GEMLI is an R package to predict cell lineages (cells with a common ancestor) from single cell RNA datasets (or any other type of single-cell gene expression data) and to call genes with a high gene expression memory on the predicted cell lineages. It is described in A.S. Eisele*, M. Tarbier*, A.A. Dormann, V. Pelechano, D.M. Suter | "Gene-expression memory-based prediction of cell lineages from scRNA-seq datasets" | Nature Communications 15, 2744 (2024). https://doi.org/10.1038/s41467-024-47158-y
The approach is based on findings of Phillips et al. 2019 (doi.org/10.1038/s41467-019-09189-8) where it was shown that some genes show varying gene expression across cell lineages that is stable over multiple cell generations.
Simply run library(devtools) and then install_github("UPSUTER/GEMLI", subdir="GEMLI_package_v0"). If devtools is not installed yet, you can do so with install.packages("devtools"). GEMLI is now installed and can be used via library(GEMLI). Dependencies will be installed with the package.
We are still working to make GEMLI more intuitive, user-friendly, faster and versatile. Therefore existing functions will still be updated and new functionalities will be added. We'll publish a list of changes for each version for you to keep track. Your feedback is very welcome and will help us make GEMLI even better. What do you like about GEMLI? Something not working? What functions are you missing? Let us know! Contact: marcel.tarbier@scilifelab.se or almut.eisele@epfl.ch
For our first example we'll be looking at mouse embryonic stem cells that have been barcoded and cultured for 48h. We'll be working with a subset of this data for fast processing. In our subset we find 'family sizes' ranging from just two up to five related cells.
First we load the example data.
> load('GEMLI_example_data_matrix.RData')
> load('GEMLI_example_barcode_information.RData')
GEMLI's inputs and outputs are stored in a list of objects with predefined names. To run GEMLI you need at least a quality controlled and normalized gene expression matrix (rows = genes/features, colums = cells/samples). In this example we also provide a ground truth for lineages stemming from a barcoding experiment (values = barcode ID, names = cell IDs).
> GEMLI_items = list()
> GEMLI_items[['gene_expression']] = data_matrix
> GEMLI_items[['barcodes']] = lineage_dict_bc
>
> GEMLI_items[['gene_expression']][9:14,1:5]
AAACGAACAGGTGTGA-1 AAAGGTAGTTGCTTGA-1 AACCACAAGTTTGTCG-1 AAGCCATGTTCCACGG-1 AAGCGAGGTACGGCAA-1
ENSMUSG00000033845 14.761746 12.9570026 13.240645 8.8596794 12.791617
ENSMUSG00000025903 3.163231 1.2340002 4.878132 0.7383066 0.000000
ENSMUSG00000033813 6.326463 5.5530011 4.181256 8.8596794 5.482122
ENSMUSG00000002459 0.000000 0.0000000 0.000000 0.0000000 0.000000
ENSMUSG00000085623 0.000000 0.0000000 0.000000 0.0000000 0.000000
ENSMUSG00000033793 3.163231 0.6170001 1.393752 2.2149199 5.482122
>
> GEMLI_items[['barcodes']][1:5]
CACAGATAGTGATGGC-1 TATCTTGGTACGGGAT-1 AAACGAACAGGTGTGA-1 AGAGAATAGGTCATAA-1 GAGTGAGTCCAGTACA-1
2 2 2 7 7
We can then identify cell lineages through repeated iterative clustering (this may take 2-3min). The predict_lineages function takes our GEMLI_items as input. It outputs a matrix of all cells against all cells with values corresponding to a confidence score that they are part of the same lineage.
> GEMLI_items = predict_lineages(GEMLI_items)
>
> GEMLI_items[['prediction']][1:5,15:19]
AGAGAATAGGTCATAA-1 AGAGCAGCAAGTGATA-1 AGATGCTTCAAAGACA-1 AGGATCTGTATCGTTG-1 AGGGAGTAGACGATAT-1
AAACGAACAGGTGTGA-1 0 0 0 0 14
AAAGGTAGTTGCTTGA-1 87 0 0 0 0
AACCACAAGTTTGTCG-1 0 0 0 0 39
AAGCCATGTTCCACGG-1 0 0 0 0 0
AAGCGAGGTACGGCAA-1 0 0 0 0 0
Since we have barcoding data for this dataset we can test the predicted lineages against our ground truth.
The test_lineage_prediction function again takes our GEMLI_items as input. It's important that a predcition has been run first with predict_lineages. It outputs the number of true positive predictions (TP), false positive predictions (FP), as well as precision and sensitivity for various confidence intervals. The output can be visualized by setting plot_results to true/T.
