基于随机森林算法的2型糖尿病肠道菌群属水平特征筛选与分析

# R代码：通过curatedMetagenomicData获取数据
library(curatedMetagenomicData)

# 获取Qin 2012 T2D队列的属水平相对丰度
qin_data <- curatedMetagenomicData(
  "QinJ_2012.relative_abundance", dryrun = FALSE
)

# 提取样本元数据和丰度矩阵
se <- qin_data[[1]]
metadata <- colData(se) # 344行
abundance <- assay(se) # 物种 x 样本 矩阵

# Python代码：log1p变换
import numpy as np

# 原始丰度矩阵 X，shape = (344, ~114)
X_transformed = np.log1p(X) # 等价于 ln(X + 1)

# 对比效果
# 原始：[0, 0.001, 0.05, 15.3, 42.7]
# log1p：[0, 0.001, 0.049, 2.79, 3.78] # 极端值被压缩

# Python代码：嵌套CV核心结构
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import StratifiedKFold, RandomizedSearchCV

outer_cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
inner_cv = StratifiedKFold(n_splits=3, shuffle=True, random_state=42)

param_dist = {
  'n_estimators': [200, 300, 500, 800],
  'max_depth': [None, 5, 8, 12],
  'min_samples_split': [2, 4, 8],
  'min_samples_leaf': [1, 2, 4],
  'max_features': ['sqrt', 'log2', 0.5]
}

# 外层循环
for train_idx, test_idx in outer_cv.split(X, y):
  # 内层：随机搜索调参
  search = RandomizedSearchCV(
    RandomForestClassifier(random_state=42),
    param_dist, n_iter=12, cv=inner_cv,
    scoring='roc_auc', random_state=42
  )
  search.fit(X[train_idx], y[train_idx])
  # 外层：用最佳模型在测试集评估
  oof_preds[test_idx] = search.predict_proba(X[test_idx])[:, 1]