cell fate probabilities in single-cell data with Palantir

Page | 001

Articles
https://doi.org/10.1038/s41587-019-0068-4
Corrected: Author Correction
Characterization of cell fate probabilities in single-cell data with Palantir
Manu Setty, Vaidotas Kiseliovas, Jacob Levine, Adam Gayoso, Linas Mazutis and Dana Pe’er*
Single-cell RNA sequencing studies of differentiating systems have raised fundamental questions regarding the discrete ver- sus continuous nature of both differentiation and cell fate. Here we present Palantir, an algorithm that models trajectories of differentiating cells by treating cell fate as a probabilistic process and leverages entropy to measure cell plasticity along the trajectory. Palantir generates a high-resolution pseudo-time ordering of cells and, for each cell state, assigns a probability of differentiating into each terminal state. We apply our algorithm to human bone marrow single-cell RNA sequencing data and detect important landmarks of hematopoietic differentiation. Palantir’s resolution enables the identification of key transcrip- tion factors that drive lineage fate choice and closely track when cells lose plasticity. We show that Palantir outperforms exist- ing algorithms in identifying cell lineages and recapitulating gene expression trends during differentiation, is generalizable to diverse tissue types, and is well-suited to resolving less-studied differentiating systems.
Differentiation is among the most fundamental processes in biology. In the traditional view, cells transition from a less- to a more-differentiated state via a series of discrete, well-
defined stages. Single-cell studies1–6 have, however, demonstrated that during differentiation cell states reside along largely continuous spaces. Despite this evolution in thinking, cell fate decisions con- tinue to be largely conceptualized as a series of discrete bifurcations along development, leading to terminal cell states7,8.
Epigenetic studies, however, support a probabilistic view of cell fate choice. Epigenomic measurements such as DNase I hypersensi- tivity site sequencing (DNase-seq) and assay for transposase-acces- sible chromatin using sequencing (ATAC-seq) suggest potential mechanisms for a continuous process by indicating that progressive enhancer restriction, coupled with pre-establishment of lineage- specifying enhancers in precursor cells, can serve as a vehicle for driving differentiation5,9,10. Indeed, in human bone marrow, we observe a lack of well-defined bifurcation points when single-cell RNA sequencing (scRNA-seq) profiles are projected along the strongest axes of variation (Fig. 1a). Even at the level of individual genes, we find a broad representation of gene ratios rather than bimodal expression states (Fig. 1a). These observations raise funda- mental questions about whether cell fates, similar to cell state transi- tions, are continuous and when and how cell fate choices are made.
To investigate these questions, we developed Palantir, an algo- rithm that leverages scRNA-seq data to model the landscape of dif- ferentiation and characterize continuity in both cell state and fate choice. As differentiation is asynchronous, sequencing a population of differentiating cells yields a snapshot representing a range of cell states. Based on scRNA-seq data from a single sample and the selec- tion of a representative early cell, Palantir generates a pseudo-time ordering of cells and, for each cell state, assigns a probability for differentiating into each terminal state. We applied Palantir to char- acterize human hematopoietic differentiation using scRNA-seq pro- files of ~25,000 cells enriched for CD34, a marker for hematopoietic stem and progenitor cells11. Palantir identified established termi- nal states and ordered cells along a pseudo-time that recapitulated known marker trends in development. Notably, Palantir identified
points along the trajectory where the differentiation potential (DP) drastically shifts. These shifts mark key events in hematopoiesis. Palantir thus provides a quantitative approach to characterizing a continuous model of cell fate choice.
Results
Development as a Markov process. Differentiation proceeds through cell divisions, where daughter cells are generally very similar to their mother cells. Thus, the population is established by incremental divergences, driven by regulatory mechanisms that create paths through the space of possible cell states (phenotypes). Regulation constrains cell states to a low-dimensional manifold of possible phenotypes12. Nearest-neighbor graphs, where each node represents a particular cell state and edges connect most similar cells, have been widely used to model this manifold1–3,13.
A single bone marrow sample contains the full spectrum of cell states in hematopoiesis and importantly the frequencies of each cell state. We leverage cell state frequencies to inform our model of possible differentiation paths in the neighbor graph and their likeli- hoods. Critically, paths along the graph represent probable trajec- tories of cells in the population rather than the path of a particular cell, and each cell state (graph node) is associated with a probabil- ity distribution for reaching the terminal states. We assert that cells traverse the manifold in small steps which can be modeled using a Markov chain to represent cell fate choices in a probabilistic manner, based on two key assumptions. Firstly, as in all pseudo-time infer- ence algorithms1,3,7,8, we assume unidirectional progression from a less- to a more-differentiated state. We posit that it is a reasonable first order approximation for healthy differentiation, but note that it fails in aberrant systems such as cancer, which require additional information (for example, mutations) to determine directionality. Second, we assume that for any node, the probability of traversing to any neighbor is independent of its history, that is, the path taken to reach that state. Note that for a particular cell, the cell’s develop- mental history is likely to be encoded in its epigenetic profile and will probably impact cell fate choices. However, nodes are abstract cell states representing multiple histories and potential trajectories
Program for Computational and Systems Biology, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA. *e-mail: peerd@mskcc.org
NAtuRe BioteChNoLoGy | VOL 37 | APRIL 2019 | 451–460 | www.nature.com/naturebiotechnology
451