> GEMLI_items = test_lineages(GEMLI_items)
>
> GEMLI_items$testing_results
TP FP precision sensitivity
0 274 24062 0.01125904 1.0000000
10 104 128 0.44827586 0.3795620
20 90 82 0.52325581 0.3284672
30 84 56 0.60000000 0.3065693
40 80 36 0.68965517 0.2919708
50 68 22 0.75555556 0.2481752
60 64 14 0.82051282 0.2335766
70 62 10 0.86111111 0.2262774
80 58 2 0.96666667 0.2116788
90 46 0 1.00000000 0.1678832
100 34 0 1.00000000 0.1240876
>
> GEMLI_items = test_lineages(GEMLI_items, plot_results=T)
We can also investigate our predictions by visualizing them as a network with the visualize_as_network function. Here we need to set a cutoff that defines which predictions we want to consider. It represents a confidence score and high values yield fewer predictions with high precision while low values yield more predictions with lower precision.
Network visualization requires the igraph library which can be loaded with library(igraph).
> visualize_as_network(GEMLI_items, cutoff=90) # top image
> visualize_as_network(GEMLI_items, cutoff=50) # lower image
If a ground truth e.g. from barcoding is available we can set ground_truth to true/T and highlight_FPs to true/T to highlight false predictions with red edges. Cells without barcode information will be displayed in white.
> visualize_as_network(GEMLI_items, cutoff=90, ground_truth=T, highlight_FPs=T) # top image
> visualize_as_network(GEMLI_items, cutoff=50, ground_truth=T, highlight_FPs=T) # lower image
Now we can extract the lineage information with the prediction_to_lineage_information function. Again we need to set a cutoff that defines which predictions we want to consider. The function outputs both a lineage table and a 'dictionary', a vector that has the lineage number as values and the cell IDs as names.
> GEMLI_items = prediction_to_lineage_information(GEMLI_items, cutoff=50)
>
> GEMLI_items$predicted_lineage_table[1:5,]
cell.ID clone.ID
[1,] "AAACGAACAGGTGTGA-1" "1"
[2,] "AAAGGTAGTTGCTTGA-1" "2"
[3,] "AACCACAAGTTTGTCG-1" "3"
[4,] "AAGCCATGTTCCACGG-1" "4"
[5,] "AAGCGAGGTACGGCAA-1" "5"
>
> GEMLI_items$predicted_lineages[1:5]
AAACGAACAGGTGTGA-1 AAAGGTAGTTGCTTGA-1 AACCACAAGTTTGTCG-1 AAGCCATGTTCCACGG-1 AAGCGAGGTACGGCAA-1
1 2 3 4 5
In some applications it may be useful to trim lineages that are too big. For instance if it is known that cells should have undergone only a certain number of divisions effectively limiting the lineage size or if you are only interested in sister cell pairs. Similarly, if you investigate large lineages you want to avoid lineages being merged due to few false predictions between otherwise well-interconnected lineages. The suggest_network_trimming_to_size function allows you to preview what a trimming to size would look like. It will again show the predicted lineages as networks but highlight all connections that would be trimmed given a certain size restriction (max_size). If you are happy with the suggested trimming you create new trimmed GEMLI_items list, in this example we called it GEMLI_items_post_processed. You can then again visualize the predictions to see how the changes have affected your predictions.
> suggest_network_trimming_to_size(GEMLI_items, max_size=2, cutoff=50) # left image
> GEMLI_items_post_processed = trim_network_to_size(GEMLI_items, max_size=2, cutoff=50)
> visualize_as_network(GEMLI_items_post_processed, cutoff=50) # right image
In this example we'll be visualizing GEMLI results from a lineage-annotated dataset of murine intestinal crypts. Intestinal crypts originate from one or few intestinal stem cells, and are composed of a number of different cell types. In the data we find cells from individual crypts that range from just three up to forty-four related cells. The scRNA-seq dataset is derived from Bues et al. 2022 (PMID 35165449) and is publically available under Gene Expression Omnibus accession number GSE148093.
First we load the example data. Here we already predicted the lineages using GEMLI and therefore do not include a count matrix, but rather start with the predictions right away.
> load('GEMLI_crypts_example_data_matrix.RData')
> load('GEMLI_crypts_example_barcode_information.RData')
We then create a GEMLI items list. This list is used to store the data, and create and store the outputs of GEMLI (for details check example one).
> GEMLI_items_crypts = list()
> GEMLI_items_crypts[['prediction']] = Crypts
> GEMLI_items_crypts[['barcodes']] = Crypts_bc_dict
To visualize large lineages we'll use three different network layout algorithms: Fruchterman-Reingold, Kamada-Kawai, and grid. Each of them has advantages and disadvantages.
(1) Fruchterman-Reingold dispalys the cells of individual predicted crypts close together with ample space between crypts. This makes it hard to see connections within individual crypts but allows to get a good overview of individual structures.
