MatthewCorney/gene-benchmark

GeneBenchmark — Gene Embedding Evaluation Suite

A curated benchmark for evaluating gene-level embeddings across four biological dimensions: disease association, pathway/function membership, protein domain/structure, and tissue expression. All tasks are anchored to NCBI Entrez gene IDs (human, Homo sapiens) so any gene representation can be evaluated without re-labelling.

---

Dataset Summary

GeneBenchmark provides two task types drawn from 13 publicly available databases:

Task Type	Category	Sources	Prediction target
Binary classification	Disease	OpenTargets, ClinVar, HPO	Is gene G associated with disease D?
Binary classification	Function / Pathway	Reactome, WikiPathways, Gene Ontology (BP)	Is gene G a member of pathway/process P?
Binary classification	Structure / Domain	Pfam, InterPro, PANTHER	Does gene G encode a protein in domain family F?
Binary classification	Expression	HPA Protein IHC	Is gene G expressed (above not-detected) in tissue T?
Regression	Structure	MobiDB	Intrinsic disorder fraction of the encoded protein (0–1)
Regression	Disease	gnomAD	6 constraint metrics: synonymous Z, missense Z, pLI, and observed counts
Regression	Expression	HPA RNA	log1p(nTPM) consensus RNA expression level per tissue

For classification, a separate binary classifier is trained for each class (positive genes vs. all others). Performance is measured per class (e.g. AUROC) and averaged across all classes to produce a dataset-level score. The binary gene × class matrix is the primary file used for this evaluation.

---

Supported Tasks

Classification

Each classification subset supports gene-level binary classification, averaged across classes. For each class (disease, pathway, or domain family) a separate binary classifier is trained on gene embeddings, with positive examples drawn from annotated genes and negatives from all other genes. The final dataset score is the macro-average metric (e.g. AUROC) across all classes.

Standard evaluation protocol:

1. Load the binary gene × class matrix for the source of interest.
2. For each class column, train a binary classifier using gene embeddings as features.
3. Compute AUROC (or another metric) per class, then average across all classes.

Regression

Twenty-seven regression targets are provided across three sources. Each is scored independently (e.g. Spearman r or R²) and results are averaged within each source to produce a source-level score.

• MobiDB (1 target) — disorder fraction of the encoded protein (0.0–1.0 continuous). Evaluates whether embeddings capture structural flexibility.
• gnomAD (6 targets) — constraint metrics from population genetics. Each metric is a separate regression task; the six scores are averaged to give the gnomAD dataset score:
  • Synonymous Z-score
  • Missense Z-score
  • Loss-of-function pLI
  • Synonymous observed count
  • Missense observed count
  • Loss-of-function observed count
• HPA RNA (20 targets) — log1p-transformed consensus RNA expression (nTPM) per tissue from the Human Protein Atlas RNA consensus dataset. Each tissue is a separate regression task evaluated independently. Tissues covered: adipose tissue, adrenal gland, bone marrow, cerebellum, cerebral cortex, colon, heart muscle, kidney, liver, lung, lymph node, ovary, pancreas, skeletal muscle, skin, small intestine, spleen, stomach, testis, thyroid gland.

---

Dataset Structure

Data Sources and Versions

Source	Category	Version / Date	License
OpenTargets	Disease	Platform 26.03	CC-BY 4.0
ClinVar	Disease	Current (GRCh38)	Public domain
HPO	Disease	Latest release	HPO License (free academic use)
gnomAD	Disease	v4.1.1	CC-BY 4.0
Reactome	Function	v92	CC-BY 4.0
WikiPathways	Function	2026-03-10	CC0 1.0
Gene Ontology	Function	go-basic.obo (current)	CC-BY 4.0
Pfam	Structure	Current	CC0 1.0
InterPro	Structure	Current	CC0 1.0
PANTHER	Structure	PTHR19.0	CC-BY 4.0
MobiDB	Structure	Current	CC-BY 4.0
Human Protein Atlas (IHC)	Expression	Current	CC-BY 4.0
Human Protein Atlas (RNA)	Expression	Current	CC-BY 4.0

Classification File Schema

Primary evaluation file: the binary gene × class matrix (*_matrix.csv) — entrez_id as rows, one binary column per class. This is the file used directly for embedding evaluation.

