Defining an mxFDA object

The mxfda package contains tools for analyzing spatial single-cell data, including multiplex imaging and spatial transcriptomics, using methods from functional data analysis. Analyses for this package are executed and stored using an S4 object of class mxFDA. This vignette outlines how to set up an mxFDA object from spatial single cell imaging data, how to calculate spatial summary functions, and exploratory data analysis and visualization of these spatial summary functions. Details on how to perform downstream analysis and feature extraction using functional principal component analysis can be found in the separate vignette vignette("mx_fpca"). To perform functional regression on spatial summary functions from multiplex imaging data, see vignette("mx_funreg").

library(mxfda)
library(tidyverse)

VectraPolarisData

Examples in this package use data adapted from the VectraPolarisData package on Bioconductor’s ExperimentHub. This package contains data from two multiplex imaging experiments conducted at the University of Colorado Anschutz Medical Campus. A shortcourse on single-cell multiplex imaging using these data is available here.

Data has been preprocessed and stored directly in the {mxfda} package. Available datasets are ovarian_FDA and lung_df. This vignette will focus on the lung_df dataset, which contains a subset of 50 subjects from a multiplex imaging study of non-small cell lung carcinoma described in Johnson et al. (2021). Each subject has 3-5 multiplex images, which come from different regions of interest (ROIs) in the tumor. Each ROI will be considered a “sample” when constructing the {mxfda} object. We load the lung cancer data below.

data(lung_df)

Setting up the mxFDA object

The central object used with the {mxfda} package is the mxFDA object. These objects are created with make_mxfda() and hold everything from the raw spatial data to fit functional data models using derived spatial summary functions. To save space on large samples, the metadata is kept separately from the spatial data and when needed, is exported and merged together. Slots in the mxFDA object are designated as follows:

Make mxFDAobject

Using the ?lung_df in the {mxfda} package, columns with repeated data pertaining to the sample-level information are extracted and stored in a data frame called clinical while the cell-level information is kept in a long data frame (make_mxfda() also accepts cell-level information as a list of data frames). The spatial parameter in the make_mxfda() function can be left blank if using a spatial metric derived from external functions. The final 2 parameters for the ?lung_df mxFDA object are the subject_key and the sample_key. The sample_key is a column name that appears both in the metadata and spatial and denotes unique samples while the subject_key is a column name in metadata that ties the samples to metadata; if the data contains one sample per subject then sample_key and subject_key is a 1:1, but if multiple samples per subject, subject_id will be repeated.

clinical = lung_df %>%
  select(image_id, patient_id, patientImage_id, gender, age, survival_days, survival_status, stage) %>%
  distinct()

spatial = lung_df %>%
  select(-image_id, -gender, -age, -survival_days, -survival_status, -stage)

mxFDAobject = make_mxfda(metadata = clinical,
                         spatial = spatial,
                         subject_key = "patient_id",
                         sample_key = "patientImage_id")

Note that the object created has class mxFDA.

class(mxFDAobject)
#> [1] "mxFDA"
#> attr(,"package")
#> [1] "mxfda"

Spatial summary functions based on point processes

The {mxfda} package provides methods for analyzing spatial relationships between cell types in single cell imaging data based on point process theory. The location of cells in image samples are treated as following a point process, realizations of a point process are called “point patterns”, and point process models seek to understand correlations in the spatial distributions of cells. Under the assumption that the rate of a cell is constant over an entire region of interest a point pattern will exhibit complete spatial randomness (CSR), and it is often of interest to model whether cells deviate from CSR either through clustering or repulsion. When the rate of a cell is not constant, this CSR assumption is violated and to estimate CSR it is recommended to use permutations (see C. Wilson et al. (2022) for explanation).

