现在我有4个R数据框： S_GB：单性状、仅加性效应的预测准确度。 M_GB：多性状、仅加性效应的预...

下面给你一套**“可直接跑”的 R 代码，把 4 个数据框统一整理成长表，然后输出 3 张顶刊友好**的图（你可任选其一或同时发表）：

• Fig1（推荐）：每个性状在不同条件下的“性状均值准确度”的雨云/小提琴+箱线对比（避免单性状因为有 k-fold 而“样本数更大”导致偏置）
• Fig2：按时间点（ph/ear/mu + 编号）画趋势线 + 95%CI（单性状用 k-fold 计算 CI，多性状若只有 1 个值则 CI 为 0）
• Fig3：显性效应带来的增益（GBD − GB）的差值分布图（非常适合写结果段落）

假设：

• S_GB/S_GBD 列名：k_Fold Acc Trait
• M_GB/M_GBD：只有一列 ACC，行名是 Trait
• Trait 命名如 ph, ph1...ph24 等

r
## =========================
## 0) Packages
## =========================
suppressPackageStartupMessages({
  library(dplyr)
  library(tidyr)
  library(stringr)
  library(tibble)
  library(ggplot2)
})

## 如果你希望 PDF 输出更“出版级”，建议装 showtext（可选）
## install.packages("showtext")
## library(showtext)
## showtext_auto()

## 色盲友好 Okabe-Ito 调色（顶刊常用）
okabe_ito <- c(
  "Single (A)"   = "#0072B2",
  "Single (A+D)" = "#D55E00",
  "Multi (A)"    = "#009E73",
  "Multi (A+D)"  = "#CC79A7"
)

## =========================
## 1) Helper functions
## =========================
.get_acc_col <- function(df) {
  nms <- names(df)
  idx <- which(tolower(nms) %in% c("acc", "accuracy"))
  if (length(idx) == 0) stop("找不到 Acc/ACC/Accuracy 列，请检查列名。")
  nms[idx[1]]
}

prep_single <- function(df, approach = "Single", model = "A") {
  stopifnot(all(c("k_Fold", "Acc", "Trait") %in% names(df)))
  df %>%
    transmute(
      Trait = as.character(Trait),
      Fold  = as.factor(k_Fold),
      Acc   = as.numeric(Acc),
      Approach = approach,
      Model = model
    )
}

prep_multi <- function(df, approach = "Multi", model = "A") {
  acc_col <- .get_acc_col(df)
  df %>%
    rownames_to_column("Trait") %>%
    transmute(
      Trait = as.character(Trait),
      Fold  = NA,
      Acc   = as.numeric(.data[[acc_col]]),
      Approach = approach,
      Model = model
    )
}

parse_trait <- function(trait_chr) {
  base <- str_extract(trait_chr, "^[A-Za-z]+")
  tp   <- str_extract(trait_chr, "\\d+$")
  tp_i <- suppressWarnings(as.integer(tp))
  tp_i[is.na(tp_i)] <- 0L  # ph/ear/mu 视为 timepoint=0
  tibble(base_trait = base, timepoint = tp_i)
}

## =========================
## 2) Combine four dataframes into long format
## =========================
acc_long <- bind_rows(
  prep_single(S_GB,  approach = "Single", model = "A"),
  prep_single(S_GBD, approach = "Single", model = "A+D"),
  prep_multi(M_GB,   approach = "Multi",  model = "A"),
  prep_multi(M_GBD,  approach = "Multi",  model = "A+D")
) %>%
  bind_cols(parse_trait(.$Trait)) %>%
  mutate(
    Condition = case_when(
      Approach == "Single" & Model == "A"   ~ "Single (A)",
      Approach == "Single" & Model == "A+D" ~ "Single (A+D)",
      Approach == "Multi"  & Model == "A"   ~ "Multi (A)",
      Approach == "Multi"  & Model == "A+D" ~ "Multi (A+D)"
    ),
    Condition = factor(Condition, levels = names(okabe_ito)),
    base_trait = factor(base_trait, levels = c("ph","ear","mu"))
  ) %>%
  filter(!is.na(Acc))

## =========================
## 3) Trait-level summary (关键：避免 k-fold 让单性状“权重更大”)
## =========================
trait_sum <- acc_long %>%
  group_by(Condition, Approach, Model, Trait, base_trait, timepoint) %>%
  summarise(
    mean_acc = mean(Acc, na.rm = TRUE),
    sd_acc   = sd(Acc, na.rm = TRUE),
    n        = sum(!is.na(Acc)),
    .groups  = "drop"
  ) %>%
  mutate(
    se  = ifelse(is.na(sd_acc) | n <= 1, 0, sd_acc / sqrt(n)),
    ci  = ifelse(n <= 1, 0, qt(0.975, df = n - 1) * se)
  )

## =========================
## 4) Publication theme
## =========================
theme_pub <- theme_classic(base_size = 13) +
  theme(
    legend.title = element_blank(),
    legend.position = "top",
    axis.title.x = element_text(margin = margin(t = 8)),
    axis.title.y = element_text(margin = margin(r = 8)),
    strip.background = element_blank(),
    strip.text = element_text(face = "bold"),
    plot.title = element_text(face = "bold", hjust = 0),
    plot.margin = margin(8, 10, 8, 10)
  )

