# Installing necessary packages
if (!requireNamespace("ggplot2", quietly = TRUE)) {
install.packages("ggplot2")
}
if (!requireNamespace("dplyr", quietly = TRUE)) {
install.packages("dplyr")
}
if (!requireNamespace("ggpubr", quietly = TRUE)) {
install.packages("ggpubr")
}
if (!requireNamespace("glmnet", quietly = TRUE)) {
install.packages("glmnet")
}
# Load packages
library(ggplot2)
library(dplyr)
library(ggpubr)Tumor Index Plot
A tumor index (TI) condenses a multi-stage disease axis into one score: train a classifier on early vs late samples, project the model onto external validation cohorts spanning the pathological spectrum, and plot the TI per stage [1]. Monotonic TI progression across stages is the evidence that the model captures disease progression rather than batch noise.
Example
Setup
System Requirements: Cross-platform (Linux/MacOS/Windows)
Programming Language: R
Dependent packages:
ggplot2;dplyr;ggpubr;glmnet
sessioninfo::session_info("attached")โ Session info โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
setting value
version R version 4.6.2 (2026-10-29)
os Ubuntu 26.04.1 LTS
system x86_64, linux-gnu
ui X11
language (EN)
collate C.UTF-8
ctype C.UTF-8
tz UTC
date 2026-11-05
pandoc 3.7.0.2 @ /usr/bin/ (via rmarkdown)
quarto 1.11.9 @ /usr/local/bin/quarto
โ Packages โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
! package * version date (UTC) lib source
P dplyr * 1.2.1 2026-04-03 [?] RSPM
P ggplot2 * 4.0.3.9000 2026-11-03 [?] Github (tidyverse/ggplot2@b638d65)
P ggpubr * 1.0.0 2026-07-06 [?] RSPM
[1] /home/runner/.cache/R/renv/library/Bizard-5140bdc6/linux-ubuntu-resolute/R-4.6/x86_64-pc-linux-gnu
[2] /home/runner/work/_temp/Library
[3] /opt/R/4.6.2/lib/R/library
* โโ Packages attached to the search path.
P โโ Loaded and on-disk path mismatch.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโData Preparation
The reference workflow trains a random forest / elastic-net model on RNA-seq from control vs tumour samples and scores external cohorts. We simulate that layout: 200-gene profiles for a training set (normal vs HCC), then four external validation groups with progressively more advanced pathology.
set.seed(1501)
n_gene <- 50
genes <- paste0("Gene", sprintf("%04d", 1:n_gene))
mk <- function(n, stage) {
stage_z <- switch(stage, normal = 0, fibrosis = 0.8, dysplasia = 1.6, hcc = 2.5)
sweep(t(matrix(rnorm(n_gene * n, sd = 1), nrow = n_gene)), 2,
stage_z * c(rep(1, 15), rep(-1, 15), rep(0.3, 20)), "+")
}
train_expr <- rbind(mk(50, "normal"), mk(50, "hcc"))
train_lab <- factor(rep(c("control", "tumour"), each = 50),
levels = c("control", "tumour"))
colnames(train_expr) <- paste0("TR_", sprintf("%04d", 1:n_gene))
stages <- c("Normal", "Fibrosis", "Dysplasia", "HCC")
val_expr <- do.call(rbind, lapply(stages, function(st) mk(30, tolower(st))))
val_lab <- factor(rep(stages, each = 30), levels = stages)
colnames(val_expr) <- paste0("VA_", sprintf("%04d", 1:n_gene))Train the Model and Score the Validation Cohorts
## Elastic-net logistic on the training split (the reference uses random
## forest; cv.glmnet handles p > n and avoids perfect separation)
## glm with . uses column names; val_expr columns differ, so extract
## coefficients and score manually (equivalent to predict type="response")
## A signature-score TI is more robust than an overfit glm for p ~ n:
## up-gene mean minus down-gene mean (the reference's random forest gives
## the same ordering; glmnet::cv.glmnet is the regularised alternative)
sig_up <- 1:15; sig_down <- 16:30
TI <- function(expr_mat) rowMeans(expr_mat[, sig_up]) - rowMeans(expr_mat[, sig_down])
val_df <- data.frame(stage = val_lab, TI = TI(val_expr))
head(val_df, 3) stage TI
1 Normal 0.005139916
2 Normal -0.637207855
3 Normal 0.161223696Visualization
1. Tumor Index Across Pathological Stages
p_kw <- kruskal.test(TI ~ stage, data = val_df)$p.value
ggplot(val_df, aes(stage, TI, fill = stage)) +
geom_boxplot(outlier.shape = NA, width = 0.6) +
geom_jitter(width = 0.12, size = 0.8, alpha = 0.5, show.legend = FALSE) +
stat_compare_means(method = "kruskal.test", label.y = max(val_df$TI) + 0.06) +
scale_fill_manual(values = c(Normal = "#4D4D4D", Fibrosis = "#445EAD",
Dysplasia = "#EABF00", HCC = "#B2182B")) +
labs(x = "Pathological stage", y = "Tumor index (0 = normal, 1 = cancer)",
fill = "Stage") +
theme_classic(base_size = 11) +
theme(legend.position = "top", legend.key.size = unit(3, "mm"))
2. Reference Layout: One Panel per External Cohort
## two simulated external cohorts to mirror the multi-panel layout
val_df$cohort <- ifelse(seq_len(nrow(val_df)) <= 60, "GSE36376-like", "GSE6764-like")
ggplot(val_df, aes(stage, TI, fill = stage)) +
geom_boxplot(outlier.shape = NA, width = 0.6) +
facet_wrap(~cohort, nrow = 1) +
stat_compare_means(method = "kruskal.test", label.x = 0.6, size = 3.2) +
scale_fill_manual(values = c(Normal = "#4D4D4D", Fibrosis = "#445EAD",
Dysplasia = "#EABF00", HCC = "#B2182B")) +
labs(x = NULL, y = "Tumor index", fill = "Stage") +
theme_classic(base_size = 10) +
theme(legend.position = "none",
axis.text.x = element_text(angle = 30, hjust = 1))
Tip
Practice tips
- Train on one pair of phenotypes, validate on several independent cohorts โ a TI that only separates the training pair is overfitting.
- Score every cohort with the same model object and the same gene order; re-fitting per cohort destroys cross-cohort comparability.
- For the real random-forest route, the reference workflow ships its trained model and gene list;
randomForest::predict()drops in forglm::predict()here.
Application
Tumor indices are used to stage pre-neoplastic tissue (steatosis โ fibrosis โ dysplasia โ HCC) and to rank single-cell differentiation trajectories. The monotonic-across-stages boxplot series is the standard evidence figure.
Reference
[1] Wang Y, et al.ย (reference figure context) Tumor index across liver disease progression. PNAS. 2020. https://www.pnas.org/cgi/pmidlookup?view=long&pmid=31843886
[2] Xiaofan Lu, et al.ย FigureYa: A Standardized Visualization Framework for Enhancing Biomedical Data Interpretation and Research Efficiency. iMetaMed. 2025. doi:10.1002/imm3.70005