SubcellPert

SubcellPert
Benchmarking Subcellular Perturbation Prediction

AUTHORS & AFFILIATIONS

Jinghao Wang1 Hanqun Cao1 Xilin Dang1 Tianhui Zhou2 Yiyang Zhao3,4 Chang-Yu Hsieh3 Chunbin Gu1,* Pheng Ann Heng1

1 The Chinese University of Hong Kong 2 Duke University 3 Zhejiang University 4 Shanghai Innovation Institute

[email protected]

SubcellPert benchmarks perturbation-response prediction across 14 imaging datasets and eight baselines, extending virtual cell evaluation from gene counts to intracellular RNA organization.

A woman wraps herself in wind-blown blue-grey drapery in Waterhouse’s Boreas.
Boreas, 1903 · John William Waterhouse
14 datasets
13.88M cells in prepared corpus
3.65B retained RNA molecules
8 evaluated baselines
608 completed runs
The missing prediction object

A cell is more than a vector.

Standard perturbation models predict one number per gene. Imaging-based spatial transcriptomics records something richer: the identity and intracellular coordinate of every detected RNA molecule.

The target is an unpaired RNA population after an intervention or in a held-out cellular context.

FROM COUNTS TO COORDINATES

The same perturbation response. A new dimension.

Source and perturbed gene-count charts becoming intracellular RNA patterns A source profile and perturbed profile begin as gene-count dot bars. The dots move into two cells while an arrow labeled Perturbation connects the source cell to the perturbed cell. Each dot represents one RNA molecule and color indicates gene identity. Gene AGene B Gene CGene D Gene AGene B Gene CGene D RNA count RNA count SOURCE PERTURBED PERTURBATION Source cell · observed RNA coordinates Perturbed cell · predicted RNA coordinates
Gene AGene BGene CGene D

One dot = one RNA molecule. Color marks gene identity; perturbation changes abundance and intracellular organization.

The benchmark corpus

Benchmark

Three datasets evaluate perturbation OOD; eleven evaluate cellular-context OOD. Each dataset retains its own gene panel and grouping metadata.

Dataset selection and diversity
Dataset selection flow with organ and perturbation diversity
14 retained datasets: three Perturbation OOD and eleven Context OOD. Tissue slices count datasets; perturbation slices count dataset–type memberships.

Corpus prepared 24 Sep 2026: ≥50 cells per context–condition group. Dataset release: Dataset snapshot ↗.

3 DATASETS
Perturbation-OOD

Unseen intervention identities and registered combination stress tests.

01PF-Astro

Brain

Brain-derived astrocyte culture; single-guide CRISPRi.

RNA
10,670,839
Cells
11,454
Genes
485

02ATM-RT

Brain

Diffuse midline glioma; ATM loss and focal radiotherapy.

RNA
147,722,071
Cells
790,166
Genes
298

03Ddx54-PD1

Lung

Lung-tumor model; Ddx54 knockdown and anti-PD-1 antibody.

RNA
494,265,984
Cells
1,155,233
Genes
5,006
11 DATASETS
Context-OOD

Known responses evaluated in held-out cellular contexts.

04Lung-Drug

Lung

Formal drug family only: AZD3965 and VB124 added after bleomycin.

RNA
681,015,170
Cells
1,561,700
Genes
5,006

05Sickness

Brain

Hypothalamus; DSS chemical challenge and LPS/poly(I:C)/PLA2/STAg immune stimuli.

RNA
1,101,298,121
Cells
1,765,443
Genes
941

06DSS-Colitis

Intestine

Colon; DSS-induced colitis and recovery.

RNA
119,553,365
Cells
1,439,542
Genes
940

07Adipose

Adipose

Subcutaneous adipose before and after bariatric-surgery-associated weight loss.

RNA
1,907,755
Cells
13,972
Genes
477

08UC-VDZ

Intestine

Colon; vedolizumab antibody treatment.

RNA
19,552,710
Cells
80,562
Genes
1,000

09Kidney-IRI

Kidney

Kidney; surgical ischemia-reperfusion injury.

RNA
168,519,194
Cells
1,374,915
Genes
300

10Gut-Microbiota

Intestine

Small-intestine and colon regions; SPF to germ-free condition.

RNA
189,197,264
Cells
2,057,563
Genes
1,815

11Cardiac-MI

Heart

Heart; myocardial infarction induced by surgical coronary ligation.

RNA
61,134,068
Cells
225,023
Genes
475

12Breast-FUS

Breast

Breast cancer; focused ultrasound with or without gemcitabine.

RNA
102,102,171
Cells
880,565
Genes
1,000

13GBM-oHSV

Brain

Glioblastoma; oncolytic HSV treatment-associated recurrence.

RNA
520,974,197
Cells
2,401,940
Genes
480

14Spinal-KO

Spinal cord

Spinal cord; four developmental knockout responses.

RNA
36,073,815
Cells
122,645
Genes
347

Evaluation

TARGET COUNTS GIVEN

Oracle

Spatial W1 ↓ measures RNA localization error with exact target counts supplied.