Spatial summary statistics can be calculated to quantify the clustering and co-occurrence of cells in a circular region with a particular radius r. Typically univariate (one cell type) or bivariate (two cell types) summary statistics are reported, and inference is obtained by comparing the observed spatial summary statistic to that obtained under CSR. A popular quantity is Ripley’s K(r), which studies the number of neighbors to a particular point within radius r, and has univariate and bivariate implementations in the {spatstat} package (Baddeley, Rubak, and Turner (2015)). Ripley’s K is characterized by clustering or repulsion depending on whether it is above or below the theoretical value of \(\pi r^2\). Other spatial summary statistics analyze the distance to a neighbor, and can be interpreted as probabilities of observing a particular cell type within a radius r. One of these metrics, G(r), or the nearest neighbor distance distribution, is the cumulative distribution function of an exponential random variable. More detailed overviews of spatial summary functions for multiplex imaging data are provided in C. M. Wilson et al. (2021) and Wrobel, Harris, and Vandekar (2023).

Univariate summary functions

Below we calculate univariate Ripley’s K to summarize the spatial relationship among immune cells in each image. The {mxfda} package accomplishes this with the function extract_summary_functions(). Either univariate or bivariate can be calculated with this function depending on the choice supplied to the extract_func argument. To calculate a univariate spatial summary we supply univariate to the extract_func argument. The summary function that is calculated depends on the function supplied to summary_fun which is one of Kest(), Gest(), or Lest() from the {spatstat.explore} package (Kcross(), Gcross(), or Lcross() for bivariate methods). Other options include supplying a vector of radius values through r_vec, and the a specific edge correction (see Baddeley, Rubak, and Turner (2015)). We calculate the K function across a range of radii from 0 to 100 and use the isotropic (“iso”) edge correction. See Baddeley, Rubak, and Turner (2015) for more details on edge corrections for Ripley’s K and nearest neighbor G. Options permute_CSR and permutations can be used if interested in using a measure of the sample-specific CSR instead of the theoretical, the column specified with markvar just needs to contain 1 level more than used for summary_func (2 levels for univariate and 3 levels for bivariate) otherwise falls back to theoretical.

mxFDAobject = extract_summary_functions(mxFDAobject,
                                        extract_func = univariate,
                                        summary_func = Kest,
                                        r_vec = seq(0, 100, by = 1),
                                        edge_correction = "iso",
                                        markvar = "immune",
                                        mark1 = "immune")
#> Using Theoretical Complete Spatial Randomness for Ripley's K
#>  ■■■■■■■■■■■■■■■                   46% |  ETA:  3s
#>  ■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■    98% |  ETA:  0s

Running this code will calculate univariate Ripley’s K function to measure spatial clustering of immune cells for each sample, and will store these spatial summary functions in the univariate_summaries slot of the mxFDAobject. To access this slot and view the extracted summary functions, type:

mxFDAobject@univariate_summaries$Kest
#> # A tibble: 24,947 × 6
#>    patientImage_id     r sumfun    csr fundiff `immune cells`
#>    <chr>           <dbl>  <dbl>  <dbl>   <dbl>          <int>
#>  1 2_1                 0      0   0       0                 6
#>  2 2_1                 1      0   3.14   -3.14              6
#>  3 2_1                 2      0  12.6   -12.6               6
#>  4 2_1                 3      0  28.3   -28.3               6
#>  5 2_1                 4      0  50.3   -50.3               6
#>  6 2_1                 5      0  78.5   -78.5               6
#>  7 2_1                 6      0 113.   -113.                6
#>  8 2_1                 7      0 154.   -154.                6
#>  9 2_1                 8      0 201.   -201.                6
#> 10 2_1                 9      0 254.   -254.                6
#> # ℹ 24,937 more rows

Note that the summaries are returned as a dataframe. The variable sumfun is the estimated summary function value, csr is the theoretical value under complete spatial randomness, and fundiff = sumfun-csr describes the “degree of clustering beyond what is expected due to chance; in downstream analysis we will use the fundiff covariate.

Plotting the mxFDA object

{mxfda} has S4 methods for visualization implemented via the plot() function (see ?plot.mxFDA for details). The first argument is the mxFDA object followed by a few options that depend on what plot output is desired. Here, we want to plot the univariate summary that we just calculated, which was the K function. By passing in what = 'uni k', the plot function will extract the univariate K results. We also need to tell plot() what column is the y-axis which can be 'sumfun' for the observed value, 'csr' for the theoretical value of complete spatial randomness (CSR), or 'fundiff' which is the difference between the observed K measure and the theoretical CSR. The output of plot() is a {ggplot2} object which can then be easily added to/manipulated as any ggplot plot would.

