Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; the references it cites are not checked yet.

PerturbNet predicts single-cell responses to unseen chemical and genetic perturbations

Open access
Yu, Hengshi, Qian, Weizhou, Song, Yuxuan, Welch, Joshua D · 2025 · Molecular Systems Biology

Published version of a preprint

Abstract

Chemical and genetic perturbations, such as those induced by small molecules and CRISPR, effect complex changes in the molecular states of cells. Despite advances in high-throughput single-cell perturbation screening technology, the space of possible perturbations is far too large to measure exhaustively. Here, we introduce PerturbNet, a flexible deep generative model designed to predict the distribution of cell states induced by unseen chemical or genetic perturbations. PerturbNet accurately predicts gene expression changes in response to unseen small molecules based on their chemical structures while also accounting for key covariates such as dosage and cell type. Moreover, PerturbNet accurately predicts the distribution of single-cell gene expression states following CRISPR activation or CRISPR interference by leveraging gene functional annotations. Our approach significantly outperforms previous methods, particularly for predicting the effects of perturbing completely unseen genes. Finally, we demonstrate for the first time that amino acid sequence embeddings can be used to predict gene expression changes induced by missense mutations. We use PerturbNet to predict the effects of all point mutations in GATA1 and nominate variants that significantly impact the cell state distribution of human hematopoietic stem cells. Using a crystal structure of GATA1 bound to DNA, we validate that these large-effect variants occur in the core DNA-contact region of GATA1 and tend to involve large changes in amino acid side-chain volume.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s44320-025-00131-3

