In wvictor14/plmec: Placental DNA methylation analysis tools

Installation

if(!requireNamespace("BiocManager", quietly = TRUE))
  install.packages("BiocManager")
BiocManager::install("planet")

Cell composition

To infer cell composition on placental villi DNAm samples, we can need to use placental reference cpgs (Yuan 2021). These are provided in this package as plCellCpGsThird and plCellCpGsFirst for third trimester (term) and first trimester samples, respectively.

In this example we are using term villi DNAm data, so we first load the reference cpgs plCellCpGsThird. This is a data frame of 600 cpgs, with mean methylation levels for each cell type.

# cell deconvolution packages
library(minfi)
library(EpiDISH)

# data wrangling and plotting
library(dplyr)
library(ggplot2)
library(tidyr)
library(planet)

# load example data
data("plBetas")
data("plCellCpGsThird")
head(plCellCpGsThird)

After our reference cpg data is loaded, we can estimate cell composition by applying either the Constrained Projection approach implemented by the R packages minfi or EpiDISH, or a non-constrained approach by EpiDish. I demonstrate how to do both.

Minfi

houseman_estimates <- minfi:::projectCellType(
  plBetas[rownames(plCellCpGsThird), ],
  plCellCpGsThird,
  lessThanOne = FALSE
)

head(houseman_estimates)

EpiDISH

# robust partial correlations
epidish_RPC <- epidish(
  beta.m = plBetas[rownames(plCellCpGsThird), ],
  ref.m = plCellCpGsThird,
  method = "RPC"
)

# CIBERSORT
epidish_CBS <- epidish(
  beta.m = plBetas[rownames(plCellCpGsThird), ],
  ref.m = plCellCpGsThird,
  method = "CBS"
)

# constrained projection (houseman 2012)
epidish_CP <- epidish(
  beta.m = plBetas[rownames(plCellCpGsThird), ],
  ref.m = plCellCpGsThird,
  method = "CP"
)

Compare

Below, I demonstrate how we can visually compare the different cell composition estimates.

data("plColors")

# bind estimate data frames and reshape for plotting
bind_rows(
  houseman_estimates %>% as.data.frame() %>% mutate(algorithm = "CP (Houseman)"),
  epidish_RPC$estF %>% as.data.frame() %>% mutate(algorithm = "RPC"),
  epidish_CBS$estF %>% as.data.frame() %>% mutate(algorithm = "CBS"),
  epidish_CP$estF %>% as.data.frame() %>% mutate(algorithm = "CP (EpiDISH)")
) %>%
  mutate(sample = rep(rownames(houseman_estimates), 4)) %>%
  as_tibble() %>%
  pivot_longer(
    cols = -c(algorithm, sample),
    names_to = "component",
    values_to = "estimate"
  ) %>%

  # relevel for plot
  mutate(component = factor(component,
                            levels = c(
                              "nRBC", "Endothelial", "Hofbauer",
                              "Stromal", "Trophoblasts",
                              "Syncytiotrophoblast"
                            )
  )) %>%

  # plot
  ggplot(aes(x = sample, y = estimate, fill = component)) +
  geom_bar(stat = "identity") +
  facet_wrap(~algorithm, ncol = 1) +
  scale_fill_manual(values = plColors) +
  scale_y_continuous(
    limits = c(-0.1, 1.1), breaks = c(0, 0.5, 1),
    labels = scales::percent
  ) +
  theme(axis.text.x = element_text(angle = 90, vjust = 0.5)) +
  coord_cartesian(ylim = c(0, 1)) +
  labs(x = "", fill = "")

Some notes:

Normalize your data with minfi::preprocessNoob and BMIQ
Use all cell CpGs - if some are missing, estimates may vary
If your samples have been processed in a particular manner, (e.g. sampling from maternal side) expect cell composition to reflect that

Gestational age

Example Data

For demonstration, I use 24 samples from a placental DNAm dataset from GEO, (GSE7519), which contains samples collected in an Australian population. The DNA methylation data (in betas) can be accessed with data(plBetas) and corresponding sample information from data(plPhenoData). Note that for demonstration purposes, the cpgs have been filtered to a random \~10,000 CpGs, plus the CpGs used in all of the functions from this package.

# load example data
data(plBetas)
data(plPhenoData)

dim(plBetas)
#> [1] 13918    24
head(plPhenoData)
#> # A tibble: 6 x 7
#>   sample_id  sex   disease   gestation_wk ga_RPC ga_CPC ga_RRPC
#>   <fct>      <chr> <chr>            <dbl>  <dbl>  <dbl>   <dbl>
#> 1 GSM1944936 Male  preeclam~           36   38.5   38.7    38.7
#> 2 GSM1944939 Male  preeclam~           32   33.1   34.2    32.6
#> 3 GSM1944942 Fema~ preeclam~           32   34.3   35.1    33.3
#> 4 GSM1944944 Male  preeclam~           35   35.5   36.7    35.5
#> 5 GSM1944946 Fema~ preeclam~           38   37.6   37.6    36.6
#> 6 GSM1944948 Fema~ preeclam~           36   36.8   38.4    36.7

There are 3 gestational age clocks for placental DNA methylation data from (Lee 2019):

Robust Placental Clock (RPC)
Control Placental Clock (CPC)
Refined Robust Placental Clock (RRPC)

To predict gestational, we load the example data:

plBetas - DNAm data for 24 placental samples
plPhenoData - Matching sample information

Predict Gestational Age

To select the type of clock, we can specify the type argument in predictAge.

