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

The DOI matches its Crossref record; 55 of the 55 references it cites were found in a registry or an index.

Generative modeling of single-cell time series with PRESCIENT enables prediction of cell trajectories with interventions

Open access
Yeo, Grace Hui Ting, Saksena, Sachit D., Gifford, David K. · 2021 · Nature Communications

Abstract

Existing computational methods that use single-cell RNA-sequencing (scRNA-seq) for cell fate prediction do not model how cells evolve stochastically and in physical time, nor can they predict how differentiation trajectories are altered by proposed interventions. We introduce PRESCIENT (Potential eneRgy undErlying Single Cell gradIENTs), a generative modeling framework that learns an underlying differentiation landscape from time-series scRNA-seq data. We validate PRESCIENT on an experimental lineage tracing dataset, where we show that PRESCIENT is able to predict the fate biases of progenitor cells in hematopoiesis when accounting for cell proliferation, improving upon the best-performing existing method. We demonstrate how PRESCIENT can simulate trajectories for perturbed cells, recovering the expected effects of known modulators of cell fate in hematopoiesis and pancreatic β cell differentiation. PRESCIENT is able to accommodate complex perturbations of multiple genes, at different time points and from different starting cell populations, and is available at https://github.com/gifford-lab/prescient .

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/s41467-021-23518-w

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 A generative model of cellular differentiation. a Existing single-cell models of development can be described as operating in pseudo-time or real time (x-axis), and by the extent to which they model the underlying differentiation process (y-axis). PRESCIENT is highlighted in red. b Observations of population-level time-series data are used in a generative framework that models the underlying dynamic process in physical time. Evolution of a cell’s state is governed by a drift term and a noise term. The drift, depicted by solid arrows, is defined as the negative gradient of the potential function, depicted by the color gradient in the background. Dashed lines correspond to noise. The model is fit using observations of population-level time-series data, depicted as solid circles. Simulations of cell states are depicted as dashed circles. c Cartoon depicting model fitting process. The neural network parameterizing the underlying drift function μ takes as input the PCA projections of gene expression data at observed time points (again depicted as solid lines). The stochastic process is then simulated via first-order time discretization to produce a population at the next time step, and so on. This proceeds until the next observed time point, at which the loss between the simulated and predicted population is minimized. The model was validated using two tasks. d Held-out recovery, where the model was asked to predict the marginal distribution of a held-out time point, and e–f, fate prediction, where the model was asked to predict the fate distribution outcome of a given progenitor cell. Fate prediction can be applied to cells observed in the dataset (e) or cell states in which some perturbation has been applied in silico (f). As shown, the perturbation results in a significant shift of fate distribution outcomes.
Fig. 1 A generative model of cellular differentiation. a Existing single-cell models of development can be described as operating in pseudo-time or real time (x-axis), and by the extent to which they model the underlying differentiation process (y-axis). PRESCIENT is highlighted in red. b Observations of population-level time-series data are used in a generative framework that models the underlying dynamic process in physical time. Evolution of a cell’s state is governed by a drift term and a noise term. The drift, depicted by solid arrows, is defined as the negative gradient of the potential function, depicted by the color gradient in the background. Dashed lines correspond to noise. The model is fit using observations of population-level time-series data, depicted as solid circles. Simulations of cell states are depicted as dashed circles. c Cartoon depicting model fitting process. The neural network parameterizing the underlying drift function μ takes as input the PCA projections of gene expression data at observed time points (again depicted as solid lines). The stochastic process is then simulated via first-order time discretization to produce a population at the next time step, and so on. This proceeds until the next observed time point, at which the loss between the simulated and predicted population is minimized. The model was validated using two tasks. d Held-out recovery, where the model was asked to predict the marginal distribution of a held-out time point, and e–f, fate prediction, where the model was asked to predict the fate distribution outcome of a given progenitor cell. Fate prediction can be applied to cells observed in the dataset (e) or cell states in which some perturbation has been applied in silico (f). As shown, the perturbation results in a significant shift of fate distribution outcomes.CC BY · PubMed Central
