Introduction to spatialTIME

```{css, echo=FALSE} body .main-container { max-width: 1500px !important; width: 1500px !important; } body { max-width: 1500px !important; }

```r
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

set.seed(126)
library(spatialTIME)
library(dplyr)
library(ggplot2)
library(viridis)
library(tidyr)
library(tidyselect)

Background for Tissue Microarray (TMA) Data

A tissue microarray (TMA) is an array of samples which are obtained by taking a slice of a biopsied FFPE tumor. Each individual slice is referred to as a core. Each core is placed on a TMA and is then stained with multiple antibodies and fluorophores which illuminate when a laser is shined at them with varying wavelengths. The intensity is measured and then a random forest algorithm is used to classify the cells as being positive for a particular marker which allows us to phenotype cells. A schematic of this process is provided in Figure 1.

Create Multiplex ImmunoFlourescent (mif) Object

spatialTIME functions use a custom mif object which can be created using create_mif. The mif object has 6 slots storing the:

We include one example of a clinical and sample dataset which have a total of 229 patients with one core. Out of those 229 samples, only 5 are included in our package.

# Make sure the variable types are the same for deidentified_id and 
# deidentified_sample in their corresponding datasets
x <- create_mif(clinical_data = example_clinical %>% 
                  mutate(deidentified_id = as.character(deidentified_id)),
                sample_data = example_summary %>% 
                  mutate(deidentified_id = as.character(deidentified_id)),
                spatial_list = example_spatial,
                patient_id = "deidentified_id", 
                sample_id = "deidentified_sample")

x #prints a summary of how many patients, samples, and spatial files are present

Plotting Cores

An individual plot for each core (each sample) is created. Plots can be assigned to an R object, such as within the empty derived slot and printed to a PDF if a file name is provided.

When studying phenotype and individual markers, note that it is important to have the individual before the phenotype markers. This will ensure that the phenotype that are derived by multiple markers are not plotted over by the individual marker. For instance, below the the first plot appears to have no cytotoxic T cells (CD3+ and CD8+), but then the order is changed we see the cytotoxic T cells. Moral of the story: Put the marker combinations before the single markers.

mnames_bad <- c("CD3..CD8.","CD3..FOXP3.","CD3..Opal.570..Positive",
                "CD8..Opal.520..Positive","FOXP3..Opal.620..Positive", 
                "PDL1..Opal.540..Positive", "PD1..Opal.650..Positive")

# Used to make the legends in both plots below be in same order and use the 
# same coloring scheme for the purpose making a common legend

values = turbo(length(mnames_bad))
names(values) = mnames_bad

#add an element in the `derived` object position
x<- plot_immunoflo(x, plot_title = "deidentified_sample",  mnames = mnames_bad,
                   cell_type = "Classifier.Label")

bad_names <- x[["derived"]][["spatial_plots"]][[4]] + 
  theme(legend.position = 'bottom') + 
  scale_color_manual(breaks = mnames_bad,
                     values = values,
                     labels = mnames_bad %>%
                       gsub("..Opal.*", "+", .) %>% 
                       gsub("\\.\\.", "+", .) %>% 
                       gsub("\\.", "+", .))

mnames_good <- c("CD3..Opal.570..Positive","CD8..Opal.520..Positive",
                 "FOXP3..Opal.620..Positive","PDL1..Opal.540..Positive",
                 "PD1..Opal.650..Positive","CD3..CD8.","CD3..FOXP3.")

x <- plot_immunoflo(x, plot_title = "deidentified_sample", mnames = mnames_good, 
                    cell_type = "Classifier.Label")

good_names <- x[["derived"]][["spatial_plots"]][[4]] + 
  theme(legend.position = 'bottom') + 
  scale_color_manual(breaks = mnames_good, 
                     values = values[match(mnames_good, names(values))],
                     labels = mnames_good %>%
                       gsub("..Opal.*", "+", .) %>% 
                       gsub("\\.\\.", "+", .) %>% 
                       gsub("\\.", "+", .))

x$sample %>% filter(deidentified_sample == 'TMA3_[9,K].tif') %>% select(c(2, 4:15)) %>%
  pivot_longer(cols = 2:13, names_to = 'Marker', values_to = 'Count')

ggpubr::ggarrange(plotlist = list(bad_names, good_names), labels = c('B', 'G'),
                  common.legend = TRUE, legend = 'bottom')

Estimating the degree of spatial clustering with Count Based Methods

Univariate

Count Based Methods

Ripley's $K$ measures the average number of neighboring cells across each cell. That is, the average (over all cells) number of cells within a specified radius of a cell. Ripley's $K$ is computed as follows:

$$\hat{K}(r) = \frac{1}{n(n-1)}\sum_{i=1}^{n}\sum_{j\neq i}w_{ij}{\bf 1}{(d(x_i,x_j)\le r)},$$

where $r$ is the specified radius, $d(x_i,x_j)$ is the distance between the $i^{th}$ and $j^{th}$ cell, ${\bf 1}A$ is indicator function of event $A$, and $w{ij}$ is the weights that are assigned for border corrections. The expected value of $\hat{K}(r)$ is $\pi r^2$, thus $\hat{K}$ is expected to grow as a quadratic function of $r$.