Cites

Statistics

Figures

Image 1
CC BY · PubMed Central
Image 2
CC BY · PubMed Central
Figure 1 PerturbNet maps perturbation representations to cell states. (A) PerturbNet uses two neural networks separately trained to encode large numbers of chemical or genetic perturbations (left) and cell profiles (right) into latent spaces. A conditional invertible neural network (cINN) learns to map points in perturbation space to cell state space using high-throughput measurements of perturbation effect. (B) PerturbNet can then predict the gene expression changes induced by an unseen perturbation by encoding the perturbation, passing its representation through the cINN, and decoding the resulting cell states.
Figure 1 PerturbNet maps perturbation representations to cell states. (A) PerturbNet uses two neural networks separately trained to encode large numbers of chemical or genetic perturbations (left) and cell profiles (right) into latent spaces. A conditional invertible neural network (cINN) learns to map points in perturbation space to cell state space using high-throughput measurements of perturbation effect. (B) PerturbNet can then predict the gene expression changes induced by an unseen perturbation by encoding the perturbation, passing its representation through the cINN, and decoding the resulting cell states.CC BY · PubMed Central
Figure 2 PerturbNet predicts response to small-molecule treatment. (A) Diagram of the chemical variational autoencoder (ChemicalVAE) architecture for encoding small molecules represented as SMILES strings. The network was trained on the ZINC dataset, consisting of approximately 250,000 drug-like molecules. No gene expression information is used for training the drug representation network. Note that a cell representation network is also trained on gene expression values (not shown here). (B) Visualization of PerturbNet predictions for two distinct perturbations from LINCS-Drug dataset. The UMAP coordinates are computed from the latent spaces of the perturbation network (left) and cell state network (right). The mapping function learned by the cINN is indicated with lines connecting the perturbation and cell state representations. The predicted cell state distributions are also indicated with contour lines. (C) Box plots of R2 calculated on all genes and top 50 differentially expressed genes (DEGs) in one test split of LINCS-Drug dataset for chemCPA and PerturbNet with (-iso) and without stereoisomers shared between train and test sets (n = 100, plots for additional splits are shown in the supplement). (D) Heatmap showing the proportion of train/test splits in which PerturbNet or chemCPA (trained with stereoisomers) outperforms the same model trained without stereoisomers across five test splits of the LINCS-Drug dataset. “Outperforms” here means a significant difference by a one-sided Wilcoxon test. (E) Box plots of evaluation metrics in one test split of LINCS-Drug dataset (n = 1889, plots for additional splits are shown in the supplement). (F) Box plots of evaluation metrics in one test split of sci-Plex dataset (n = 38, plots for additional splits are shown in the supplement). The y axis is truncated at 0. (G) Heatmap showing the proportion of train/test splits in which PerturbNet achieves significantly better performance (based on the one-sided Wilcoxon tests) compared to competing models (columns) across five train/test splits of sci-Plex and LINCS-Drug dataset. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.
Figure 2 PerturbNet predicts response to small-molecule treatment. (A) Diagram of the chemical variational autoencoder (ChemicalVAE) architecture for encoding small molecules represented as SMILES strings. The network was trained on the ZINC dataset, consisting of approximately 250,000 drug-like molecules. No gene expression information is used for training the drug representation network. Note that a cell representation network is also trained on gene expression values (not shown here). (B) Visualization of PerturbNet predictions for two distinct perturbations from LINCS-Drug dataset. The UMAP coordinates are computed from the latent spaces of the perturbation network (left) and cell state network (right). The mapping function learned by the cINN is indicated with lines connecting the perturbation and cell state representations. The predicted cell state distributions are also indicated with contour lines. (C) Box plots of R2 calculated on all genes and top 50 differentially expressed genes (DEGs) in one test split of LINCS-Drug dataset for chemCPA and PerturbNet with (-iso) and without stereoisomers shared between train and test sets (n = 100, plots for additional splits are shown in the supplement). (D) Heatmap showing the proportion of train/test splits in which PerturbNet or chemCPA (trained with stereoisomers) outperforms the same model trained without stereoisomers across five test splits of the LINCS-Drug dataset. “Outperforms” here means a significant difference by a one-sided Wilcoxon test. (E) Box plots of evaluation metrics in one test split of LINCS-Drug dataset (n = 1889, plots for additional splits are shown in the supplement). (F) Box plots of evaluation metrics in one test split of sci-Plex dataset (n = 38, plots for additional splits are shown in the supplement). The y axis is truncated at 0. (G) Heatmap showing the proportion of train/test splits in which PerturbNet achieves significantly better performance (based on the one-sided Wilcoxon tests) compared to competing models (columns) across five train/test splits of sci-Plex and LINCS-Drug dataset. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.CC BY · PubMed Central