Fig. 2 Fate outcomes generated by PRESCIENT align with experimental lineage tracing when taking into account cell proliferation. a Testing performance of models trained with and without proliferation on the task of recovering the held-out time point, day 4. Performance is reported as Wasserstein distance of simulated population with respect to actual populations at the given time points for models trained without and with proliferation over n = 5 seeds. Dashed line indicates distance of linearly interpolated population with respect to the actual population on day 4. b Correlation of actual and estimated proliferation rates on day 2 cells with lineage tracing data (n = 4638). c Examples of n = 50 trajectories with starting cells assigned different clonal fate biases. The color-scale indicates time corresponding to 100 time steps of dt = 0.1 starting from t = 2. Clonal fate bias is computed with respect to monocyte/neutrophil populations at the final time point. d Visualizations of underlying potential and drift functions learned by models with and without cell proliferation. Drift is visualized for a random sample of cells. Dotted circle indicates qualitative differences in potential landscape. e Summary of training/test splits of lineage tracing dataset. Training sets either include only cells with lineage tracing data, all cells, or cells without lineage tracing data. f–h, Performance of other methods (far left, Smoothed fate probabilities given by held-out clonal data, PBA: population balance analysis10, WOT: Waddington-OT8, FateID9 evaluated using predictions provided by Weinreb et al.) in comparison to PRESCIENT (left-right) on predicting clonal fate bias for the same test set (n = 335) given different training sets. Training sets either consisted of only cells for which lineage tracing data was available (left), all cells (middle), or cells without lineage tracing data (right). Performance metrics evaluated include f Pearson r, g AUROC, and h fraction of test set in which at least 1 simulated cell at the final time point is either classified as a neutrophil or a monocyte over n = 5 seeds. In b, f–g, boxplots indicate median (middle line), first and third quartiles (box), and the upper whisker extends from the edges to the largest value no further than 1.5 × IQR (interquartile range) from the quartiles and the lower whisker extends from the edge to the smallest value at most 1.5 × IQR of the edge, while data beyond the end of the whiskers are outlying points that are plotted individually as diamonds. In h, bar plots show the average fraction of cells with error bars representing the 95% CI.
Fig. 2 Fate outcomes generated by PRESCIENT align with experimental lineage tracing when taking into account cell proliferation. a Testing performance of models trained with and without proliferation on the task of recovering the held-out time point, day 4. Performance is reported as Wasserstein distance of simulated population with respect to actual populations at the given time points for models trained without and with proliferation over n = 5 seeds. Dashed line indicates distance of linearly interpolated population with respect to the actual population on day 4. b Correlation of actual and estimated proliferation rates on day 2 cells with lineage tracing data (n = 4638). c Examples of n = 50 trajectories with starting cells assigned different clonal fate biases. The color-scale indicates time corresponding to 100 time steps of dt = 0.1 starting from t = 2. Clonal fate bias is computed with respect to monocyte/neutrophil populations at the final time point. d Visualizations of underlying potential and drift functions learned by models with and without cell proliferation. Drift is visualized for a random sample of cells. Dotted circle indicates qualitative differences in potential landscape. e Summary of training/test splits of lineage tracing dataset. Training sets either include only cells with lineage tracing data, all cells, or cells without lineage tracing data. f–h, Performance of other methods (far left, Smoothed fate probabilities given by held-out clonal data, PBA: population balance analysis10, WOT: Waddington-OT8, FateID9 evaluated using predictions provided by Weinreb et al.) in comparison to PRESCIENT (left-right) on predicting clonal fate bias for the same test set (n = 335) given different training sets. Training sets either consisted of only cells for which lineage tracing data was available (left), all cells (middle), or cells without lineage tracing data (right). Performance metrics evaluated include f Pearson r, g AUROC, and h fraction of test set in which at least 1 simulated cell at the final time point is either classified as a neutrophil or a monocyte over n = 5 seeds. In b, f–g, boxplots indicate median (middle line), first and third quartiles (box), and the upper whisker extends from the edges to the largest value no further than 1.5 × IQR (interquartile range) from the quartiles and the lower whisker extends from the edge to the smallest value at most 1.5 × IQR of the edge, while data beyond the end of the whiskers are outlying points that are plotted individually as diamonds. In h, bar plots show the average fraction of cells with error bars representing the 95% CI.CC BY · PubMed Central
