gene_x 0 like s 539 view s
Tags: plot, R, metagenomics, 16S
Plot Chao1 richness estimator, Observed OTUs, Shannon index, and Phylogenetic diversity. Regroup together samples from the same group.
library("readxl") # necessary to import the data from Excel file
library("ggplot2") # graphics
library("picante")
library("microbiome") # data analysis and visualisation
library("phyloseq") # also the basis of data object. Data analysis and visualisation
library("ggpubr") # publication quality figures, based on ggplot2
library("dplyr") # data handling, filter and reformat data frames
library("RColorBrewer") # nice color options
library("heatmaply")
library(vegan)
library(gplots)
ps.ng.tax <- readRDS("ps.ng.tax.rds")
hmp.meta <- meta(ps.ng.tax)
hmp.meta$sam_name <- rownames(hmp.meta)
hmp.div_qiime <- read.csv("adiv_even.txt", sep="\t")
colnames(hmp.div_qiime) <- c("sam_name", "chao1", "observed_otus", "shannon", "PD_whole_tree")
row.names(hmp.div_qiime) <- hmp.div_qiime$sam_name
div.df <- merge(hmp.div_qiime, hmp.meta, by = "sam_name")
div.df2 <- div.df[, c("Group", "chao1", "shannon", "observed_otus", "PD_whole_tree")]
colnames(div.df2) <- c("Group", "Chao-1", "Shannon", "OTU", "Phylogenetic Diversity")
#options(max.print=999999)
stat.test.Shannon <- compare_means(
Shannon ~ Group, data = div.df2,
method = "t.test"
)
div_df_melt <- reshape2::melt(div.df2)
p <- ggboxplot(div_df_melt, x = "Group", y = "value",
facet.by = "variable",
scales = "free",
width = 0.5,
fill = "gray", legend= "right")
p3 <- p +
stat_compare_means(
method="t.test",
comparisons = list(c("P16-P20.Foot", "P16-P20.Nose"), c("AH-XN.LH", "AH-XN.Nose"), c("AH-XN.NLH", "AH-XN.Nose")),
label = "p.signif",
symnum.args <- list(cutpoints = c(0, 0.0001, 0.001, 0.01, 0.05, 1), symbols = c("****", "***", "**", "*", "ns")),
) +
theme(axis.text.x = element_text(angle = 30, hjust = 1)) # add this line to rotate x-axis text
png("alpha_diversity.png", width=1000, height=800)
print(p3)
dev.off()
# Filter the data frame for each plot
groups_plot_1 <- c("AH-XN.LH", "AH-XN.NLH", "AH-XN.Nose")
groups_plot_2 <- setdiff(unique(div_df_melt$Group), groups_plot_1)
div_df_melt_plot_1 <- div_df_melt %>% filter(Group %in% groups_plot_1)
div_df_melt_plot_2 <- div_df_melt %>% filter(Group %in% groups_plot_2)
# Create and save Plot 1
div_df_melt_plot_1 <- div_df_melt_plot_1 %>% rename(Value = value)
p <- ggboxplot(div_df_melt_plot_1, x = "Group", y = "Value",
facet.by = "variable",
scales = "free",
width = 0.5,
fill = "gray", legend= "right")
p1 <- p +
stat_compare_means(
method="t.test",
comparisons = list(c("AH-XN.LH", "AH-XN.Nose"), c("AH-XN.NLH", "AH-XN.Nose")),
label = "p.signif",
symnum.args <- list(cutpoints = c(0, 0.0001, 0.001, 0.01, 0.05, 1), symbols = c("****", "***", "**", "*", "ns")),
) +
theme(axis.text.x = element_text(angle = 15, hjust = 1)) # add this line to rotate x-axis text
#png("alpha_diversity1.png", width=800, height=600)
print(p1)
#dev.off()
ggsave("alpha_diversity1.png", device="png", height = 8, width = 8)
#/usr/bin/convert alpha_diversity1.png -resize 800x800 alpha_diversity1_resized.png
# Create and save Plot 2
div_df_melt_plot_2 <- div_df_melt_plot_2 %>% rename(Value = value)
p <- ggboxplot(div_df_melt_plot_2, x = "Group", y = "Value",
facet.by = "variable",
scales = "free",
width = 0.5,
fill = "gray", legend= "right")
p2 <- p +
stat_compare_means(
method="t.test",
comparisons = list(c("P16-P20.Foot", "P16-P20.Nose")),
label = "p.signif",
symnum.args <- list(cutpoints = c(0, 0.0001, 0.001, 0.01, 0.05, 1), symbols = c("****", "***", "**", "*", "ns")),
) +
theme(axis.text.x = element_text(angle = 15, hjust = 1)) # add this line to rotate x-axis text
ggsave("alpha_diversity2.png", device="png", height = 8, width = 8)
#/usr/bin/convert alpha_diversity2.png -resize 800x800 alpha_diversity2_resized.png
#Using the function create_plot (Deprecated!)
#create_plot <- function(data, comparisons = NULL) {
# p <- ggboxplot(data, x = "Group", y = "Value",
# facet.by = "variable",
# scales = "free",
# width = 0.5,
# fill = "gray", legend = "right") +
# theme(axis.text.x = element_text(angle = 30, hjust = 1), strip.text = element_text(size = 15))
# #theme(axis.title.x = element_text(angle=30, hjust=1, size = 15), # Increase size of X-axis title
# # axis.title.y = element_text(size = 15), # Increase size of Y-axis title
# # strip.text = element_text(size = 15), # Increase size of Facet labels
# # axis.text.x = element_text(size = 12), # Increase size of X-axis text
# # axis.text.y = element_text(size = 12)) # Increase size of Y-axis text
#
# if (!is.null(comparisons)) {
# for (comp in comparisons) {
# if (all(comp %in% unique(data$Group))) {
# p <- p +
# stat_compare_means(
# method = "t.test",
# comparisons = list(comp),
# label = "p.signif",
# symnum.args = list(cutpoints = c(0, 0.0001, 0.001, 0.01, 0.05, 1), symbols = c("****", "***", "**", "*", "ns"))
# )
# } else {
# warning("One or more groups in the comparison are not present in the data: ", paste(comp, collapse = ", "))
# }
# }
# }
#
# return(p)
#}
#div_df_melt_plot_2 <- div_df_melt_plot_2 %>% rename(Value = value)
#p2 <- create_plot(div_df_melt_plot_2, list(c("P16-P20.Foot", "P16-P20.Nose")))
#png("alpha_diversity_plot_2.png", width=1000, height=800)
#print(p2)
#dev.off()
点赞本文的读者
还没有人对此文章表态
没有评论
QIIME + Phyloseq + MicrobiotaProcess (v1)
© 2023 XGenes.com Impressum