jang1563/genelab-benchmark

GeneLab Spaceflight Transcriptomics Benchmark

A comprehensive benchmark for evaluating ML models and foundation models on NASA spaceflight transcriptomics data.

Version: v6.0 | Dataset freeze: 2026-03-01 | Code: GitHub

---

Overview

GeneLab Benchmark provides standardized train/test splits for evaluating how well machine learning models generalize spaceflight transcriptomic signatures across ISS missions.

Core challenge: Given RNA-seq from one spaceflight mission, can a classifier detect spaceflight vs. ground control in samples from a different mission it has never seen?

Data source: NASA Open Science Data Repository (OSDR) -- mouse multi-tissue bulk RNA-seq from ISS rodent research missions (C57BL/6J strain).

Scope

Dimension	Coverage
Tissues	8 (Liver, Gastrocnemius, Kidney, Thymus, Skin, Eye, Lung, Colon)
ISS Missions	9 (RR-1, RR-3, RR-5, RR-6, RR-7, RR-8, RR-9, MHU-1, MHU-2)
OSD Studies	24
Samples	660+ (binary: Flight vs. Ground Control)
Classifiers	8 (PCA-LR, ElasticNet-LR, RF, XGBoost, SVM-RBF, KNN, MLP, TabNet)
Feature types	4 (Gene, Hallmark pathways, KEGG pathways, Combined)
Foundation Models	5 (Geneformer, scGPT, UCE, scFoundation, Text LLMs)

---

Dataset Structure

This HuggingFace repository contains feature matrices (train_X.csv, test_X.csv) for 4 benchmark tasks that passed significance thresholds. Labels and metadata are in the GitHub repository.

genelab-benchmark/
├── A2_gastrocnemius_lomo/        <- 3 missions, 32 samples
│   ├── fold_RR-1_test/
│   │   ├── train_X.csv           <- Training features (samples x genes)
│   │   └── test_X.csv            <- Test features
│   ├── fold_RR-5_test/
│   └── fold_RR-9_test/
│
├── A4_thymus_lomo/               <- 4 missions, 67 samples
│   ├── fold_MHU-1_test/
│   ├── fold_MHU-2_test/
│   ├── fold_RR-6_test/
│   └── fold_RR-9_test/
│
├── A5_skin_lomo/                 <- 3 missions, 102 samples
│   ├── fold_MHU-2_test/
│   ├── fold_RR-6_test/
│   └── fold_RR-7_test/
│
├── A6_eye_lomo/                  <- 3 missions, 37 samples
│   ├── fold_RR-1_test/
│   ├── fold_RR-3_test/
│   └── fold_TBD_test/            <- OSD-397, no official mission name
│
├── v4/evaluation/                <- Multi-method evaluation results (JSON)
├── v5/evaluation/                <- Systems biology analysis results
└── v6/evaluation/                <- Human translation analysis results

Each fold holds out one mission as the test set and trains on the remaining missions. This Leave-One-Mission-Out (LOMO) cross-validation evaluates true cross-mission generalization.

---

File Format

Feature matrix (train_X.csv, test_X.csv)

• Rows: Sample IDs
• Columns: Ensembl mouse gene IDs (e.g., ENSMUSG00000021969)
• Values: Log2(DESeq2 size-factor normalized counts + 1)
• Gene selection: Top 75th percentile variance, computed on training missions only (no test leakage)
• Typical shape: ~20,000 genes per sample

Labels (in GitHub repo)

Value	Meaning
1	Flight (spaceflight / microgravity)
0	Ground Control (vivarium / ground control)

---

Results

Multi-Method Evaluation (256 evaluations: 8 tissues x 8 methods x 4 feature types)

Best AUROC per tissue across all method-feature combinations, evaluated via LOMO cross-validation (6 tissues) or 5-fold stratified CV (Lung, Colon -- single-mission datasets):

Tissue	Best AUROC	Method	Feature	perm_p	Significant
Thymus	0.948	PCA-LR	KEGG	0.031	Yes
Colon	0.921	PCA-LR	KEGG	0.033	Yes
Lung	0.901	ElasticNet-LR	Gene	0.028	Yes
Gastrocnemius	0.898	ElasticNet-LR	Gene	0.058	-
Kidney	0.829	PCA-LR	Hallmark	0.010	Yes
Eye	0.823	PCA-LR	Hallmark	0.042	Yes
Skin	0.819	ElasticNet-LR	Gene	0.004	Yes
Liver	0.766	ElasticNet-LR	KEGG	0.093	-

Significance: permutation p < 0.05. Overall, 6/8 tissues significant; 40/256 individual evaluations significant.

PCA-LR on gene features provides a strong baseline (8-tissue mean AUROC = 0.776).