NOTE: These are the columns when calculating using the extract_summary_function() of {mxfda} but if summary data is added from elsewhere with add_summary_function() then those column names will have to be used.

plot(mxFDAobject, y = "fundiff", what = "uni k") +
  geom_hline(yintercept = 0, color = "red", linetype = 2)

Bivariate summary functions

The extract_summary_functions() function can also be used to extract bivariate summaries comparing spatial clustering of 2 cell types. We will look at relationship between T-cells and macrophages. There are a few images that have fewer than 5 T-cells or macrophages, which makes estimation of spatial summary functions less stable for those images. To look at T-cells and macrophages, the data phenotypes and cell locations have to be in long format so we first create a variable with the cell types ('phenotype') from the lung_df.

lung_df = lung_df %>%
  mutate(phenotype = case_when(phenotype_cd8 == "CD8+" ~ "T-cell",
                               phenotype_cd14 == "CD14+" ~ "macrophage",
                               TRUE ~ "other"),
         phenotype = factor(phenotype))

We then recreate the mxFDAobject

spatial = lung_df %>%
  select(-image_id, -gender, -age, -survival_days, -survival_status, -stage)

mxFDAobject = make_mxfda(metadata = clinical,
                         spatial = spatial,
                         subject_key = "patient_id",
                         sample_key = "patientImage_id")

Now we calculate the bivariate G function, but can replace Gcross() with Lcross() or Kcross() to estimate the L or K bivariate functions instead. The argument markvar takes the variable that we created above called 'phenotype', and the 2 cell types that we are interested in calculating the bivariate G for are 'T-cell' and 'macrophage' so we provide them to mark1 and mark2, respectively.

mxFDAobject = extract_summary_functions(mxFDAobject,
                summary_func = Gcross,
                extract_func = bivariate,
                r_vec = seq(0, 100, by = 1),
                edge_correction = "rs",
                markvar = "phenotype",
                mark1 = "T-cell",
                mark2 = "macrophage")
#> Using Theoretical Complete Spatial Randomness for Bivariate Nearest Neighbor G

Plotting bivariate G

Just like with the univariate plots, we can use the plot() function to plot our mxFDA object results. The what now is 'bi g', 'bivar g', or 'bivariate g'.

plot(mxFDAobject, y = "fundiff", what = "bi g") +
  geom_hline(yintercept = 0, color = "red", linetype = 2)
#> Warning: Removed 712 rows containing missing values or values outside the scale range
#> (`geom_line()`).

Entropy

Below we show how to calculate spatial summary functions using a multivariate metrics based on spatial entropy from Vu et al. (2023). Briefly, this approach measures the heterogeneity in the tissue sample by considering all cell types in a given radius.

This framework is able to capture the diversity in cellular composition, such as similar proportions across cell types or dominance of a single type, at a specific distance range.Spatial patterns, including clustered, independent, or regular, among cell types can also be acquired. If cells of different types are randomly scattered on an image, the spatial entropy at any given distance is close to zero. The spatial entropy deviates from zero if spatial patterns of cell types in a given local neighborhood are different from the global pattern. See Vu et al. (2023) for a mathematical representation of this index.

Below we calculate spatial summary functions based on entropy in the lung cancer dataset.

#load in lung DF
data(lung_df)
#filter to only the clinical information
clinical = lung_df %>%
  select(image_id, patient_id, patientImage_id, gender, age, survival_days, survival_status, stage) %>%
  distinct()
#filter to cell information
spatial = lung_df %>%
  select(-image_id, -gender, -age, -survival_days, -survival_status, -stage)%>%
  mutate(phenotype = case_when(phenotype_cd8 == "CD8+" ~ "T-cell",
                               phenotype_cd14 == "CD14+" ~ "macrophage",
                               TRUE ~ "other"),
         phenotype = factor(phenotype))
#create the mxfda object
mxFDAobject = make_mxfda(metadata = clinical,
                         spatial = spatial,
                         subject_key = "patient_id",
                         sample_key = "patientImage_id")
#run entropy
mxFDAobject = extract_summary_functions(mxFDAobject,
                                        extract_func = bivariate,
                                        summary_func = entropy,
                                        r_vec = seq(0, 100, by = 1),
                                        edge_correction = "iso",
                                        markvar = "phenotype",
                                        mark1 = "T-cell",
                                        mark2 = "macrophage")
#plot
mxFDAobject@bivariate_summaries$entropy %>%
  ggplot() +
  geom_line(aes(x = r, y = spatial_entropy,
                group = patientImage_id, color = patientImage_id),
            alpha = 0.2) +
  theme(legend.position = "none")