We will apply all three clocks on this data, and add the predicted age to the sample information data.frame, plPhenoData.

plPhenoData <- plPhenoData %>%
  mutate(
    ga_RPC = predictAge(plBetas, type = "RPC"),
    ga_CPC = predictAge(plBetas, type = "CPC"),
    ga_RRPC = predictAge(plBetas, type = "RRPC")
  )

Note that the number of predictors (CpGs) that were used in our data are printed. It's important to take note if a significant number of predictive CpGs are missing in your data, as this can affect the predicted gestational age accuracy.

Next, I plot the difference between predicted and reported gestational age, for each of the 3 gestational age predictors.

plPhenoData %>%

  # reshape, to plot
  pivot_longer(
    cols = contains("ga"),
    names_to = "clock_type",
    names_prefix = "ga_",
    values_to = "ga"
  ) %>%

  # plot code
  ggplot(aes(x = gestation_wk, y = ga, col = disease)) +
  geom_point() +
  geom_smooth(method = "lm", se = FALSE) +
  facet_wrap(~clock_type) +
  theme(legend.position = "top") +
  labs(x = "Reported GA (weeks)", y = "Inferred GA (weeks)", col = "")

Ethnicity

Before predicting ethnicity You can ensure that you have all features using the ethnicityCpGs vector:

data(ethnicityCpGs)
all(ethnicityCpGs %in% rownames(plBetas))

results <- predictEthnicity(plBetas)
results %>%
  tail(8)

predictEthnicity returns probabilities corresponding to each ethnicity for each sample (e.g Prob_Caucasian, Prob_African, Prob_Asian). This applies a glmnet model described in (Yuan 2019). A final classification is determined in two ways:

Predicted_ethnicity_nothresh - returns a classification corresponding to the highest class-specific probability.
Predicted_ethnicity - if the highest class-specific probability is below 0.75, then the the sample is assigned an Amibiguous label. This threshold can be adjusted with the threshold argument. Samples with this label might require special attention in downstream analyses.

results %>%
  ggplot(aes(
    x = Prob_Caucasian, y = Prob_African,
    col = Predicted_ethnicity
  )) +
  geom_point(alpha = 0.7) +
  coord_cartesian(xlim = c(0, 1), ylim = c(0, 1)) +
  scale_x_continuous(labels = scales::percent) +
  scale_y_continuous(labels = scales::percent) +
  labs(x = "P(Caucasian)", y = "P(African)")

results %>%
  ggplot(aes(
    x = Prob_Caucasian, y = Prob_Asian,
    col = Predicted_ethnicity
  )) +
  geom_point(alpha = 0.7) +
  coord_cartesian(xlim = c(0, 1), ylim = c(0, 1)) +
  scale_x_continuous(labels = scales::percent) +
  scale_y_continuous(labels = scales::percent) +
  labs(x = "P(Caucasian)", y = "P(Asian)")

We can't compare this to self-reported ethnicity as it is unavailable. But we know these samples were collected in Sydney, Australia, and are therefore likely mostly European with some East Asian participants.

table(results$Predicted_ethnicity)

A note on adjustment in differential methylation analysis

Because 'Ambiguous' samples might have different mixtures of ancestries, it might be inadequate to adjust for them as one group in an analysis of admixed populations (e.g. 50/50 Asian/African should not be considered the same group as 50/50 Caucasian/African). One solution would be to simply remove these samples. Another would be to adjust for the raw probabilities-in this case, use only two of the three probabilities, since the third will be redundant (probabilities sum to 1). If sample numbers are large enough in each group, stratifying downstream analyses by ethnicity might also be a valid option.

Early-onset preeclampsia (EOPE)

To calculate the probability of EOPE of placental DNAm chorionic villi samples, we rely on 45 predictive CpGs.

In this example, we load the validation data used in Fernández-Boyano 2023] and estimate the EOPE probability of the samples. Note that the function must be run on a matrix with the full set of CpG probes in either the 450K or 850K arrays - the reason for this is that all 45 predictive CpGs must be present for prediction to be completed.

It is recommended that data is normalized using BMIQ prior to prediction.

Note that samples must be rows and CpGs must be columns. The default threshold for classification used to assign labels is 55%; if the users wishes to use other threshold, different labels can be assigned based on the output probabilities.

library(ExperimentHub)
eh <- ExperimentHub()
query(eh, "eoPredData")

# download BMIQ normalized 450k data
x_test <- eh[['EH8403']]
preds <- x_test %>% predictPreeclampsia()

Inspect the results:

head(preds)

# join with metadata
valMeta <- eh[['EH8404']]
valMeta <- left_join(valMeta, preds, by="Sample_ID")

# visualize results
plot_predictions <- function(df, predicted_class_column) {
  df %>% 
    ggplot() +
    geom_boxplot(
      aes(x = Class, y = EOPE),
      width = 0.3,
      alpha = 0.5,
      color = "darkgrey"
    ) +
    geom_jitter(
      aes(x = Class, y = EOPE, color = {{predicted_class_column}}), 
      alpha = 0.75, 
      position = position_jitter(width = .1)
    ) + 
    coord_flip() +
    ylab("Class Probability of EOPE") +
    xlab("True Class") +
    ylim(0,1) +
    scale_color_manual(
      name = "Predicted Class", 
      values = c("#E69F00", "skyblue", "#999999")
    ) + 
    theme_minimal() 
}

valMeta %>% plot_predictions(PE_Status)

if user wishes to use different threshold from 55% probability, as an example 70%

valMeta %>% mutate(
  Pred_Class = case_when(
    EOPE > 0.7 ~ "EOPE",
    `Non-PE Preterm` > 0.7 ~ "Non-PE Preterm",
    .default = 'low-confidence'
  )) %>% plot_predictions(Pred_Class)