OmicsGBench: Benchmarking Graph Neural Networks on Omics Data

OmicsGBench is the first benchmark suite for graph-level prediction in the n ≪ p regime characteristic of omics data, where the number of patient samples n is much smaller than the number of nodes (genes or proteins) p per graph.

Datasets

This repository contains six preprocessed omics graph classification datasets, grouped by modality:

Dataset	Modality	n	p	Task
HERITAGE	Proteomics	654	4,976	Exercise responder (binary)
Tuberculosis	Proteomics	561	3,814	Culture negative vs positive
Parkinson's	Transcriptomics	535	21,755	Cognitive status (binary)
AddNeuroMed	Transcriptomics	711	17,197	Clinical diagnosis (3-class)
Smoking	Epigenomics	464	139,125	Never- vs ever-smoker
BRCA	Epigenomics	640	19,049	Cancer subtype (4-class)

Tuberculosis is a full-proteome antibody microarray (immunoproteomics), where each feature is one M. tuberculosis protein. Smoking stores 139,125 TSS1500/TSS200 methylation probes; training keeps one probe per gene, about 20,763 genes.

Source Data

• HERITAGE: Robbins et al. (2021), Nature Metabolism. Available via MoTrPAC Data Hub (motrpac-data.org) under CC-BY 4.0.
• Tuberculosis: Kunnath-Velayudhan et al. (2010), PNAS. Available via NCBI GEO (GSE19433) under GEO public data access policy.
• Parkinson's: Shamir et al. (2017), Neurology. Available via NCBI GEO (GSE99039) under GEO public data access policy.
• AddNeuroMed: Lovestone et al. (2009). Available via NCBI GEO (GSE63063) under GEO public data access policy.
• Smoking: Tsaprouni et al. (2014), Epigenetics. Available via NCBI GEO (GSE50660) under GEO public data access policy.
• BRCA: Yang et al. (2025), MLOmics, Scientific Data. Available on Figshare/Hugging Face under CC-BY 4.0.

Preprocessing

Datasets share one ingest pipeline: probe-to-gene aggregation where the source is probe-level, and the normalization described in Appendix B of the accompanying paper. Matrices on the Hub are uncorrected. HERITAGE covariate adjustment, AddNeuroMed ComBat, and smoking promoter-probe selection plus median-centering are fit on the training split only.

The default evaluation is stratified 5-fold. A 70/15/15 train/val/test split is available as split_type=fixed. Parkinson's and Tuberculosis keep each batch inside one fold: hybridization date for Parkinson's, country of serum collection for Tuberculosis.