Exploring the S4 object

Another useful function is ?summary.mxFDA which feeds into the summary() method. Either typing the name of the object or wrapping it in the summary function will provide information like the number of subjects, samples, if spatial summary functions have been calculated, and functional data analyses that have been run.

mxFDAobject
#> mxFDA Object:
#>  Subjects: 50
#>  Samples: 247
#>  Has spatial data
#>  Univariate Summaries: None
#>  Bivariate Summaries: Gcross
#>  Multivariate Summaries: None
#> FPCs not yet calculated
#> MFPCs not yet calculated
#> FCMs not yet calculated
#> MFCMs not yet calculated
#> Scalar on Functional Regression not calculated

SpatialTIME

Sometimes other summary functions or normalizations are run outside of the {mxfda} package but the end goal is to still run functional data analysis. Other packages, such as {spatialTIME} (Creed et al. (2021)) provide methods for fast calculation of functions in {spatstat} with permutation estimates of complete spatial randomness that are more robust than theoretical CSR estimates, especially when tissue samples have holes that violate the assumption of a homogeneous point pattern (see C. Wilson et al. (2022)). Lets look at how to perform the estimation of univariate nearest neighbor G with {spatialTIME}.

The central object of spatialTIME is the mIF object, that contains a list of spatial data frames, a data frame of sample-level summaries, and a data frame for the metadata (‘clinical’). From creating the mxFDA object, we have a spatial data frame and the clinical data, now we have to convert them into something that works with spatialTIME. The steps below will be:

  1. convert positive/negative to 1/0 integers
  2. identify which columns in the spatial data frame are cell types
  3. convert spatial data frame to a list, where each element is a unique sample
  4. create a summary data frame from the spatial list for the number and proportion of positive cells for each phenotype
#Step 1
spatialTIME_spatial_df = spatial %>% 
  select(-phenotype) %>%
  mutate(across(phenotype_ck:phenotype_cd4, ~ ifelse(grepl("\\+", .x), 1, 0))) %>%
  relocate(patientImage_id, .before = 1)

#Step 2
cell_types = colnames(spatialTIME_spatial_df) %>% grep("phenotype", ., value = TRUE)

#Step 3
spatial_list = split(spatialTIME_spatial_df, spatial$patientImage_id)

#Step 4
summary_data = lapply(spatial_list, function(df){
  df %>%
    #group by sample ID to maintain ID column
    group_by(patient_id, patientImage_id) %>%
    #find number of positive
    reframe(across(!!cell_types, ~ sum(.x)),
              `Total Cells` = n()) %>%
    #calculate proportion
    mutate(across(!!cell_types, ~.x/`Total Cells` * 100, .names = "{.col} %"))
}) %>%
  #bind the rows together
  do.call(bind_rows, .)

With the spatial list, clinical, and summary data the mIF object can be constructed. For best computation efficiency, use >v1.3.4.


library(spatialTIME)
#> spatialTIME version:
#> 1.3.4.5
#> If using for publication, please cite our manuscript:
#> https://doi.org/10.1093/bioinformatics/btab757

#make mif
mif = create_mif(clinical_data = clinical,
                 sample_data = summary_data,
                 spatial_list = spatial_list[1:50],
                 patient_id = "patient_id",
                 sample_id = "patientImage_id")

Deriving spatial metrics with the mIF object is really easy but does take some time. Will only do 10 permutations here to estimate the complete spatial randomness measure of nearest neighbor G and a reduced sampling, or 'rs', edge correction. To make the run faster, will look at only cytotoxic T cells (CD8+) and helper T cells (CD4+).

mif = NN_G(mif, mnames = cell_types[c(2, 6)],
           r_range = 0:100, num_permutations = 10, 
           edge_correction = "rs", keep_perm_dis = FALSE,
           workers = 1, overwrite = TRUE, xloc = "x", yloc = "y")

With spatialTIME, all cell types (markers) are added to the data frame. We can visualize both CD8+ and CD4+ with ggplot.

mif$derived$univariate_NN %>%
    ggplot() +
    geom_line(aes(x = r, y = `Degree of Clustering Permutation`, color = patientImage_id), alpha = 0.4) +
    facet_grid(~Marker) +
  theme(legend.position = "none")
#> Warning: Removed 1086 rows containing missing values or values outside the scale range
#> (`geom_line()`).