Figure EV1 Box plots of evaluation metrics on the LINCS-Drug and sci-Plex datasets. (A) Comparison between PerturbNet and chemCPA trained with and without stereoisomers. (B) Benchmark results on the LINCS-Drug dataset. (C) Benchmark results on the sci-Plex dataset. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers.
Figure EV1 Box plots of evaluation metrics on the LINCS-Drug and sci-Plex datasets. (A) Comparison between PerturbNet and chemCPA trained with and without stereoisomers. (B) Benchmark results on the LINCS-Drug dataset. (C) Benchmark results on the sci-Plex dataset. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers.CC BY · PubMed Central
Figure 3 PerturbNet predicts response to genetic perturbation. (A) Diagram of the GenotypeVAE architecture. Perturbations are represented in terms of their gene ontology annotations. The network is trained on all one- and two-gene combinations (~177 million). (B) Box plots of R2 values for unseen genetic perturbations, calculated on all genes and the top 50 differentially expressed genes (DEGs) on one test split of the Norman et al dataset (n = 33, plots for additional splits are in the supplement). “PerturbNet-raw-count” utilizes a zero-inflated negative binomial (ZINB) likelihood for the cell representation network, while “PerturbNet-normalized” employs a Gaussian likelihood. (C) Box plots of R2 values for unseen genetic perturbations, calculated on all genes and the top 50 DEGs on one test split of the Norman et al dataset. (n = 46, plots for additional splits are shown in the supplement). (D) Heatmap showing the fraction of train/test splits in which PerturbNet achieves significantly better performance (one-sided Wilcoxon tests) compared to competing models (rows) on the Norman et al dataset. (E, F) Scatter plots of R2 values for unseen genetic perturbations, calculated across all genes, the top 50 differentially expressed genes (DEGs), and large-effect genes from the Norman et al dataset. Data points are aggregated from all five test splits. Different colors represent the “number of seen genes” Labels such as “0/1” indicates that the test perturbation affects one unseen gene, while “0/2” indicates that two unseen genes are perturbed. “1/2” denotes that two genes are perturbed, but one of the target effects has already been observed independently or in combination with other genetic perturbations. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.
Figure 3 PerturbNet predicts response to genetic perturbation. (A) Diagram of the GenotypeVAE architecture. Perturbations are represented in terms of their gene ontology annotations. The network is trained on all one- and two-gene combinations (~177 million). (B) Box plots of R2 values for unseen genetic perturbations, calculated on all genes and the top 50 differentially expressed genes (DEGs) on one test split of the Norman et al dataset (n = 33, plots for additional splits are in the supplement). “PerturbNet-raw-count” utilizes a zero-inflated negative binomial (ZINB) likelihood for the cell representation network, while “PerturbNet-normalized” employs a Gaussian likelihood. (C) Box plots of R2 values for unseen genetic perturbations, calculated on all genes and the top 50 DEGs on one test split of the Norman et al dataset. (n = 46, plots for additional splits are shown in the supplement). (D) Heatmap showing the fraction of train/test splits in which PerturbNet achieves significantly better performance (one-sided Wilcoxon tests) compared to competing models (rows) on the Norman et al dataset. (E, F) Scatter plots of R2 values for unseen genetic perturbations, calculated across all genes, the top 50 differentially expressed genes (DEGs), and large-effect genes from the Norman et al dataset. Data points are aggregated from all five test splits. Different colors represent the “number of seen genes” Labels such as “0/1” indicates that the test perturbation affects one unseen gene, while “0/2” indicates that two unseen genes are perturbed. “1/2” denotes that two genes are perturbed, but one of the target effects has already been observed independently or in combination with other genetic perturbations. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.CC BY · PubMed Central
Figure EV2 Box plots of evaluation metrics on the genetic perturbation datasets. (A) Comparison between PerturbNet trained on raw counts and on normalized expression. (B) Benchmark results on the Norman et al dataset. (C) Benchmark results on the Ursu et al and Jorge et al datasets. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers.
Figure EV2 Box plots of evaluation metrics on the genetic perturbation datasets. (A) Comparison between PerturbNet trained on raw counts and on normalized expression. (B) Benchmark results on the Norman et al dataset. (C) Benchmark results on the Ursu et al and Jorge et al datasets. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers.CC BY · PubMed Central
Figure EV3 Scatter plots of R2 values for unseen genetic perturbations, calculated across all genes, the top 50 differentially expressed genes (DEGs), and large-effect genes from the Norman et al dataset. Data points are aggregated from all five test splits. Different colors represent the “number of seen genes.” Labels such as “0/1” indicate that the test perturbation affects one unseen gene, while “0/2” indicates two unseen genes are perturbed. “1/2” denotes that two genes are perturbed, but one of the target effects has already been observed independently or in combination with other genetic perturbations. (A) PerturbNet v.s linear baseline. (B) PerturbNet v.s KNN. (C) PerturbNet v.s. training sample.