(2) Kamada-Kawai spaces individual cells well, so we can see individual connections between them. It may, however, happen that two different predicted crypts are partially overlayed, as can be seen for dark red and bright green lineages on the right side of the plot.
(3) When the network is layed out as a grid, one gets generally a good overview of the predicted lineages and their connections, but it's hard to see which connections belongs to which cell in the same row.
> visualize_as_network(GEMLI_items_crypts, cutoff=70, display_orphan=F, max_edge_width=1, ground_truth=T, include_labels=F, layout_style="fr") # first image
> visualize_as_network(GEMLI_items_crypts, cutoff=70, display_orphan=F, max_edge_width=1, ground_truth=T, include_labels=F, layout_style="kk") # second image
> visualize_as_network(GEMLI_items_crypts, cutoff=70, display_orphan=F, max_edge_width=1, ground_truth=T, include_labels=F, layout_style="grid") # third image
The cells of individual intestinal crypts can be assigned to different cell types. This information can be added to the GEMLI items list as 'cell_type' slot in the form of a dataframe with column 'cell.ID' and 'cell.type'.
> load('GEMLI_crypts_example_cell_type_annotation.RData')
> GEMLI_items_crypts[['cell_type']] = Crypts_annotation
The visualization of the lineage predictions can now be colored by the cell type annotation. This allows to see the composition of individual intestinal crypts.
> visualize_as_network(GEMLI_items_crypts, cutoff=70, max_edge_width=5, display_orphan=F, include_labels=F, ground_truth=T, highlight_FPs=T, layout_style="kk", cell_type_colors=T)
Specific colors can be assigned to specific cell types by adding a dataframe with column 'cell.type' and 'color' in the GEMLI items list slot 'cell_type_color'. This can also allow to highlight just one or two selected cell types.
> cell.type <- unique(GEMLI_items_crypts[['cell_type']]$cell.type)
> color <- c("#5386BD", "skyblue1", "darkgreen", "gold", "red", "darkred", "black")
> Cell_type_color <- data.frame(cell.type, color)
> GEMLI_items_crypts[['cell_type_color']] = Cell_type_color
>
> visualize_as_network(GEMLI_items_crypts, cutoff=70, max_edge_width=5, display_orphan=F, include_labels=F, ground_truth=T, highlight_FPs=T, layout_style="kk", cell_type_colors=T)
To get an even more quantitative overview of the cell type composition of individual intestinal crypts, we can use the function 'cell_type_composition_plot'. To do so we have to run the 'prediction_to_lineage_information' function first.
> GEMLI_items_crypts = prediction_to_lineage_information(GEMLI_items_crypts, cutoff=50)
> cell_type_composition_plot(GEMLI_items_crypts, cell_type_colors=T, type=c("bubble"))
With other 'type' parameters we can output using the same function, and upsetR or plain table of the lineage numbers with different cell type compositions.
> cell_type_composition_plot(GEMLI_items_crypts, ground_truth=F, cell_type_colors=T, type=c("upsetR"))
> cell_type_composition_plot(GEMLI_items_crypts, ground_truth=F, cell_type_colors=T, type=c("plain"))
combi n
Entero 1
Entero__Goblet__PIC__Stem 1
Entero__Goblet__PIC__Stem__TA 1
Entero__PIC__Paneth__Regstem__Stem__TA 1
Entero__Stem__TA 1
Goblet__PIC__Regstem__Stem__TA 1
Goblet__PIC__Stem__TA 1
Goblet__Regstem__Stem__TA 1
Goblet__Stem__TA 2
PIC 3
PIC__Stem__TA 2
Paneth 2
Paneth__Regstem__Stem 1
Paneth__Stem 1
Regstem__Stem 1
Stem 3
TA 1
For our third example we'll be looking at cell fate decisions in a scRNA-seq dataset of human breast cancer encompassing both ductal carcinoma in situ (DCIS) and invasive tumor (inv_tumor) cells. We'll be working with a subset of this data for fast processing. No ground truth is available. We will study the fate transition from DCIS to invasive tumor cells. The data is derived from a public 10X Genomics dataset associated to the following preprint bioRxiv 2022.10.06.510405; doi: https://doi.org/10.1101/2022.10.06.510405 and downloaded from https://www.10xgenomics.com/products/xenium-in-situ/preview-dataset-human-breast (April 2023).
First we load the example data. We already predicted the lineages and extracted the predicted lineages using GEMLI. Furthermore we load the previously generated cell type information for all cells in the dataset.
> load('GEMLI_cancer_example_norm_count.RData')
> load('GEMLI_cancer_example_predicted_lineages.RData')
> load('GEMLI_cancer_example_cell_type_annotation.RData')
We then create a GEMLI items list. This list is used to store the data, and create and store the outputs of GEMLI (for details check example one).