Fig. 3 In silico perturbations of hematopoiesis results in expected shifts in fate distribution. a The distribution of cells at the final time-point generated by the model initialized with unperturbed cells (left) and cells with perturbations of Lmo4, Cebpe, Mxd1, and Dach1 upregulated (z = 10) during neutrophil differentiation (right). b Proportions of generated cell types from day 2 to 6 initialized with unperturbed cells (left) and cells with perturbations of transcription factors upregulated during neutrophil differentiation (right). c Fraction of neutrophil and monocyte cells at final time point with single-gene perturbations. Transcription factors involved in monocyte development are indicated in green, while transcription factors involved in neutrophil development are indicated in orange. Control genes (in gray) indicate experiments when perturbing genes from a random set of non-TFs as in (h) d, e, Individual genetic perturbations made to transcription factors involved in neutrophil and monocyte differentiation have an increased effect at higher dosages. f–g Ensemble perturbations of transcription factors involved in neutrophil (orange) and monocyte (green) differentiation have a stronger effect. h Ensemble random perturbations of non-transcription factors without proliferative signatures (gray). In c–h, boxplots are of randomly initialized unperturbed vs. perturbed simulations (n = 10) with 200 cells for each initialization, and red asterisks indicate Welch’s independent two-sided t-test at p < 0.05. Boxplots indicate median (middle line), first and third quartiles (box), and the upper whisker extends from the edges to the largest value no further than 1.5 × IQR (interquartile range) from the quartiles and the lower whisker extends from the edge to the smallest value at most 1.5 × IQR of the edge, while data beyond the end of the whiskers are outlying points that are plotted individually as diamonds.
Fig. 3 In silico perturbations of hematopoiesis results in expected shifts in fate distribution. a The distribution of cells at the final time-point generated by the model initialized with unperturbed cells (left) and cells with perturbations of Lmo4, Cebpe, Mxd1, and Dach1 upregulated (z = 10) during neutrophil differentiation (right). b Proportions of generated cell types from day 2 to 6 initialized with unperturbed cells (left) and cells with perturbations of transcription factors upregulated during neutrophil differentiation (right). c Fraction of neutrophil and monocyte cells at final time point with single-gene perturbations. Transcription factors involved in monocyte development are indicated in green, while transcription factors involved in neutrophil development are indicated in orange. Control genes (in gray) indicate experiments when perturbing genes from a random set of non-TFs as in (h) d, e, Individual genetic perturbations made to transcription factors involved in neutrophil and monocyte differentiation have an increased effect at higher dosages. f–g Ensemble perturbations of transcription factors involved in neutrophil (orange) and monocyte (green) differentiation have a stronger effect. h Ensemble random perturbations of non-transcription factors without proliferative signatures (gray). In c–h, boxplots are of randomly initialized unperturbed vs. perturbed simulations (n = 10) with 200 cells for each initialization, and red asterisks indicate Welch’s independent two-sided t-test at p < 0.05. Boxplots indicate median (middle line), first and third quartiles (box), and the upper whisker extends from the edges to the largest value no further than 1.5 × IQR (interquartile range) from the quartiles and the lower whisker extends from the edge to the smallest value at most 1.5 × IQR of the edge, while data beyond the end of the whiskers are outlying points that are plotted individually as diamonds.CC BY · PubMed Central
Fig. 4 PRESCIENT predicts expected temporal dynamics of in silico perturbations of the endocrine/exocrine axis. a Simplified model of in vitro endocrine induction. Progenitor cells first bifurcate along the endocrine-exocrine axis, and then are specified into more specific pancreatic cell fates. b Schematic of steps for different perturbational experiments. First, cells are sampled from either a specific timepoint in the scRNA-seq timecourse or by a specific subtype label. Next, either individual genes or ensembles of genes are perturbed by setting target genes to a higher or lower z-score in scaled gene expression space. This profile is then reduced to principal components. These perturbed cells are then used to initialize a simulation of an already trained PRESCIENT stochastic process, which is simulated forward to the final timepoint. Finally, a pre-trained ANN classifier is used to evaluate the cell-type distribution at the final time-step of the simulation. This distribution is then compared to an unperturbed simulation via a two-sided paired t-test. c Final fractions of endocrine and exocrine cells as a result of ensembled perturbations of endocrine- and exocrine- associated TFs (top, bottom) on day 0 and day 4 (left, right) with increasing magnitude. Boxplots (as described previously) are of randomly initialized unperturbed vs. perturbed simulations (n = 10) with 200 cells for each initialization, and red asterisks indicate Welch’s independent two-sided t-test at p < 0.05. d In silico perturbations are introduced at different time points to the corresponding unperturbed population. The different outcomes of the cell type of interest are then calculated as the difference in fraction at the final time point starting from the perturbed and unperturbed populations. Bar plots show average differences over n = 10 randomly sampled cell populations of 200 cells each from each timepoint with error bars representing the 95% CI.
