SenoMorph

SenoMorph is a senescence-conditioned cell morphology benchmark of single-cell DAPI crops paired with continuous, transcriptome-derived senescence scores. The dataset accompanies an anonymous NeurIPS 2026 Datasets & Benchmarks submission.

Composition

Subset	Cells	Score-conditioned crops
Real (lung)	41,702	—
Real (prostate)	28,845	—
Real (skin)	16,844	—
Real total	87,391	—
Companion (lung)	41,702	417,020 (×10)
Companion (prost.)	28,845	288,450 (×10)
Companion (skin)	16,844	168,440 (×10)
Companion total	87,391	873,910

Each instance carries five attributes (paper §2 attribute card): tissue, cell_type_name, senescence_score, nuclear_area (mask pixel count, threshold > 10), and dapi_intensity (mean intensity within the mask). The image is a 64×64 grayscale uint8 DAPI crop.

Splits

70 / 10 / 20 train / val / test, stratified by cell type within each tissue. The companion set inherits its split from the source real cell.

Real subset

Each real cell has a senescence score produced by DeepScence from its matched single-cell transcriptome and min-max normalized to [0,1] within its tissue. Cell-type labels are produced by CellTypist.

Companion subset

For each real source cell, the model generates ten score-conditioned images at target scores evenly spaced over [0,1]. The companion subset is intended for downstream augmentation tasks (e.g., senescent-cell classification) and is not used as ground truth for any evaluation metric in the accompanying paper.

Loading

Available configs:

name=	Description
real	All three tissues, real DAPI cells
companion	All three tissues, generated companion cells
real_lung / real_prostate / real_skin	Real cells, single tissue
companion_lung / companion_prostate / companion_skin	Companion cells, single tissue

Each config has train, validation, and test splits.

from datasets import load_dataset

# Full real benchmark, training split (3 tissues combined)
real_train = load_dataset('anon-neuripsdb26/SenoMorph', name='real', split='train')

# Real lung test split only
lung_test = load_dataset('anon-neuripsdb26/SenoMorph', name='real_lung', split='test')

# Companion data (873,910 generated cells), train split
comp_train = load_dataset('anon-neuripsdb26/SenoMorph', name='companion', split='train')

# Companion skin only, validation split
skin_val_comp = load_dataset('anon-neuripsdb26/SenoMorph', name='companion_skin', split='validation')

Schema

image            PIL.Image (grayscale 64×64, PNG-encoded)
tissue           str  ('lung' | 'prostate' | 'skin')
cell_type_id     int32 (raw CellTypist integer)
cell_type_name   str  ('Epithelial cells' | 'T cells' | 'Fibroblasts' | ...)
senescence_score float32  in [0, 1]
                     real: DeepScence min-max norm within tissue
                     companion: target conditioning score from linspace(0,1,10)
nuclear_area     int32  (mask pixel count, threshold > 10)
dapi_intensity   float32 (mean intensity within mask)
split            str  ('train' | 'val' | 'test')
cell_id          int64
                     real: row index in source set
                     companion: source real cell id (10 companion rows share a cell_id)

License

Released under CC-BY-4.0. Raw Xenium outputs are publicly available from the 10x Genomics dataset portal. The dataset contains no personally identifiable information; 64×64 DAPI nuclear crops cannot be traced back to individual donors.

Anonymous

This release accompanies an anonymous NeurIPS 2026 Datasets & Benchmarks submission. Author identification is intentionally suppressed during review.