> GEMLI_items = list()
> GEMLI_items[['gene_expression']] = Cancer_norm_count
> GEMLI_items[['predicted_lineage_table']] = Cancer_predicted_lineages
> GEMLI_items[['cell_type']] = Cancer_annotation
To get an overview of the symmetric and asymmetric cell lineages present in the data, we can use the function 'cell_type_composition_plot' with parameter 'type' ="plain" or "upsetR".
> cell_type_composition_plot(GEMLI_items, type=c("plain"))
combi n
DCIS 250
DCIS__inv_tumor 25
inv_tumor 203
>
> cell_type_composition_plot(GEMLI_items, type=c("upsetR"))
We extract now predicted cell lineages with members in only one cell type (symmetric) or in two or more cell types (asymmetric). To analyze the transition from DCIS to invasive breast cancer we will extract symmetric DCIS, asymmetric DCIS and invasive tumor, and symmetric inv_tumor lineages. To exclude lineages with a too large asymmetry, we set a threshold to extract asymmetric lineages containing at least 10% of each cell type. The function output is stored in GEMLI_items 'cell_fate_analysis' item. It is a data frame with a column cell.fate with label sym or asym and cell type separated by an underscore. This cell.fate designation allows to subsequently analyze only a specific cell type in asymmetric cell lineages.
> GEMLI_items<-extract_cell_fate_lineages(GEMLI_items, selection=c("inv_tumor", "DCIS"), unique=FALSE, threshold=c(10,10))
>
> GEMLI_items[['cell_fate_analysis']][1:10,]
cell.ID clone.ID cell.type cell.fate
AAACCCACATCCGTGG-3 2826 DCIS NA_DCIS
AAACCCATCCTTATAC-4 54 DCIS asym_DCIS
AAACGAACAACACGTT-2 1466 inv_tumor sym_inv_tumor
AAAGAACCAACAGCTT-3 726 DCIS sym_DCIS
AAAGAACGTCGAATGG-2 1467 inv_tumor sym_inv_tumor
AAAGGATCAGAGTTCT-4 4383 inv_tumor sym_inv_tumor
AAAGGATTCGCCAATA-4 4385 inv_tumor sym_inv_tumor
AAAGGGCCAGTCGGAA-3 754 inv_tumor sym_inv_tumor
AAAGGGCGTAAGAACT-1 10 inv_tumor sym_inv_tumor
AAAGGGCGTAGTTCCA-1 11 inv_tumor sym_inv_tumor
>
> table(GEMLI_items[['cell_fate_analysis']]$cell.fate)
asym_DCIS asym_inv_tumor NA_DCIS sym_DCIS sym_inv_tumor
77 96 225 273 744
Based on the symmetric and asymmetric lineages we extracted, we will now call differentially expressed genes (DEG) specific for cells of specific cell types in specific lineages types. To analyze the transition from DCIS to invasive breast cancer, we notable call DEG for DCIS cells in asymmetric and symmetric lineages. These are genes specific to DCIS cells at the start of the transition.
> GEMLI_items<-cell_fate_DEG_calling(GEMLI_items, ident1="sym_DCIS", ident2="asym_DCIS", min.pct=0.05, logfc.threshold=0.1)
>
> GEMLI_items[['DEG']][1:10,]
p_val avg_log2FC pct.1 pct.2 p_val_adj
DCAF7 4.308339e-15 -1.0275658 0.985 1.000 1.086520e-10
NRAS 2.630887e-11 -1.2821469 0.516 0.766 6.634834e-07
LINC01999 2.824883e-11 -1.4817245 0.736 0.961 7.124072e-07
CSDE1 7.762498e-10 -0.8055031 0.952 0.974 1.957624e-05
S100A10 3.676826e-09 1.1729996 0.839 0.610 9.272588e-05
MAN1A2 8.900499e-09 -0.8487466 0.868 0.961 2.244617e-04
APPBP2-DT 9.767327e-09 -1.0797417 0.495 0.792 2.463222e-04
PVALB 9.882215e-09 -1.1913536 0.304 0.623 2.492196e-04
CDH2 1.169570e-08 -1.6410588 0.264 0.558 2.949538e-04
CSTA 2.204180e-08 -1.0506112 0.198 0.506 5.558721e-04
>
> DEG_volcano_plot(GEMLI_items, name1="Sym_DCIS", name2="Asym_DCIS")
If you use the package, please cite A.S. Eisele*, M. Tarbier*, A.A. Dormann, V. Pelechano, D.M. Suter | "Gene-expression memory-based prediction of cell lineages from scRNA-seq datasets" | Nature Communications 15, 2744 (2024). https://doi.org/10.1038/s41467-024-47158-y