiers added by our team |
| `obs["ontology_disease_matched_name"]` | Corresponding disease names added by our team |
| `obs` | Selected source-derived tissue, histology and disease descriptors; available fields vary by dataset |
| `var_names` | Harmonized Ensembl gene identifiers |
| `var["artificial_gene"]` | Boolean flag identifying genes appended to complete the common panel |

**Zero padding:** genes missing from a dataset's standardized gene set were added with all-zero expression and `artificial_gene=True`. A value of `False` means that the gene was already present; it does not guarantee nonzero expression.

### Preparation

**Gene harmonization.** Gene identifiers were standardized. Genes were filtered against a common reference panel of 19,260 genes. Genes outside the reference panel were excluded. Missing panel genes were appended with zeros, and the final columns were sorted by Ensembl identifier.

**Expression processing.** For PDXE, human transcript TPM values were aggregated to genes and rescaled to one million per sample over the selected panel. META-PRISM TPM measurements were matched to clinical records. TCGA used GDC unstranded gene TPM values. DepMap 24Q4 log2(TPM+1) values were converted back to TPM. GSE317901 TPM measurements were associated with patient/PDX sample metadata.

**Disease annotation.** MONDO labels were added by our team based on the authors' metadata columns. Samples without a validly formatted MONDO identifier were excluded. Expression values in the final export were preserved from the corresponding standardized matrices.

## Mutation data

Three complementary tables connect mutation data directly to the expression samples. The `dataset` column contains `tcga`, `metaprism` or `depmap`.

| File | One row represents | Columns | Rows |
|---|---|---|---:|
| [mutated_genes.parquet](mutations/mutated_genes.parquet) | A sample–gene pair with a qualifying mutation | `dataset`, `sample_id`, `gene_id_ensembl`, `gene_symbol` | 345,620 |
| [assayed_samples.parquet](mutations/assayed_samples.parquet) | An expression sample represented in the mutation matrices | `dataset`, `sample_id` | 10,696 |
| [assayed_genes.parquet](mutations/assayed_genes.parquet) | A gene represented in a dataset's mutation matrix | `dataset`, `gene_id_ensembl` | 37,298 |

| Dataset | Samples with mutation data | Mutated sample–gene pairs | Genes in the mutation matrix |
|---|---:|---:|---:|
| TCGA | 8,877 | 244,203 | 18,566 |
| META-PRISM | 354 | 567 | 124 |
| DepMap | 1,465 | 100,850 | 18,608 |

**Mutation definition** 
Variants with `vep_impact` equal to `HIGH` or `MODERATE` were retained. Multiple qualifying variants in a gene were collapsed to binary presence. These tables describe mutation presence, rather than individual variants or allele frequencies.

**Sample matching** 
Join by `dataset` and `sample_id`; `sample_id` matches the observation identifier in the corresponding H5AD. TCGA patient-level calls and META-PRISM subject-level calls were assigned to their expression samples. DepMap calls were matched to model identifiers. DepMap mutation data use **25Q3**, while expression data use **24Q4**.

**Missing data** 
A sample present in `assayed_samples` but absent from `mutated_genes` has no qualifying mutation in the represented gene set. A sample absent from `assayed_samples` has no mutation data in this export.

**Comparison panel** 
The intersection of the three mutation gene sets contains **124 genes**. The `assayed_genes` table records columns present in the input mutation matrices. 

## MONDO disease hierarchy

[mondo_ontology/train-00000-of-00001.parquet](mondo_ontology/train-00000-of-00001.parquet) contains **15,481 terms**, **25,555 parent relations** and ten columns.

| Columns | Description |
|---|---|
| `mondo_id`, `name` | Disease identifier and name |
| `mondo_level` | Minimum distance from the root |
| `parent_mondo_id` | One selected direct parent, provided for convenience |
| `direct_ancestors_ids`, `direct_ancestors_names`, `len_direct_ancestors` | All direct parents and their count |
| `all_ancestors_ids`, `all_ancestors_names`, `len_all_ancestors` | Ancestors and their count |

Use `direct_ancestors_ids` when reconstructing the graph, because terms may have more than one parent.

The non-human animal disease branch (`MONDO:0005583` and 62 exclusive descendants) was removed. The resulting graph is acyclic and has one root, `MONDO:0000001`.

## Sources and attribution

| Component | Original source |
|---|---|
| PDXE expression | NIBR; Gao et al., *Nature Medicine* (2015), [doi:10.1038/nm.3954](https://doi.org/10.1038/nm.3954); [PDXE data deposit](https://figshare.com/articles/dataset/pdxe_sample_annotations_txt/13331072) |
| META-PRISM expression | Gustave Roussy; [META-PRISM processed data](https://github.com/gustaveroussy/MetaPRISM_Public/tree/master/data) |
| TCGA expression | NCI/NHGRI; [Genomic Data Commons](https://portal.gdc.cancer.gov/) |
| DepMap expression | Broad Institute; [DepMap 24Q4 Public](https://doi.org/10.25452/figshare.plus.27993248) |
| GSE317901 expression | Queen's University Belfast and collaborators; [GEO GSE317901](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE317901) |
| TCGA mutations | [cBioPortal Datahub](https://github.com/cBioPortal/datahub/tree/master/public), `*tcga_pan_can_atlas_2018*/data_mutations.txt` |
| META-PRISM mutations | [Publication supplementary material](https://pmc.ncbi.nlm.nih.gov/articles/PMC10157368/#sec63), Table S6 |
| DepMap mutations | Broad Institute; [DepMap Public 25Q3](https://depmap.org/portal/download/all/), `OmicsSomaticMutations.csv` |
| MONDO | Monarch Initiative; [Mondo Disease Ontology](https://mondo.monarchinitiative.org/); [official OBO reference](https://purl.obolibrary.org/obo/mondo.obo) |

### Licence and reuse

Source-specific licences and attribution requirements apply. 
The collection does not assign a single licence to all components.

The PDXE deposit, DepMap 24Q4 expression release and MONDO ontology identify **CC BY 4.0**. META-PRISM processed expression tables were made publicly available by the authors. TCGA expression was obtained through GDC open-access resources. GEO data are subject to [NCBI's data-use notice](https://www.ncbi.nlm.nih.gov/geo/info/disclaimer.html).