Exporting the spatial summary function data from the mIF object is the same as accessing the list object. However, we need to make sure that the data that we use with the mxFDA object contains only a single cell types results. This is to make sure that when modeling we aren’t mixing up different cells. Below is the extraction and filtering of the new univariate G results and keeping only the cytotoxic T cell results.

uni_g = mif$derived$univariate_NN %>%
  filter(grepl("cd8", Marker))

With the derived univariate nearest neighbor G for CD8+, it can be added to an mxFDA object with add_summary_function(). To show this, first will create a new mxFDA object with an empty spatial slot then add the new summary function results.

#make mxFDA object 
mxFDA_spatialTIME = make_mxfda(metadata = clinical,
                               spatial = NULL,
                               subject_key = "patient_id",
                               sample_key = "patientImage_id")
#add summary data
mxFDA_spatialTIME = add_summary_function(mxFDAobject,
                                         summary_function_data = uni_g,
                                         metric = "uni g")

Can now use the mxfda plot method with the new data and continue with analyses as would be done if using the internal extract_summary_function().

plot(mxFDA_spatialTIME, y = "Degree of Clustering Permutation", what = "uni g")
#> Warning: Removed 480 rows containing missing values or values outside the scale range
#> (`geom_line()`).

References

Baddeley, Adrian, Ege Rubak, and Rolf Turner. 2015. Spatial Point Patterns: Methodology and Applications with r. CRC press.
Creed, Jordan H, Christopher M Wilson, Alex C Soupir, Christelle M Colin-Leitzinger, Gregory J Kimmel, Oscar E Ospina, Nicholas H Chakiryan, et al. 2021. spatialTIME and iTIME: R package and Shiny application for visualization and analysis of immunofluorescence data.” Bioinformatics 37 (23): 4584–86. https://doi.org/10.1093/bioinformatics/btab757.
Johnson, Amber M, Jennifer M Boland, Julia Wrobel, Emily K Klezcko, Mary Weiser-Evans, Katharina Hopp, Lynn Heasley, et al. 2021. “Cancer Cell-Specific Major Histocompatibility Complex II Expression as a Determinant of the Immune Infiltrate Organization and Function in the NSCLC Tumor Microenvironment.” Journal of Thoracic Oncology 16 (10): 1694–704.
Vu, Thao, Souvik Seal, Tusharkanti Ghosh, Mansooreh Ahmadian, Julia Wrobel, and Debashis Ghosh. 2023. “FunSpace: A Functional and Spatial Analytic Approach to Cell Imaging Data Using Entropy Measures.” PLOS Computational Biology 19 (9): e1011490.
Wilson, Christopher M, Oscar E Ospina, Mary K Townsend, Jonathan Nguyen, Carlos Moran Segura, Joellen M Schildkraut, Shelley S Tworoger, Lauren C Peres, and Brooke L Fridley. 2021. “Challenges and Opportunities in the Statistical Analysis of Multiplex Immunofluorescence Data.” Cancers 13 (12): 3031.
Wilson, Christopher, Alex C Soupir, Ram Thapa, Jordan Creed, Jonathan Nguyen, Carlos Moran Segura, Travis Gerke, Joellen M Schildkraut, Lauren C Peres, and Brooke L Fridley. 2022. “Tumor Immune Cell Clustering and Its Association with Survival in African American Women with Ovarian Cancer.” PLoS Computational Biology 18 (3): e1009900.
Wrobel, Julia, Coleman Harris, and Simon Vandekar. 2023. “Statistical Analysis of Multiplex Immunofluorescence and Immunohistochemistry Imaging Data.” In Statistical Genomics, 141–68. Springer.