The long-format TSV (below) is an intermediate representation used to generate the matrix. It is provided for reference and for users who want to construct alternative subsets or splits.

All long-format classification files share this four-column schema:

Column	Type	Description
entrez_id	string	NCBI Entrez gene ID
id	string	Class identifier (e.g. R-HSA-5673001, GO:0007186, DOID:14330, cerebral_cortex)
*_name	string	Human-readable class name (disease_name, pathway_name, domain_name, family_name, or tissue_name)
source	string	Source label (e.g. reactome, clinvar, pfam, hpa_protein)

Example rows:

entrez_id  id              pathway_name                       source
1956       R-HSA-1257604   PIP3 activates AKT signalling      reactome
1956       GO:0008284      positive regulation of cell pop...  gene_ontology

entrez_id  id              tissue_name        source
1956       cerebral_cortex Cerebral cortex    hpa_protein

Regression File Schema

Column	Type	Description
entrez_id	string	NCBI Entrez gene ID
value	float	Continuous target value
source	string	Source label (mobidb or gnomad)

gnomAD is split into six separate files, one per metric:

File	Metric
syn_z_score_matrix.csv	Synonymous Z-score
mis_z_score_matrix.csv	Missense Z-score
lof_pLI_matrix.csv	Loss-of-function pLI
syn_obs_matrix.csv	Synonymous observed count
mis_obs_matrix.csv	Missense observed count
lof_obs_matrix.csv	Loss-of-function observed count

HPA RNA is split into 20 separate files, one per tissue (e.g. adipose_tissue_matrix.csv, cerebral_cortex_matrix.csv). Each file uses the same three-column schema with source = hpa_rna.

Curated Class Lists

For each classification source a curated list of class IDs is provided (under config/classes/). These lists filter to classes that are well-populated (≥20 genes), biologically interpretable, and non-redundant. Using these lists is recommended for fair benchmarking comparisons.

---

Dataset Creation

Source Selection

Sources were selected to cover complementary aspects of gene biology:

• Disease: Three complementary databases capturing clinically curated (ClinVar), expert-curated phenotype (HPO), and large-scale data-integrated (OpenTargets) disease associations.
• Function: Curated pathway databases (Reactome, WikiPathways) plus the Gene Ontology biological process hierarchy, which provides fine-grained functional annotation.
• Structure: Pfam and InterPro capture protein domain families; PANTHER captures broader evolutionary family groupings; MobiDB provides continuous disorder predictions.
• Expression: The Human Protein Atlas provides two complementary expression readouts — IHC protein detection at the tissue level (classification) and RNA consensus expression as log1p(nTPM) values (regression). Both are restricted to 20 major tissues for coverage and comparability.

Preprocessing

• OpenTargets: Filtered to association score ≥ 0.7 to retain high-confidence gene–disease links only.
• ClinVar: Filtered to GRCh38 assembly, pathogenic and likely-pathogenic classifications (ClinSigSimple == 1).
• HPO: NOT-qualified annotations excluded; restricted to OMIM disease IDs.
• Gene Ontology: Biological process (BP) aspect only; NOT-qualified annotations excluded.
• Pfam / InterPro: Streamed from large source files (4.7 GB / 16 GB respectively); filtered to human UniProt accessions.
• PANTHER: Subfamily suffixes stripped to family-level grouping.
• gnomAD: Six constraint metrics decomposed into separate regression targets.
• HPA Protein IHC: Restricted to 20 major tissues. For each (gene, tissue) pair, the maximum IHC level across all cell types is used. Ambiguous or non-standard level values (Ascending, Descending, N/A, Not representative) are dropped before aggregation. Genes where the maximum level is Not detected in a tissue are excluded from that tissue's positive set.
• HPA RNA: Restricted to 20 major tissues (matching HPA Protein). nTPM values are log1p-transformed (log1p(nTPM)) to reduce the dynamic range. One file is produced per tissue.

Known Limitations

• Human only — all data is restricted to Homo sapiens (taxon 9606).
• No train/validation/test splits are pre-computed — splits should be constructed by downstream users.
• Annotation incompleteness — negative labels are implicit (absence of annotation ≠ confirmed non-association); true negatives are not available.
• Identifier mapping loss — a fraction of genes cannot be mapped to Entrez IDs and are excluded.