Foundation Model Comparison (7 tissues)

Model	Best Single-Tissue AUROC	vs PCA-LR Baseline (0.776)
scGPT	0.667 (mean)	Below baseline
scFoundation	0.635 (liver, p<0.01)	Below baseline
UCE	0.632 (thymus, p=0.031)	Below baseline
Mouse-Geneformer	0.476 (mean)	Below baseline
Text LLMs (GPT-4o, Claude, Llama 3)	0.47-0.51	Chance level

All foundation models underperform classical PCA-LR. Pre-trained cell atlas representations do not improve spaceflight detection.

Negative Controls

• Permutation control: AUROC = 0.50 +/- 0.03 (label shuffling)
• Housekeeping gene control: AUROC = 0.49-0.55 (non-informative features)
• Held-out validation: Thymus RR-23 (0.905), Skin RR-7 (0.885)

---

Downloading

Option A: Python API (recommended)

from huggingface_hub import hf_hub_download
import pandas as pd

# Download one fold's feature matrix
train_X = pd.read_csv(
    hf_hub_download(
        repo_id="jang1563/genelab-benchmark",
        filename="A5_skin_lomo/fold_RR-7_test/train_X.csv",
        repo_type="dataset",
    ),
    index_col=0
)
print(train_X.shape)  # (72, 20110)

Option B: Download full task

from huggingface_hub import snapshot_download

# Download all A5 skin files
snapshot_download(
    repo_id="jang1563/genelab-benchmark",
    repo_type="dataset",
    allow_patterns="A5_skin_lomo/**",
    local_dir="./data/benchmark",
)

---

Evaluation

Evaluate predictions using the included script (GitHub):

# Prepare submission JSON
submission = {
    "task_id": "A5",
    "model_name": "MyModel_v1",
    "predictions": {
        "fold_MHU-2_test": {"sample_id_1": 0.92, "...": "..."},
        "fold_RR-6_test":  {"...": "..."},
        "fold_RR-7_test":  {"...": "..."}
    }
}

# Run evaluation (requires cloning GitHub repo for labels)
# python scripts/evaluate_submission.py --submission my_submission.json --task A5

Model Tracks

Track	Examples	Input Format
Classical ML	LR, RF, XGBoost, PCA-LR, SVM, KNN, MLP, TabNet	Tabular (gene x sample)
Foundation Models	Geneformer, scGPT, UCE, scFoundation	Gene rank order / embeddings
Text LLMs	GPT-4o, Claude, Llama 3	Natural language gene list

---

Source Data

All data derived from publicly available NASA OSDR datasets:

OSD ID	Tissue	Mission	n (Flight + Ground)
OSD-48	Liver	RR-1	18
OSD-137	Liver	RR-3	20
OSD-245	Liver	RR-6	48
OSD-379	Liver	RR-8	40
OSD-242	Liver	RR-9	39
OSD-686	Liver	MHU-2	28
OSD-101	Gastrocnemius	RR-1	12
OSD-401	Gastrocnemius	RR-5	12
OSD-326	Gastrocnemius	RR-9	8
OSD-102	Kidney	RR-1	47
OSD-163	Kidney	RR-3	32
OSD-253	Kidney	RR-7	39
OSD-289	Thymus	MHU-2	12
OSD-244	Thymus	RR-6	35
OSD-421	Thymus	RR-9	20
OSD-238	Skin (dorsal)	MHU-2	18
OSD-239	Skin (femoral)	MHU-2	17
OSD-243	Skin	RR-6	37
OSD-254	Skin	RR-7	30
OSD-100	Eye	RR-1	12
OSD-194	Eye	RR-3	9
OSD-397	Eye	OSD-397	16
OSD-248	Lung	RR-6	39
OSD-247	Colon	RR-6	36

Lung and Colon additionally include Basal Control samples treated as ground control (+19 and +18 respectively).

---

Preprocessing

1. DESeq2 size-factor normalization (per mission)
2. Log2(counts + 1) transformation
3. Low-expression filter (>=20% samples with count > 1)
4. Top 75th percentile variance gene selection per fold, training missions only (prevents test leakage)
5. Pathway scores: gseapy ssGSEA (MSigDB Hallmark, KEGG gene sets)

---

Citation

(Manuscript in preparation)

@dataset{kang2026genelab,
  title   = {GeneLab Benchmark: A Multi-Tissue Spaceflight Transcriptomics
             Benchmark for AI/ML Models},
  author  = {Kang, Jaeyoung},
  year    = {2026},
  url     = {https://huggingface.co/datasets/jang1563/genelab-benchmark},
  note    = {v6.0}
}

Data source: NASA Open Science Data Repository (OSDR) -- https://osdr.nasa.gov/bio/repo/

---

License

• Dataset: CC-BY-4.0
• Code: MIT (GitHub repository)
• Source data: NASA OSDR (public domain)