COUNTS + COORDINATES PREDICTED

Joint

MSE ↓ and PCC-delta ↑ measure abundance. Spatial W1 ↓ measures localization on recovered genes, alongside gene coverage ↑.

Eight baselines · 38 frozen split instances · 304 Oracle + 304 Joint runs. Foundation-model encoders are frozen; adapters train for five epochs.

Inside the datasets

Within-dataset statistics

Observed RNA responses reveal how abundance, spatial organization, and population composition vary within each dataset.

Population comparisons
1,628
Condition-decoding contexts
426
Cell splits per decoding context
3
Six panels comparing abundance and spatial response magnitudes, condition-decoding gains, radial and angular changes, cluster contributions, and nuclear RNA fraction versus area-normalized enrichment.
Each point is a context–condition comparison, except panel b (one point per context). Colors identify datasets. Hollow circles flag limited observability, clustering stability, or gene support; they do not mark significance. View full figure ↗
A

Abundance and localization

Response magnitudes vary on matched genes. Separate gene-level correlations are +0.264 in Adipose adipocytes, +0.086 in GBM tumor cells, and −0.139 in GBM macrophages/monocytes: weak associations with different signs.

B

What spatial information adds

Median condition-decoding gains are +4.85 pp in Sickness (101/102 contexts improve), +0.67 pp in GBM-oHSV (7/8), and −3.01 pp in Spinal-KO.

C–D

Radius and angle respond differently

1,622 comparisons across 12 datasets show dataset-specific radial and angular patterns. DSS-Colitis has low angular observability, limiting interpretation of its apparent direction.

E

Composition versus within-cluster change

In Breast-FUS, within-cluster terms dominate radius changes in 66/68 comparisons; composition terms dominate angle in 51/68. These spatial clusters do not establish biological states.

F

Nuclear fraction depends on geometry

Four Adipose contexts have higher nuclear RNA fractions but lower enrichment after normalization by projected nuclear area. Fraction and area-normalized enrichment describe different aspects of the response.

Condition-decoding statistics · all 14 datasets

Adding spatial principal components (PCs) to abundance features; identical three stratified cell splits, λ = 1. Gains are medians of context-level mean balanced-accuracy differences, in percentage points (pp).

Download CSV ↓
DatasetContexts Mean gain > 064 PCs All splits > 064 PCs Gain: 64 PCspp Gain: 32 PCspp
Adipose600-1.672-1.540
ATM-RT100-0.115-0.123
Breast-FUS46198-0.901+0.033
Cardiac-MI50187-0.664-0.283
Ddx54-PD1100-0.120-0.174
DSS-Colitis800-0.091-0.027
GBM-oHSV875+0.666+0.591
Gut-Microbiota6652-1.240-0.589
Kidney-IRI20136+0.036+0.107
Lung-Drug1122-0.105-0.050
PF-Astro100-0.371-0.261
Sickness10210199+4.851+5.581
Spinal-KO8391-3.006-1.859
UC-VDZ23146+0.357+0.401

Positive counts are descriptive, with differences within 10⁻¹² balanced-accuracy units treated as ties. This analysis classifies observed conditions from held-out cells; it does not test prediction for unseen donors or perturbations.

Every dataset · every baseline · every metric

Results

608 runs across 14 datasets and eight baselines. Compare all-gene and top-100 DEG scores.

PF-Astro · Perturbation OOD · All genes · 3 data splits

PF-Astro · All genes · scores for all eight baselines
Baseline Oracle W1 ↓ Joint W1 ↓Recovered genes Coverage ↑ Joint MSE ↓ PCC-delta ↑
Resampling0.251 ± 0.00810.254 ± 0.008299.79% ± 0.042 pp0.00646 ± 0.00025-0.00043 ± 0.032
Gaussian0.253 ± 0.00690.257 ± 0.006399.79% ± 0.042 pp0.00834 ± 0.000540.033 ± 0.021
Radial0.298 ± 0.00250.267 ± 0.006999.79% ± 0.042 pp0.00834 ± 0.000540.033 ± 0.021
scGPT-spatial0.249 ± 0.00840.263 ± 0.01299.61% ± 0.29 pp0.0109 ± 0.0040.0347 ± 0.0059
SpatialFormer0.249 ± 0.00630.263 ± 0.006399.43% ± 0.17 pp0.0109 ± 0.00150.0385 ± 0.016
scGPT0.249 ± 0.00630.267 ± 0.01399.42% ± 0.26 pp0.0133 ± 0.00320.0482 ± 0.022
scFoundation0.249 ± 0.00630.254 ± 0.008299.79% ± 0.042 pp0.00646 ± 0.00025-0.00043 ± 0.032
STATE-SE0.248 ± 0.00660.256 ± 0.0199.67% ± 0.2 pp0.00778 ± 0.0022-0.00753 ± 0.02
Expanded SubcellPert benchmark overview

SubcellPert benchmark overview: 14 active datasets, eight evaluated baselines, and two evaluation modes.

Copied to clipboard