CytoWave: Perturbation-Centric Pretraining for Single-Cell Response Prediction
Authors: Yuwei Miao, Jingquan Yan, Hehuan Ma, Thao M. Dang, Lin Xu, Siyuan Zhang, Junzhou Huang
Organizations: Department of Computer Science and Engineering, University of Texas at Arlington, Arlington, TX, USA · UT Southwestern Medical Center, Dallas, TX, USA
Predicting gene expression under perturbations from single-cell data is a fundamental task for modeling cellular responses, with applications in discovering gene function and designing therapeutics. However, existing methods often struggle to generalize across datasets due to differences in experimental protocols, biological contexts, and species. We present CytoWave, a perturbation-centric pretraining framework for single-cell genetic perturbation response prediction. We leverage large-scale cross-species perturbation datasets unified via orthology and train CytoWave in a two-stage manner. In the first stage, we learn reference control cell states via masked gene expression recovery on unperturbed cells. In the second stage, we pretrain the model to predict perturbed gene expression given control states and perturbation signals, together with a contrastive objective that aligns latent representations of perturbation responses with perturbation signals. We perform strict dataset-held-out evaluation across multiple condition combinations and case studies of pathway activity changes, demonstrating consistent and effective performance of CytoWave in predicting single-cell genetic perturbation responses.
Figures & tables
Figure 1: CytoWave architecture and two-stage self-supervised pretraining. Stage 1 : Learning control representations via masked gene expression and recovery using a joint Transformer encoder with FiLM-conditioned gene embeddings. Stage 2 : Learning a genetic perturbation operator by jointly predicting cellular responses after perturbation (MSE loss) and contrastive perturbation alignment, where predicted perturbation embeddings are aligned with ground-truth masked perturbation embeddings with in-batch InfoNCE loss.
Figure 2: Overview of datasets collected for CytoWave. Bar lengths represent the number of control cells (lighter color) and perturbed cells (darker color) in each dataset. The central marker indicates the year of dataset publication (2016–2025).
MAE (lower is better)
Method
DatlingerBock2017
DatlingerBock2017(Stim)
PapalexiSatija2021
GasperiniShendure2019
SantinhaPlatt2023
BaseReg
2.016
1.996
2.200
2.547
2.057
Linear
2.038
2.003
2.200
2.627
2.057
TrainMean
2.007
2.000
2.214
2.519
2.057
BaseMLP
2.017 ± 0.015
1.976 ± 0.030
2.314 ± 0.006
2.632 ± 0.031
2.134 ± 0.014
CPA
1.080 ± 0.009
0.982 ± 0.008
0.997 ± 0.004
1.332 ± 0.005
1.001 ± 0.004
Table 1: Single-gene perturbation prediction across all datasets. We report MAE (lower is better) together with Δ PCC and Δ PCC 20 (higher is better).
Figure 3: Ground-truth and predicted pathway activity changes across multiple single-cell perturbation datasets. An illustrative subset of high-agreement pathways and perturbations is shown. Top panels denote ground truth, and bottom panels denote predictions.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: (A) shows the distribution of different datasets, the size of the circle represents the number of unique perturbations in this dataset. (B) The heatmap between lineage and cell context, numbers represent how many datasets are in each category combination. (C) The histogram for the number of datasets in each cell line. Colors indicate disease status, including cancer, healthy, and unknown.
Figure 5: The gene coverage of training data with sparse tokenization
#Unique
Dataset
#Cells
#Genes
#Orthologs
Perturbations
#Control Cells
#Perturbed Cells
ZhuMarson2025_D1_Rest
1633989
16801
16801
11332
76543
1557446
ZhuMarson2025_D1_Stim48h
1534912
16815
16815
11403
73534
1461378
ZhuMarson2025_D1_Stim8h
1498205
16804
16804
11417
69632
1428573
ZhuMarson2025_D2_Rest
1851226
16805
16805
11307
83987
1767239
ZhuMarson2025_D2_Stim48h
1912736
16816
16816
11316
87747
1824989
Appendix
Table 2: Scale and perturbation coverage statistics for the full curated corpus before the dataset-held-out partitioning. Datasets used for downstream evaluation are listed for completeness but are excluded from all pretraining stages.
Dataset
Species
Cell Context
Cell Line
Lineage
Disease
ZhuMarson2025_D1_Rest
Human
Primary
–
Hematopoietic
Healthy
ZhuMarson2025_D1_Stim48h
Human
Primary
–
Hematopoietic
Healthy
ZhuMarson2025_D1_Stim8h
Human
Primary
–
Hematopoietic
Healthy
ZhuMarson2025_D2_Rest
Human
Primary
–
Hematopoietic
Healthy
ZhuMarson2025_D2_Stim48h
Human
Primary
–
Hematopoietic
Healthy
ZhuMarson2025_D2_Stim8h
Human
Primary
–
Hematopoietic
Healthy
Appendix
Table 3: Biological and experimental context metadata across all curated datasets (with evaluation datasets held out from pretraining).
Dataset
Perturbation Technology
Stimulation
Stimulation Time
ZhuMarson2025_D1_Rest
CRISPRi
None
–
ZhuMarson2025_D1_Stim48h
CRISPRi
TCR
48h
ZhuMarson2025_D1_Stim8h
CRISPRi
TCR
8h
ZhuMarson2025_D2_Rest
CRISPRi
None
–
ZhuMarson2025_D2_Stim48h
CRISPRi
TCR
48h
ZhuMarson2025_D2_Stim8h
CRISPRi
TCR
8h
Appendix
Table 4: Perturbation technologies and stimulation conditions across all curated datasets (with evaluation datasets held out from pretraining)
Dataset
Train
Valid
Test
#Unique Perturbations
Mean Cells / Perturb
DatlingerBock2017
1634
191
206
32
63.47
DatlingerBock2017 (stim)
2022
255
277
32
79.81
GasperiniShendure2019
3982
397
747
286
17.92
PapalexiSatija2021
3377
420
427
6
704.00
SantinhaPlatt2023
30910
3854
3889
25
1546.12
Appendix
Table 5: Statistics of evaluation datasets used for downstream perturbation prediction tasks.
Figure 6: The top hallmark 50 absolute pathway activity of the perturbed ground truth and the predicted results, y-axis is hallmarks and x-axis is perturbation group in each dataset
1L3S Research Center, Leibniz Universität Hannover, Germany · School of Clinical Medicine & Laboratory Medicine, Jiangsu University, China · Institute for Information Processing (tnt), Leibniz Universität Hannover, Germany.
School of Computer Science and Engineering, South China University of Technology, Guangzhou, China · School of Future Technology, South China University of Technology, Guangzhou, China
Australian Institute for Machine Learning, Adelaide University, Australia · Responsible AI Research Centre, Australia · College of Computer Science and Artificial Intelligence, Fudan University, China +3