Generative modeling of single-cell population time series for inferring cell differentiation landscapes

Abstract

SummaryExisting computational methods that use single-cell RNA-sequencing for cell fate prediction either summarize observations of cell states and their couplings without modeling the underlying differentiation process, or are limited in their capacity to model complex differentiation landscapes. Thus, contemporary methods cannot predict how cells evolve stochastically and in physical time from an arbitrary starting expression state, nor can they model the cell fate consequences of gene expression perturbations. We introduce PRESCIENT (Potential eneRgy undErlying Single Cell gradIENTs), a generative modeling framework that learns an underlying differentiation landscape from single-cell time-series gene expression data. Our generative model framework provides insight into the process of differentiation and can simulate differentiation trajectories for arbitrary gene expression progenitor states. We validate our method on a recently published experimental lineage tracing dataset that provides observed trajectories. We show that this model is able to predict the fate biases of progenitor cells in neutrophil/macrophage lineages when accounting for cell proliferation, improving upon the best-performing existing method. We also show how a model can predict trajectories for cells not found in the model’s training set, including cells in which genes or sets of genes have been perturbed. PRESCIENT is able to accommodate complex perturbations of multiple genes, at different time points and from different starting cell populations. PRESCIENT models are able to recover the expected effects of known modulators of cell fate in hematopoiesis and pancreatic β cell differentiation.

A systematic comparison of computational methods for expression forecasting

“…Large-scale perturbation datasets display a high rate of success in targeting individual genes, modest correlation across replicates, and mostly small global effects Table 1 Figure S1 Figure 2 Regression-based analyses do not out-perform uninformed baselines Figure 3 Figure S2 Table 4 Diverse published methods do not out-perform simple baselines Figure 5 A grammar of gene regulatory network (GRN) inference synthesizes diverse methods withinIntroduction Systematically determining causal effects of transcription factor (TF) activity is a fundamental problem in systems biology, with early studies even prior to whole-genome expression profiling (Akutsu, Miyano, & Kuhara, 1999;Duggan, Bittner, Chen, Meltzer, & Trent, 1999;Liang, Fuhrman, & Somogyi, 1998). Driven by new datasets, a raft of recent computational methods now predict genetic perturbation results in silico, with applications in embryology, stem cell reprogramming, and drug discovery (Amrute et al, 2022;Burdziak et al, 2023;Cui, Wang, Maan, & Wang, 2023;Hyttinen, Eberhardt, & Hoyer, 2012;Jiang et al, 2023;Kamimoto et al, 2023;Osorio et al, 2022;Qiu et al, 2022;Roohani, Huang, & Leskovec, 2022;Theodoris et al, 2023;Tran, Yang, Yang, & Ormerod, 2022;Wang et al, 2022;Yeo, Saksena, & Gifford, 2021). These expression forecasting methods aim to serve as a new type of general-purpose screening tool.…”

“…Compared to Perturb-seq and similar assays, computer simulations are cheaper, less labor-intensive, and easier to apply to less accessible cell types. Example applications frequently include ranking or nomination of hundreds of genetic perturbations, sometimes in cell types where very few genetic perturbations have been conducted (Burdziak et al, 2023;Jiang et al, 2023;Kamimoto et al, 2023;Roohani et al, 2022;Theodoris et al, 2023;Wang et al, 2022;Yeo et al, 2021). This could enable more efficient research in a variety of areas; examples might include determining genetic dependencies of cell fate transitions in cancer (Baca et al, 2021;Reddy et al, 2021) or using reprogramming to reverse molecular indicators of aging (Lee et al, 2021).…”

“…The utility of these approaches hinges on their accuracy. A number of rationales are provided for the feasibility of expression forecasting under novel genetic perturbations, including mathematical identifiability guarantees (Yeo et al, 2021), use of prior functional annotations (Roohani et al, 2022), support for TF-to-target binding via ChIP or motif analysis (Burdziak et al, 2023;Kamimoto et al, 2023;Tran et al, 2022;Wang et al, 2022), and empirical tests against real perturbation outcomes (Cui et al, 2023;Jiang et al, 2023;Kamimoto, Hoffmann, & Morris, 2020;Qiu et al, 2022;Roohani et al, 2022;Theodoris et al, 2023;Wang et al, 2022;Yeo et al, 2021). However, most expression forecasting methods are very new, and most have been tested on only a small number of perturbations, with few direct comparisons across methods (Cui et al, 2023;Jiang et al, 2023;Lopez, Hütter, Pritchard, & Regev, 2022;Roohani et al, 2022;Theodoris et al, 2023;Tran et al, 2022;Wang et al, 2022).…”

“…Methods for expression forecasting differ in terms of expected output. Some methods provide only cell-type classifications or low-dimensional embeddings (Cui et al, 2023;Kamimoto et al, 2023;Theodoris et al, 2023;Yeo et al, 2021), while others provide gene-specific predictions of expression, velocity, or fold-change (Burdziak et al, 2023;Qiu et al, 2022;Roohani et al, 2022;Wang et al, 2022). To build flexible and extensible benchmarking infrastructure, we have collected datasets with transcriptome-wide profiles of many genetic perturbations.…”

Single-Cell RNA Sequencing Procedures and Data Analysis

“…This model generalizes RNA velocity computations with transient cell states, which are common in the development of and the response to perturbations. Another interesting approach to simulate time-series trajectories is proposed by Yeo et al (76), who are using a generative model framework that is able to predict trajectories for cells, which are not found in the model' s training set (including cells in which genes or sets of genes have been perturbed).…”