Figure EV3 Scatter plots of R2 values for unseen genetic perturbations, calculated across all genes, the top 50 differentially expressed genes (DEGs), and large-effect genes from the Norman et al dataset. Data points are aggregated from all five test splits. Different colors represent the “number of seen genes.” Labels such as “0/1” indicate that the test perturbation affects one unseen gene, while “0/2” indicates two unseen genes are perturbed. “1/2” denotes that two genes are perturbed, but one of the target effects has already been observed independently or in combination with other genetic perturbations. (A) PerturbNet v.s linear baseline. (B) PerturbNet v.s KNN. (C) PerturbNet v.s. training sample.CC BY · PubMed Central
Figure 4 PerturbNet predicts response to coding sequence mutation. (A) Diagram of the approach for training a representation network for coding sequence mutations. Each perturbation is an amino acid sequence edited by CRISPR. We used an evolutionary scale modeling (ESM) transformer pretrained on the UniParc database, containing 250M sequences. (B) Workflow for predicting the effects of all possible point coding variants along a specific protein sequence. PerturbNet is trained using all available coding variants and then used to generate cell distributions for every possible protein variant at each position within the protein sequence. (C) Bar plots showing the cell-type proportions (left) and mean volume changes induced by coding variants (right) for the three predicted variant clusters along the GATA1 sequences. (D) Box plots of R2 values for unseen genetic perturbations (coding variants) calculated on the large-effect genes in the dataset from Ursu et al (n = 10, TP53 + KRAS) and Jorge et al (n = 35, GATA1). The y axis is truncated at 0. (E) Heatmap showing the proportion of train/test splits for which PerturbNet significantly outperforms (one-sided Wilcoxon tests) other models (columns). (F) UMAP visualization of predicted cell distributions for all possible point coding variants and observed coding variants on the GATA1 protein, with three distinct perturbation clusters highlighted. (G) Cell density of each perturbation cluster in the UMAP. (H) The top 10 predicted large-effect coding variants are located at or near critical GATA1–DNA interacting regions. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.
Figure 4 PerturbNet predicts response to coding sequence mutation. (A) Diagram of the approach for training a representation network for coding sequence mutations. Each perturbation is an amino acid sequence edited by CRISPR. We used an evolutionary scale modeling (ESM) transformer pretrained on the UniParc database, containing 250M sequences. (B) Workflow for predicting the effects of all possible point coding variants along a specific protein sequence. PerturbNet is trained using all available coding variants and then used to generate cell distributions for every possible protein variant at each position within the protein sequence. (C) Bar plots showing the cell-type proportions (left) and mean volume changes induced by coding variants (right) for the three predicted variant clusters along the GATA1 sequences. (D) Box plots of R2 values for unseen genetic perturbations (coding variants) calculated on the large-effect genes in the dataset from Ursu et al (n = 10, TP53 + KRAS) and Jorge et al (n = 35, GATA1). The y axis is truncated at 0. (E) Heatmap showing the proportion of train/test splits for which PerturbNet significantly outperforms (one-sided Wilcoxon tests) other models (columns). (F) UMAP visualization of predicted cell distributions for all possible point coding variants and observed coding variants on the GATA1 protein, with three distinct perturbation clusters highlighted. (G) Cell density of each perturbation cluster in the UMAP. (H) The top 10 predicted large-effect coding variants are located at or near critical GATA1–DNA interacting regions. Note: for all box plots in this panel, the box plots show the median (center line), the 25th and 75th percentiles (box bounds), and 1.5× the interquartile range (whiskers). Points beyond the whiskers are plotted as outliers. Source data are available online for this figure.CC BY · PubMed Central
Figure EV4 Attributing cell state shifts to specific features of perturbations. (A) Diagram of approach for attributing perturbation outcomes to specific perturbation features. We attach a binary classifier after the cINN to classify cells into discrete types. By comparing the classification results of an input perturbation and a baseline (random perturbations), we can determine which perturbation features increase classification probability. (B) UMAP plots of cells from LINCS-Drug colored by cluster label. (C, D) UMAP plots of LINCS-Drug with selected clusters and the selected drug colored by attribution scores for each atom. The bottom is the UMAP plot colored by the cell density from selected drugs.
Figure EV4 Attributing cell state shifts to specific features of perturbations. (A) Diagram of approach for attributing perturbation outcomes to specific perturbation features. We attach a binary classifier after the cINN to classify cells into discrete types. By comparing the classification results of an input perturbation and a baseline (random perturbations), we can determine which perturbation features increase classification probability. (B) UMAP plots of cells from LINCS-Drug colored by cluster label. (C, D) UMAP plots of LINCS-Drug with selected clusters and the selected drug colored by attribution scores for each atom. The bottom is the UMAP plot colored by the cell density from selected drugs.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