## =========================
## FIG 1: 推荐——“雨云/小提琴 + 箱线 + 抖动点”
## 每个性状贡献一个值（mean_acc），公平对比四条件
## =========================
p1 <- ggplot(trait_sum, aes(x = Condition, y = mean_acc, fill = Condition)) +
  geom_violin(
    width = 0.95, trim = TRUE, alpha = 0.45,
    color = "grey25", linewidth = 0.35
  ) +
  geom_boxplot(
    width = 0.18, outlier.shape = NA, alpha = 0.75,
    color = "grey15", linewidth = 0.35
  ) +
  geom_jitter(
    aes(color = Condition),
    width = 0.10, height = 0, size = 1.0, alpha = 0.35
  ) +
  facet_wrap(~ base_trait, nrow = 1, scales = "free_y") +
  scale_fill_manual(values = okabe_ito) +
  scale_color_manual(values = okabe_ito) +
  labs(
    title = "Prediction accuracy comparison across conditions",
    x = NULL,
    y = "Accuracy (trait-level mean)"
  ) +
  theme_pub +
  guides(color = "none")

## =========================
## FIG 2: 趋势图（timepoint 0..24）+ 单性状 95%CI
## 多性状如果只有一个值，ci=0，误差条不会显著
## =========================
p2 <- ggplot(trait_sum, aes(x = timepoint, y = mean_acc, color = Condition)) +
  geom_line(linewidth = 0.6) +
  geom_point(size = 1.6) +
  geom_errorbar(aes(ymin = mean_acc - ci, ymax = mean_acc + ci),
                width = 0.22, linewidth = 0.35, alpha = 0.8) +
  facet_wrap(~ base_trait, ncol = 1, scales = "free_y") +
  scale_color_manual(values = okabe_ito) +
  scale_x_continuous(breaks = sort(unique(trait_sum$timepoint))) +
  labs(
    title = "Accuracy trends across timepoints (with 95% CI for k-fold single-trait)",
    x = "Timepoint index (0 = baseline trait name without suffix)",
    y = "Accuracy (mean ± 95% CI)"
  ) +
  theme_pub

## =========================
## FIG 3: 显性效应增益（A+D − A）差值图
## 强烈建议用于结果表述：dominance improves accuracy by ...
## =========================
delta <- trait_sum %>%
  select(Approach, base_trait, timepoint, Trait, Model, mean_acc) %>%
  pivot_wider(names_from = Model, values_from = mean_acc) %>%
  mutate(delta_AD = `A+D` - `A`) %>%
  filter(!is.na(delta_AD)) %>%
  mutate(Approach = factor(Approach, levels = c("Single","Multi")))

p3 <- ggplot(delta, aes(x = Approach, y = delta_AD, fill = Approach)) +
  geom_hline(yintercept = 0, linetype = "dashed", linewidth = 0.4, color = "grey35") +
  geom_violin(width = 0.9, trim = TRUE, alpha = 0.5, color = "grey25", linewidth = 0.35) +
  geom_boxplot(width = 0.18, outlier.shape = NA, alpha = 0.75, color = "grey15", linewidth = 0.35) +
  geom_jitter(width = 0.08, height = 0, size = 1.0, alpha = 0.35) +
  facet_wrap(~ base_trait, nrow = 1, scales = "free_y") +
  scale_fill_manual(values = c("Single" = "#0072B2", "Multi" = "#009E73")) +
  labs(
    title = "Dominance gain: (A+D) − A",
    x = NULL,
    y = "Δ Accuracy"
  ) +
  theme_pub +
  guides(fill = "none")

## =========================
## 5) Print plots
## =========================
p1
p2
p3

## =========================
## 6) Save (出版级：优先 PDF；位图用 600–900 dpi)
## =========================
ggsave("Fig1_violin_box_traitmean.pdf", p1, width = 9.2, height = 3.8, device = cairo_pdf)
ggsave("Fig2_trend_timepoints.pdf",      p2, width = 8.2, height = 7.5, device = cairo_pdf)
ggsave("Fig3_dominance_gain.pdf",        p3, width = 9.2, height = 3.8, device = cairo_pdf)

ggsave("Fig1_violin_box_traitmean.png", p1, width = 9.2, height = 3.8, dpi = 800)
ggsave("Fig2_trend_timepoints.png",     p2, width = 8.2, height = 7.5, dpi = 800)
ggsave("Fig3_dominance_gain.png",       p3, width = 9.2, height = 3.8, dpi = 800)

你写论文时怎么用这些图（建议）

• Fig1：一句话总结“四种条件总体分布谁更好/更稳定”
• Fig2：回答“不同时间点/阶段的优势是否一致”（审稿人很爱问）
• Fig3：直接量化“显性效应是否提升准确度、提升多少、单/多性状是否一致”

如果你希望我把 Fig2 做成更“Nature/Science 风格”的版本（比如每个 timepoint 用 paired slopegraph 或者把三类性状拼成一个 panel 并统一 y 轴、加 A/B/C 面板编号），我也可以在你现有数据结构上再给一版更“版面级”的排版代码（patchwork/cowplot）。