There are several edge corrections. Our studies have included a small number of cells and we recommend using the 'isotropic' or 'translational' edge correction, as opposed to the 'border' edge correction. The main goal of the edge correction is to account for the fact that there are unobserved points outside of the region, and the assumption is that the location of these cells has the same distribution as the study region. An excellent description of these corrections are provided here.

Distance Based Measures

The distribution of the nearest neighbor distances, $\hat{G}(r)$, can be studied and is computed by

$$\hat{G}(r) = \frac{1}{n}\sum_{i=1}^{n}{\bf 1}(\min_{j}({d(x_i,x_j)}\le r),$$

which is interpreted as the proportion of cells whose distance to its nearest neighbor is less than $r$. Notice that there is not a weighting factor for each pair of points as we saw above. The edge correction in these methods, reduced sample (rs) and Hanisch (han), simply have different cell inclusion conditions. The reduced sample correction is similar to the border correction for count based methods, where the middle chunk of the area of interest is studied. The Hanisch border correction leaves out points whose $k^{th}$ neighbor can not be in the area of interest. For more information about these border corrections, see the following article.

Need for Permutations

An underlying assumption used for many spatial clustering metrics is that the cells are randomly distributed across the region, no evidence of clustering or repulsion, and that the cell intensity is constant across the entire region. This assumption is the so-called complete spatial randomness (CSR). Damage can occur to tissue cores due to how they are collected. This damage can lead to rips and tears in the cores which results in regions where it appears that cells are not located and is not actually the case. Due to these violations of the CSR assumption, the theoretical estimate for CSR may not be accurate. To address this the cell positivity can be permuted across all observed locations and the permutation distribution of $K$, $L$, $M$, and $G$ is a core specific measure of CSR.

Implementation

Univariate Count-Based Methods

The ripleys_k function reports a permuted and theoretical estimate of CSR, the observed value for $K$ (method = 'K'), $L$ (method = L), or Marcon $M$ (method = 'M'), and the full permutation distribution of $K$, $L$, and $M$ of keep_perm_dis = TRUE.

Currently, the number of permutations is 10, but this should be increased to at least 100 for a more reliable estimate of the mean.

x <- ripleys_k(mif = x, mnames = mnames_good, 
               num_permutations = 10, method = "K",
               edge_correction = 'translation', r = seq(0,100,10),
               keep_perm_dis = FALSE, workers = 1)

# This will keeps the colors in evx$ery plot for the remainder of the vignette compatible 
values = turbo(length(unique(x$derived$univariate_Count$deidentified_sample)))
names(values) = unique(x$derived$univariate_Count$deidentified_sample)

x$derived$univariate_Count  %>%
  filter(Marker != 'PDL1..Opal.540..Positive') %>%
  ggplot(aes(x = r, y = `Degree of Clustering Permutation`)) +
  geom_line(aes(color = deidentified_sample), show.legend = FALSE) +
  facet_wrap(Marker~., scales = 'free') + theme_bw() + 
  scale_color_manual(values = values)

Positive values for degree of cluster when using method = 'K' and method = 'L' indicates evidence of spatial clustering, while negative values correspond to spatial regularity. On the other hand, if using method = 'M' then values larger than 1 correspond to clustering and values less than one correspond to regularity. Also, these values can be interpreted as the the percent difference from spatial clustering, for example if $M=0.5$ that means there is 50% less spatial clustering than expected under CSR.

