marcin119a/ISB-CGC-MultiOmics

ISB-CGC MultiOmics

A multi-omics cancer dataset derived from ISB-CGC open-access data (9 programs, 54 annotated projects, 20,825 cases) with missing modalities, prepared for benchmarking multimodal imputation and missing-modality models (SoftMoE, MIMIR).

Structure

Main_Dataset/
├── Imputation_datasets/
│   ├── ISB-All/    # 20,825 samples: rna, mirna, methylation (M-value), cnv, metadata
│   ├── ISB-TCGA/   # subset matching the TCGA benchmark (8,034 samples)
│   └── ISB-RMM/    # rna + miRNA + methylation (20,671 samples)
├── Splits/         # splits_isb*.json: train/val/test lists of case_barcode
└── Clinical/       # clinical_survival.csv (TCGA only)
load_isb.py         # loader

Format

• Each modality is a separate Parquet file (float32, z-scored features): row = case_barcode, column = feature. Samples lacking a modality have no row in that file.
• metadata.parquet: case_barcode, program, project_id, cancer_type, disease_type, primary_site, split (train/val/test), and presence flags has_rna, has_mirna, has_meth, has_cnv. 2,541 cases have no project annotation (null program / project_id).
• Splits (ISB-All: 14,575 / 3,134 / 3,116) are ~70/15/15, stratified by project_id (seed 42). Projects with at least 3 samples have at least one sample in val and test; projects with 1-2 samples go to train.
• clinical_survival.csv covers TCGA only and is keyed by sample_id / patient_id (not case_barcode).

Modality	Features	Samples	Coverage
rna	3,008	20,040	96%
miRNA	262	14,445	69%
methylation (M-value)	3,141	10,979	53%
cnv	3,105	8,034	39%

CNV is available for TCGA only; methylation is available almost exclusively for TCGA.

Program	Cases
TCGA	11,150
TARGET	3,356
CPTAC	1,685
BEATAML1.0	662
CGCI	524
NCICCR	481
HCMI	278
WCDT	99
ORGANOID	49
(no annotation)	2,541

Usage

from load_isb import load_isb

ds = load_isb(".", "ISB-All")   # dict: metadata, rna, mirna, methylation, cnv
ds["rna"].shape                  # (20825, 3008); missing samples are NaN rows

Or load a single modality with 🤗 datasets:

from datasets import load_dataset
rna = load_dataset("marcin119a/ISB-CGC-MultiOmics", "rna")

Licensing and attribution

This dataset contains only open-access, de-identified, derived data (z-scored feature matrices; no raw reads, no controlled-access data), obtained via ISB-CGC from the NCI Genomic Data Commons (GDC) and Proteomic Data Commons (PDC). It carries no single license: use is governed by the NIH Genomic Data Sharing Policy and the terms of the source programs below. By using it you agree to:

1. Not attempt to identify individual research participants.
2. Acknowledge the datasets and the repositories in publications: the NCI Genomic Data Commons (GDC), the NIH, and ISB-CGC.
3. Cite the source programs:
   • TCGA: follow the NCI citation guidelines. Publication moratoria have been lifted.
   • CPTAC (CC-BY 4.0 via PDC): cite the primary CPTAC publication and include "Data used in this publication were generated by the National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC)."
   • TARGET, BEATAML1.0, HCMI, WCDT, CGCI, NCICCR, ORGANOID: NIH GDS Policy; cite the program and its primary publications as listed on the GDC.

The MMRF CoMMpass program is deliberately excluded because its terms may restrict redistribution. This is not legal advice; verify the terms of each source program before reusing the data beyond research.