From Snapshots to Trajectories: Benchmarking Senescence-Conditioned Cell Morphology Generation

AT A GLANCE

The first benchmark for senescence-conditioned cell morphology generation

SenoFlow combines 87,391 real single-cell DAPI nuclear crops from three public Xenium tissue panels with continuous senescence scores and computationally generated temporal sequences, creating a benchmark for studying how cell morphology changes along senescence progression.

0Real single-cell images

0Xenium tissue panels

0Cell types

0temporal sequences

Dataset

Dataset Demo

Real Cell Crops

TissueLung

Cell TypeEpithelial

Score0.18

TissueLung

Cell TypeT cells

Score0.47

TissueProstate

Cell TypeMyeloid

Score0.74

TissueSkin

Cell TypeMyeloid

Score0.91

Generated Temporal Sequences

Low senescence score High senescence score

sequence_id: gen_lung_epi_0001 | target scores: 0.00 to 1.00 (10 bins)

sequence_id: gen_pro_fib_0002 | target scores: 0.00 to 1.00 (10 bins)

Method

Method Summary

We adopt flow matching as the backbone for senescence-conditioned generation. We propose SD-Prototype OT to stabilize transport under heterogeneous cell types with continuous scores and Continuous-Score Velocity Correction to preserve biologically consistent trajectory progression.

Benchmark Results

Evaluation Protocol and Benchmark Comparison

Evaluation Protocols

Automatic Metrics

FID: measures realism of generated images via feature distributions.
KID: complements FID with more reliable estimation under small sample sizes.
ρ (sig): mean signed Spearman correlation on significant metric across cell types.
Effect: fraction of real-data morphological shift reproduced by the model.
Viol: mean monotonicity violation ratio.

Expert Study

Specialists in senescence and cell biology review generated trajectories in a blinded setting. Scores are reported on a 1-5 scale and averaged across experts and evaluated samples.

Realism: whether generated images resemble real cell nuclei.
Fidelity: whether progression reflects biologically plausible ageing dynamics.
preservation: whether original cell identity is maintained.

Benchmark Results

We evaluate 13 baselines and our method under the same per-cell-type protocol, grouped into non-FM generators and Flow Matching variants.

Method	FID	KID (×10³)	ρ (sig)	Effect (%)	Viol (%)	Realism	Fidelity	preservation
NON-FM GENERATORS
cVAE	71.70	44.29	0.020	3.0	-	-	-	-
CCDM	43.20	21.37	0.162	37.0	-	-	-	-
FLOW MATCHING VARIANTS
iMF	14.16	3.72	0.108	40.0	0.2	-	-	-
MeanFlow	17.95	6.05	0.115	80.0	39.3	-	-	-
Contrastive FM	26.24	10.83	0.307	55.7	38.5	-	-	-
Noise-start CFM	36.81	17.24	0.575	75.6	33.4	-	-	-
Rectified Flow	61.02	34.68	0.057	40.0	48.5	-	-	-
Untyped OT-CFM	66.94	40.15	0.202	64.6	47.1	-	-	-
I-CFM (random)	85.55	56.81	0.307	41.3	39.2	-	-	-
LogitNorm OT-CFM	92.96	62.43	0.070	48.7	48.0	-	-	-
Ours	11.11	2.14	0.570	60.2	23.2	-	-	-

Download

Core Resources

Dataset

Real nuclei crops + generated temporal sequences with per-cell-type protocol setup.

🤗 Hugging Face

Code

Training, inference, and evaluation scripts for our method.

GitHub Repo

Pretrained Models

Released checkpoints for reproducible benchmark comparison and downstream testing.

Model Weights

Citation

BibTeX

@article{zhang2026xensen,
  title   = {XenSen-Bench: A Multi-Tissue Benchmark for Cellular
             Senescence from Spatial Transcriptomics},
  author  = {Xiang Zhang and Boxuan Zhang and Ting Long and
             Haotian Zhuang and Jason Ji and Ruixiang Tang and
             Heng Fan and Dongfang Liu},
  journal = {NeurIPS Datasets and Benchmarks Track},
  year    = {2026}
}