Fig. 4 PRESCIENT predicts expected temporal dynamics of in silico perturbations of the endocrine/exocrine axis. a Simplified model of in vitro endocrine induction. Progenitor cells first bifurcate along the endocrine-exocrine axis, and then are specified into more specific pancreatic cell fates. b Schematic of steps for different perturbational experiments. First, cells are sampled from either a specific timepoint in the scRNA-seq timecourse or by a specific subtype label. Next, either individual genes or ensembles of genes are perturbed by setting target genes to a higher or lower z-score in scaled gene expression space. This profile is then reduced to principal components. These perturbed cells are then used to initialize a simulation of an already trained PRESCIENT stochastic process, which is simulated forward to the final timepoint. Finally, a pre-trained ANN classifier is used to evaluate the cell-type distribution at the final time-step of the simulation. This distribution is then compared to an unperturbed simulation via a two-sided paired t-test. c Final fractions of endocrine and exocrine cells as a result of ensembled perturbations of endocrine- and exocrine- associated TFs (top, bottom) on day 0 and day 4 (left, right) with increasing magnitude. Boxplots (as described previously) are of randomly initialized unperturbed vs. perturbed simulations (n = 10) with 200 cells for each initialization, and red asterisks indicate Welch’s independent two-sided t-test at p < 0.05. d In silico perturbations are introduced at different time points to the corresponding unperturbed population. The different outcomes of the cell type of interest are then calculated as the difference in fraction at the final time point starting from the perturbed and unperturbed populations. Bar plots show average differences over n = 10 randomly sampled cell populations of 200 cells each from each timepoint with error bars representing the 95% CI.CC BY · PubMed Central
Fig. 5 PRESCIENT predicts the outcome of transcription factor perturbations in a large perturbational screen as well as the different effects of perturbations in early progenitors vs. cells further along endocrine induction. a Perturbational screen (z = 5) of all the TFs in the highly variable gene set of in vitro β-cell differentiation. The x-axis is the log2 fold-change (FC) of the final cell-type fraction of the target cell fraction between perturbed and unperturbed simulations. The y-axis is the -log10 p values of two-sided paired t-tests between target cell fate outcomes between unperturbed and perturbed simulations over n = 10 randomly sampled starting populations at the final time step consisting of 200 cells each. Points are colored if they are a hit (FDR < 0.01 and log2(FC) > 0.5) and β-cell fractions are shown in purple, α-cells in red, and EC-cells in blue. b Difference in final β-cell populations when introducing perturbations (z = 5) in different cell populations for different TFs. c Difference in final α-cell perturbations when introducing perturbations (z = 5) in different cell populations for different TFs. In b, c, different starting populations correspond to cell stages as labeled by Veres et al.: SOX2+ progenitors (progsox2), NKX61+ progenitors (prognkx61), and cells with early/middle/late NEUROG3 signatures (neurog3early, neurog3mid, neurog3late, respectively). Bar plots show average paired differences over n = 10 randomly sampled cell populations of 200 cells from each early subpopulation with error bars representing the 95% CI. Figure S4e, f show boxplots of this data.
Fig. 5 PRESCIENT predicts the outcome of transcription factor perturbations in a large perturbational screen as well as the different effects of perturbations in early progenitors vs. cells further along endocrine induction. a Perturbational screen (z = 5) of all the TFs in the highly variable gene set of in vitro β-cell differentiation. The x-axis is the log2 fold-change (FC) of the final cell-type fraction of the target cell fraction between perturbed and unperturbed simulations. The y-axis is the -log10 p values of two-sided paired t-tests between target cell fate outcomes between unperturbed and perturbed simulations over n = 10 randomly sampled starting populations at the final time step consisting of 200 cells each. Points are colored if they are a hit (FDR < 0.01 and log2(FC) > 0.5) and β-cell fractions are shown in purple, α-cells in red, and EC-cells in blue. b Difference in final β-cell populations when introducing perturbations (z = 5) in different cell populations for different TFs. c Difference in final α-cell perturbations when introducing perturbations (z = 5) in different cell populations for different TFs. In b, c, different starting populations correspond to cell stages as labeled by Veres et al.: SOX2+ progenitors (progsox2), NKX61+ progenitors (prognkx61), and cells with early/middle/late NEUROG3 signatures (neurog3early, neurog3mid, neurog3late, respectively). Bar plots show average paired differences over n = 10 randomly sampled cell populations of 200 cells from each early subpopulation with error bars representing the 95% CI. Figure S4e, f show boxplots of this data.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

