Welcome to Escort!

Escort is a framework that evaluates various data processing decisions in terms of their effect on trajectory inference with single-cell RNA-seq data. Escort guides users through a trajectory analysis by providing evaluations of embeddings, which represent combinations of analysis choices including feature selection, dimension reduction, normalization, and/or trajectory inference-specific hyperparameters.

In Step 1, Escort will assess the evidence of a trajectory signal in the dataset. Sometimes data are not suitable for trajectory analysis, for example, when cells come from biologically distinct clusters or have insufficient heterogeneity. In these cases, Escort will alert the user and offers guidance to further investigate the appropriateness of trajectory analysis.

In Step 2, Escort will compare various embeddings, specifically in terms of how well they preserve cellular relationships and the distribution of cells in the embedding.

In Step 3, Escort evaluates how well a specific trajectory inference method interacts with a given embedding. This allows for evaluaton of additional graph structures used by specific methods and consideration of method-specific parameters. Finally, Escort provides an overall score for each option, as well as, a classification to help researchers select more optinal analysis choices for inferring a trajectory from their data.

For any questions or issues, please submit a comment to our GitHub issues page .

Additional explanations are provided in our vignettes .

Step 1: Detecting existence of a trajectory signal

Before fitting a trajectory, Escort determines whether a given dataset is appropriate for trajectory analysis. There are two common scenarios that may render trajectory analysis inappropriate or require more careful consideration:
- Datasets that contain distinct/disjoint cell types
- Datasets in which the cells are too homogeneous/similar

Disjoint cell types evaluation: If this module fails, Escort will perform cluster-specific differential expression testing and display the top cluster-specific genes. Users should further examine this list to determine whether fitting a trajectory that connects these particular cell types is biologically reasonable. If so, then users should re-examine whether sufficient intermediate cells exist between these two groups, or whether batch effects have been exist and have been addressed.

Homogenous cells evaluation: If this module fails, Escort will report the top highly variable genes in the dataset along with their enrichments. Users should further examine this list to re-examine whether an underlying trajectory is appropriate. If so, then this output should be used to identify what other biological processes may be overriding or diluting the biological signal of interest (e.g. cell cycle) or excessive signal from ribosomal or mitochondrial genes.
Disjoint cell types detected?
Homogeneous cells detected?
Please select a database that contains information about the genome and gene symbol annotations of a particular organism.

Upload scRNA-seq datasets:

Please upload either .csv files or .rds files for both raw data and normalized data. More information on how to prepare these files from a Seurat object or SingleCellExperiment is provided here.







Step 1 only needs to be run once for any particular dataset, however its results are used in Step 2. Download and save the results to skip this step in the future.


If you have already generated your own embeddings, please proceed directory to Step 2.

Embeddings are data representations that are ultimately passed onto trajectory inference packages to generate trajectories and pseudotime estimates. Escort quantitatively compares these embeddings based on trajectory properties. Typically, embeddings are generated by varying data processing choices such as feature selection and dimension reduction, however, the user may consider as many data processing choices as desired. Simple embeddings can be generated individually here (in combination with the preferred trajectory method that will be used preliminarily for Step 3), or users can generate their own embeddings following the workflow in our vignette describing how to generate additional or custom embeddings for Escort R/Shiny.

If you have already generated your own embeddings, please proceed directory to Step 2.


Embeddings are required for step 2
Download embeddings (required for Step 2!)




Step 2: Evaluating the trajectory characteristics of embeddings

Next, Escort identifies preferred embeddings for performing trajectory inference. We will evaluate three characteristics of each embedding:
- The retention of inter-cellular relationships.
- The preservation of similarity relationships.
- Distribution of cells in the embedding space.

Inter-cellular relationships

Preservation of similarity relationships

Cell spread

Load all embeddings: (multiple allowed)

Step 3: Quantifying trajectory fitting performance

Embeddings are also evaluated in the context of a trajectory inference method. A preliminary trajectory in inferred and Escort assesses the proportion of cells having an ambiguous projection to the trajectory. For example, trajectories in a U-shape tend to be inaccurate because some cells will have map to opposing pseudotimes with equal probability.

Percentage of ambiguous cells

Escort suggestions for embedding selection

Below is a table with each embedding's rating according to their overall performance. Embeddings with a score larger than zero are Recommended for trajectoy inference.
Download the table