The citation lists are queued

They are fetched later and appear here then.

References 65 · Cited by 57

Research works printed but not linked: 9

  1. DeLano WL (2002) PyMOL: an open-source molecular graphics tool. CCP4 Newsl Protein Crystallogr 40(1):82–92
  2. Devlin J, Chang M-W, Lee K, Toutanova K (2019) BERT: pre-training of deep bidirectional transformers for language understanding. NAACL-HLT 1:4171–4186 · DOI · Paper page
  3. Dinh L, Sohl-Dickstein J, Bengio S (2017) Density estimation using Real NVP. International Conference on Learning Representations. https://openreview.net/forum?id=HkpbnH9lx
  4. Hetzel L, Boehm S, Kilbertus N, Günnemann S, Theis F (2022) Predicting cellular responses to novel drug perturbations at a single-cell resolution. Adv Neural Inf Process Syst 35:26711–26722 · DOI
  5. Johansson F, Shalit U, Sontag D (2016) Learning representations for counterfactual inference. In: Proceedings of the International Conference on Machine Learning (ICML). PMLR, pp 3020–3029
  6. Kingma DP, Dhariwal P (2018) Glow: generative flow with invertible 1 × 1 convolutions. Advances in Neural Information Processing Systems (NeurIPS), 31, 10215–10224
  7. Kusner MJ, Paige B, Hernández-Lobato JM (2017) Grammar variational autoencoder. In: Proceedings of the International Conference on Machine Learning (ICML). PMLR, pp 1945–1954
  8. Shrikumar A, Greenside P, Kundaje A (2017) Learning important features through propagating activation differences. In: Proceedings of the International Conference on Machine Learning (ICML). PMLR, pp 3145–3153
  9. Sundararajan M, Taly A, Yan Q (2017) Axiomatic attribution for deep networks. In: Proceedings of the International Conference on Machine Learning (ICML). PMLR, pp 3319–3328

Metadata

Authors
Yu, H., Qian, W., Song, Y., & Welch, J. D.
Affiliations
University of Michigan (US)
Corresponding author
  • Yu, Hengshi
  • Qian, Weizhou
Venue
Molecular Systems Biology
Published
2025-07-10
Volume
21
Issue
8
Pages
960-982
ISSN
1744-4292 (ISSN-L)
OpenAlex
W4412160990
Cited by
57
References
65
Citations by year
2026: 27 · 2025: 16
FWCI
12.84
Type
journal article
Language
en
Open access
Licence
Locations
  • Molecular Systems Biology (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
HHS | NIH | National Human Genome Research Institute (R01HG010883, U01HG011952)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › CRISPR and Genetic Engineering
Keywords
deep generative models, CRISPR activation, CRISPR interference, gene expression prediction, missense mutations, GATA-1, hematopoietic stem cells, amino acid embeddings, point mutations
MeSH terms
Gene Expression Regulation, Humans, Computational Biology, Mutation, Missense, GATA1 Transcription Factor, Single-Cell Analysis, CRISPR-Cas Systems

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: PerturbNet predicts single-cell responses to unseen chemical and genetic perturbations. https://app.cochat.ai/papers/10.1038/s44320-025-00131-3
The paper
Yu, H., Qian, W., Song, Y., & Welch, J. D. (2025). PerturbNet predicts single-cell responses to unseen chemical and genetic perturbations. Molecular Systems Biology. https://doi.org/10.1038/s44320-025-00131-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.