Palantir analysis notebook

Table of contents

Introduction

Palantir is an algorithm to align cells along differentiation trajectories. Palantir models differentiation as a stochastic process where stem cells differentiate to terminally differentiated cells by a series of steps through a low dimensional phenotypic manifold. Palantir effectively captures the continuity in cell states and the stochasticity in cell fate determination.

Loading data

Palantir can read single cell RNA-seq data from a csv file, mtx file, 10x counts and HDF files. The csv file should be a cells X genes matrix of counts. This tutorial shows the usage with a csv file. The following code snippet imports the palantir package along with other plotting related imports

A sample RNA-seq csv data is available at <palantir directory>/data/marrow_sample_scseq_counts.csv.gz. This sample data will be used to demonstrate the utilization and capabilities of the Palantir package. This dataset contains ~4k cells and ~16k genes.

This is a subset of the data from the CD34+ cells from the human marrow, the dataset used in the manuscript. The data comprises of progenitor Hematopoeitic stem cells and cells of the erythroid and myeloid lineages.

Assuming that Palantir was downloaded at ~/repositories/palantir, the following snipped can be used to load the counts matrix.

For alternative file formats, please use palantir.io.from_mtx, palantir.io.from_10x, palantir.io.from_10x_HDF5 functions.

Data clean up

Palantir provides an interface for removing low molecule count cells and genes with low detection rates. The following snippet allows visualzation of molecule counts and gene detection rates

The dataset provided is already prefiltered. For other datasets, cells and genes can be filtered using

filtered_counts = palantir.preprocess.filter_counts_data(counts, cell_min_molecules=1000, genes_min_cells=10)

Normalization

The next step is to correct for molecule count distribution. Palantir divides the counts of each cell by the total molecules detected as a normalization technique

Data can be optionally log transformed. Note that, some datasets show better signal in the linear scale while others show stronger signal in the log scale. We use log transformation for this dataset

Data preprocessing

Principal component analysis

The first step in data processing for Palantir is to determine metagenes using principal component analysis. This representation is necessary to overcome the extensive dropouts that are pervasive in single cell RNA-seq data

Diffusion maps

Palantir next determines the diffusion maps of the data as an estimate of the low dimensional phenotypic manifold of the data.

The low dimensional embeddeing of the data is estimated based on the eigen gap using the following function

tSNE visualization

Palantir constructs the tSNE map in the embedded space since these maps better represent the differentiation trajectories

For consistency, the following snippet will load the precomputed tSNE projections.

tSNE results can be visualized by the plot_tsne and plot_tsne_by_cell_sizes functions. The plot_tsne_by_cell_sizes function colors the cells by their molecule counts before normalization

MAGIC imputation

MAGIC is an imputation technique developed in the Pe'er lab for single cell data imputation. Palantir uses MAGIC to impute the data for visualization and determining gene expression trends.

Gene expression can be visualized on tSNE maps using the plot_gene_expression function. The genes parameter is an string iterable of genes, which are a subset of the expression of column names. The below function plots the expression of HSC gene CD34, myeloid gene MPO and erythroid precursor gene GATA1 and dendritic cell gene IRF8.

Diffusion maps

The computed diffusion components can be visualized with the following snippet.

Running Palantir

Palantir can be run by specifying an approxiate early cell. While Palantir automatically determines the terminal states, they can also specified using the termine_states parameter.

The start cell for this dataset was chosen based on high expression of CD34.

Palantir generates the following results

1. Pseudotime: Pseudo time ordering of each cell
2. Terminal state probabilities: Matrix of cells X terminal states. Each entry represents the probability of the corresponding cell reaching the respective terminal state
3. Entropy: A quantiative measure of the differentiation potential of each cell computed as the entropy of the multinomial terminal state probabilities

The terminal states in this dataset are renamed to reflect the known biology below

Visualizing Palantir results

Palantir results can be visualized on the tSNE map using the plot_palantir_results function

Terminal state probability distributions of individual cells can be visualized using the plot_terminal_state_probs function

The cells can be highlighted on the tSNE map using the highlight_cells_on_tsne function

Gene expression trends

Palantir uses Generalized Additive Models (GAMs) to determine the gene expression trends along different lineages. The marker trends can be determined using the following snippet. This computes the trends for all lineages. A subset of lineages can be used using the lineages parameter.

The determined trends can be visualized with the plot_gene_trends function. A separate panel is generated for each gene

Alternatively, the trends can be visualized on a heatmap using

Clustering

Cells can be clustered and visualized using Phenograph with the following snippet.

Similary, gene expression trends can be clustered and visualized using the following snippet. As an example, the first 1000 genes along the erythroid genes are clustered