Cites

Show all 55 references

Cited by

See all 125

Metadata

Authors
Yeo, G. H. T., Saksena, S. D., & Gifford, D. K.
Affiliations
MIT Computer Science and Artificial Intelligence Laboratory (US)
Corresponding author
Gifford, David K.
Venue
Nature Communications
Published
2021-05-28
Volume
12
Issue
1
Pages
3222-3222
ISSN
2041-1723 (ISSN-L)
OpenAlex
W3164448939
Cited by
103
References
55
Citations by year
2026: 31 · 2025: 25 · 2024: 26 · 2023: 14 · 2022: 6 · 2021: 1
FWCI
3.69
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc)
Funders
U.S. Department of Health & Human Services | National Institutes of Health (5 R01 NS109217, 5 R01 HG008363)
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 › Developmental Biology and Gene Regulation
Keywords
generative models, cell fate prediction, cell differentiation trajectories, lineage tracing, hematopoiesis, pancreatic beta cell differentiation, cell proliferation, perturbation prediction, single-cell RNA sequencing
MeSH terms
Deep Learning, RNA-Seq, Animals, Cell Differentiation, Cells, Cultured, Computer Simulation, Hematopoiesis, Humans, Models, Genetic, Software, Stem Cells, Stochastic Processes, Cell Proliferation, Insulin-Secreting Cells, Mice, Single-Cell Analysis, Datasets as Topic

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: Generative modeling of single-cell time series with PRESCIENT enables prediction of cell trajectories with interventions. https://app.cochat.ai/papers/10.1038/s41467-021-23518-w
The paper
Yeo, G. H. T., Saksena, S. D., & Gifford, D. K. (2021). Generative modeling of single-cell time series with PRESCIENT enables prediction of cell trajectories with interventions. Nature Communications. https://doi.org/10.1038/s41467-021-23518-w

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.