Sensitivity of r

There is no clear way to select which value of $r$ to use for a particular analysis. Below we illustrate how the degree of clustering can change with the value of $r$. Notice that for values of $r > 20$ that there is very little difference in the ordering between the samples (though the degree of spatial clustering can changes dramatically), while very small values $r$ would have different results.

x$derived$univariate_Count  %>%
  filter(Marker == 'CD3..CD8.') %>%
  inner_join(x$clinical,.) %>%
  ggplot(aes(shape = status, y = `Degree of Clustering Permutation`, x =r)) +
  geom_point(aes(color = deidentified_sample)) +
  theme_bw() + scale_color_manual(values = values)

Bivariate Count-Based Methods

In the univariate case, we consider each cell of a single cell type and center circles around each cell (reference cell). In the bivariate case, we are interested in how many cells of Type 1 (Counted) are clustered in proximity to Type 2 (Anchor). Here the circles are centered around cell of Type 2 and then the cells of Type 1 are counted.

x <- bi_ripleys_k(mif = x, mnames = c("CD3..CD8.", "CD3..FOXP3."), 
                  num_permutations = 10, method = "K",
                  edge_correction = 'translation', r = seq(0,100,10),
                  keep_perm_dis = FALSE, workers = 1, exhaustive = TRUE)

x$derived$bivariate_Count  %>%
  filter(anchor == 'CD3..FOXP3.') %>%
  ggplot(aes(x = r, y = `Degree of Clustering Permutation`)) +
  geom_line(aes(color = deidentified_sample), show.legend = TRUE) +
  theme_bw() + scale_color_manual(values = values)

The interpretation of the degree of clustering is the same here. The green line shows evidence that Tregs tend to cluster around Cytotoxic T cells for all values of $r$, while the black line indicates spatial repulsion of Tregs by Cytotoxic T cells for $\gt50$.

Univariate Nearest-Neighbor Methods

The NN_G function reports a permuted and theoretical estimate of CSR, the observed value for $G$, and the full permutation distribution of $G$ when keep_perm_dis = TRUE. The degree of clustering is computed by taking the ratio of the observed $G$ and either the permutation or theoretical estimate of CSR.

Currently, the number of permutations is 10, but this should be increased to at least 100 for a more reliable estimate of the mean.

x <- NN_G(mif = x, mnames = mnames_good, num_permutations = 10,
                edge_correction = 'rs', r = seq(0,100,10),
                keep_perm_dis = FALSE, workers = 1)

x$derived$univariate_NN  %>%
  filter(Marker != 'PDL1..Opal.540..Positive') %>%
  ggplot(aes(x = r, y = `Degree of Clustering Permutation`)) +
  geom_line(aes(color = deidentified_sample)) +
  facet_wrap(Marker~., scales = 'free') + theme_bw() + 
  scale_color_manual(values = values)

The interpretation of the degree of clustering for $G$ that values greater than 0 indicate spatial clustering of the cell types of interest, while values less than 0 indicate dispersion of these cells. For example in core TMA2_[3,B], FOXP3..Opal.620..Positive shows spatial dispersion from $0\le r \le50$ (less than zero), while CD3..Opal.570..Positive for core TMA2_[3,B] shows spatial clustering $0\le r\le75$ (greater than zero).

Bivariate Nearest-Neighbor Methods

The interpretation of the bivariate nearest neighbor distribution is similar to bivariate Ripley's $K$, in that we are measuring the degree of clustering of one cell type with respect to another. The actual value of degree of clustering is interpreted in the same manner as the univariate nearest neighbor distribution.

x <- bi_NN_G(mif = x, mnames = c("CD3..CD8.", "CD3..FOXP3."), num_permutations = 10,
               edge_correction = 'rs', r = seq(0,100,10),
               keep_perm_dis = FALSE, workers = 1, overwrite = TRUE)

x$derived$bivariate_NN  %>%
  filter(anchor == 'CD3..FOXP3.') %>%
  ggplot(aes(x = r, y = `Degree of Clustering Permutation`)) +
  geom_line(aes(color = deidentified_sample), show.legend = TRUE) +
  theme_bw() +  scale_color_manual(values = values)

Permutation Distribution

For this example, we use the same data used above for plotting cell locations. Now we only focus on the same TMA core and only focus on CD3 positive cells. In this particular TMA core, we can see that the overall cell distribution is not uniform (i.e. there are many regions where cells do not occur). Also notice, that there the CD3+ cells occur mostly in two places, thus we suspect that the degree of clustering is quite large in this TMA core.

x <- plot_immunoflo(x, plot_title = "deidentified_sample", mnames = mnames_good[1], 
                    cell_type = "Classifier.Label")

x[["derived"]][["spatial_plots"]][[4]]

Below is an example that shows how to select a single value for $r$, and how one can visualize the permutation distribution as it relates to the theoretical and observed values. Here we plot the permutation distribution for the CD3+ cells with $r = 50$. Note that the range of $r$ that is included must contain 0 and the output will not include the degree of clustering at $r = 0$, which is 0 for $K$ and $L$, and $\infty$ for $M$. Notice that the theoretical value is smaller than all 100 values from the permutation distribution.

x <- ripleys_k(mif = x, mnames = mnames_good[1], num_permutations = 100,
               edge_correction = 'translation', r = c(0,50),
               keep_perm_dis = TRUE, workers = 1, overwrite = TRUE)

#Computed the mean of the permutation distribution for a particular image
perm_mean  = x$derived$univariate_Count %>%
  filter(deidentified_sample == 'TMA3_[9,K].tif')  %>%
  summarize(`Permuted K` = mean(`Permuted K`)) %>%
  unlist()

x$derived$univariate_Count %>%
  filter(deidentified_sample == 'TMA3_[9,K].tif') %>%
  ggplot(aes(x = `Permuted K`)) + 
  geom_histogram(color = 'white', bins = 50) + 
  geom_vline(aes(xintercept = `Theoretical CSR`), color = 'red', size = 1.5) + 
  geom_vline(aes(xintercept = `Observed K`), color = 'blue', size = 1.5) + 
  geom_vline(xintercept = perm_mean, color = 'green', size = 1.5) + 
  theme_bw()

Subsetting by Compartment

Many studies may only interested in the spatial location within the tumor or stroma compartment instead of studying all cells. By using the subset_mif function which requires the column (Classifier.Label in our example) and label of cells desired (Tumor here meaning that we want to focus our analysis on only the tumor compartment).

x_tumor = subset_mif(x, classifier = 'Classifier.Label', level = 'Tumor', markers = mnames_good)

table(x$spatial[[1]]$Classifier.Label)
table(x_tumor$spatial[[1]]$Classifier.Label)

Follow-up Analysis

Once the degree of spatial clustering has be computed, questions like "Is the degree of clustering for CD3+ cells between two populations different? (t-test) or "Is the pattern of degree of clustering different between two populations?" (can be checked visually with a heatmap) may be answered.

Due to space constraints, there are only 5 TMA cores included in the example dataset. Thus, visualization is included and no statistical analysis will be conducted. With this very limited number of samples it does look like there is some evidence that CD3+CD8+, CD8+, and FOXP3+ are more clustered in patients from status B than status A, while the CD3+ cells are less clustered in patients with status B.

x <- ripleys_k(mif = x, mnames = mnames_good, num_permutations = 10,
               edge_correction = 'translation', r = c(0,50),
               keep_perm_dis = FALSE, workers = 1, overwrite = TRUE)

inner_join(x$clinical, x$derived$univariate_Count) %>%
  filter(grepl('CD3..O|CD8..O', Marker)) %>% 
  ggplot(aes(x = status, y = `Degree of Clustering Permutation`)) + 
  geom_point(aes(color = deidentified_sample), show.legend = FALSE) + 
  facet_wrap(Marker ~., scales = 'free') 
Rip_K_df = x$derived$univariate_Count %>% 
  filter(!(Marker %in% c('PDL1..Opal.540..Positive', 
                         'CD3..FOXP3.',
                         'PD1..Opal.650..Positive'))) %>%
  select(deidentified_sample, Marker, `Degree of Clustering Permutation`) %>%
  pivot_wider(names_from = deidentified_sample, values_from = `Degree of Clustering Permutation`,
              id_cols = Marker)

Rip_K_matrix = Rip_K_df %>% 
  select(-Marker) %>%
  as.matrix()

rownames(Rip_K_matrix) = Rip_K_df$Marker

annotation = inner_join(x$clinical, x$derived$univariate_Count) %>%
  select(deidentified_sample, status) %>%
  filter(!duplicated(.)) %>%
  data.frame()

rownames(annotation) = annotation$deidentified_sample

annotation = annotation %>%
  select(status)

pheatmap::pheatmap(Rip_K_matrix, treeheight_row = 0, treeheight_col = 0, 
                   cluster_cols = FALSE, annotation = annotation)

Using Custom Cell Locations

If xloc and yloc are set as NULL, when computing spatial clustering than the cell location is the center of the cell (x-coordinate is the average of XMin and XMax, and y-coordinate is the average of YMin and YMax). If you desire to use custom points, this is possible by identifying which column the new coordinates are from. Here were compute Ripley's $K$ with the minimum coordinate of x and y.

x <- ripleys_k(mif = x, mnames = mnames_good[1], num_permutations = 100,
               edge_correction = 'translation', r = c(0,50),
               keep_perm_dis = TRUE, workers = 1, overwrite = TRUE,
               xloc = 'XMin', yloc = 'YMin')


Try the spatialTIME package in your browser

Any scripts or data that you put into this service are public.

spatialTIME documentation built on Sept. 11, 2021, 9:07 a.m.