@nahisaho/satori 0.1.0

package/src/.github/skills/scientific-meta-analysis/SKILL.md

@@ -0,0 +1,352 @@
+---
+name: scientific-meta-analysis
+description: |
+  メタ解析スキル。固定効果・ランダム効果モデル（DerSimonian-Laird）、Forest プロット、
+  異質性評価（I²/Q 検定/τ²）、出版バイアス検出（Funnel プロット/Egger/Begg 検定）、
+  サブグループ解析、メタ回帰、累積メタ解析のテンプレートを提供。
+---
+
+# Scientific Meta-Analysis
+
+複数の独立した研究結果を統合し、全体的なエビデンスを定量化するためのスキル。
+効果量（SMD / OR / RR / MD）の統合、異質性評価、出版バイアス検出のパイプラインを
+提供する。
+
+## When to Use
+
+- 複数の研究・実験結果を統合的に評価するとき
+- 効果量（Hedges' g / Cohen's d / SMD）を算出するとき
+- Forest プロット / Funnel プロットを描画するとき
+- 研究間の異質性を定量化するとき（I² / Q / τ²）
+- 出版バイアスの有無を検定するとき
+
+---
+
+## Quick Start
+
+## 1. 効果量の算出
+
+```python
+import numpy as np
+import pandas as pd
+from scipy.stats import norm
+
+def compute_effect_sizes(studies_df, effect_type="SMD"):
+    """
+    各研究の効果量と分散を算出する。
+
+    effect_type:
+        "SMD"  — Standardized Mean Difference (Hedges' g)
+        "MD"   — Mean Difference (同スケール)
+        "OR"   — Odds Ratio (log 変換)
+        "RR"   — Risk Ratio (log 変換)
+
+    Input columns (SMD/MD):
+        mean1, sd1, n1, mean2, sd2, n2
+
+    Input columns (OR/RR):
+        events1, total1, events2, total2
+    """
+    df = studies_df.copy()
+
+    if effect_type == "SMD":
+        # Cohen's d → Hedges' g (小標本補正)
+        pooled_sd = np.sqrt(
+            ((df["n1"]-1)*df["sd1"]**2 + (df["n2"]-1)*df["sd2"]**2) /
+            (df["n1"] + df["n2"] - 2)
+        )
+        d = (df["mean1"] - df["mean2"]) / pooled_sd
+        # Hedges' correction factor
+        J = 1 - 3 / (4*(df["n1"]+df["n2"]-2) - 1)
+        df["effect_size"] = d * J
+        df["variance"] = (df["n1"]+df["n2"])/(df["n1"]*df["n2"]) + \
+                          df["effect_size"]**2 / (2*(df["n1"]+df["n2"]))
+
+    elif effect_type == "MD":
+        df["effect_size"] = df["mean1"] - df["mean2"]
+        df["variance"] = df["sd1"]**2/df["n1"] + df["sd2"]**2/df["n2"]
+
+    elif effect_type == "OR":
+        a = df["events1"]; b = df["total1"] - df["events1"]
+        c = df["events2"]; d_val = df["total2"] - df["events2"]
+        df["effect_size"] = np.log((a * d_val) / (b * c + 1e-10) + 1e-10)
+        df["variance"] = 1/a + 1/b + 1/c + 1/d_val
+
+    elif effect_type == "RR":
+        p1 = df["events1"] / df["total1"]
+        p2 = df["events2"] / df["total2"]
+        df["effect_size"] = np.log(p1 / (p2 + 1e-10) + 1e-10)
+        df["variance"] = (1-p1)/(df["events1"]+1e-10) + \
+                          (1-p2)/(df["events2"]+1e-10)
+
+    df["se"] = np.sqrt(df["variance"])
+    df["ci_lower"] = df["effect_size"] - 1.96 * df["se"]
+    df["ci_upper"] = df["effect_size"] + 1.96 * df["se"]
+    df["weight"] = 1 / df["variance"]
+
+    return df
+```
+
+## 2. 固定効果 / ランダム効果モデル
+
+```python
+def meta_analysis(studies_df, model="random"):
+    """
+    メタ解析統合。
+
+    model:
+        "fixed"   — 固定効果モデル (Inverse-Variance weighted)
+        "random"  — ランダム効果モデル (DerSimonian-Laird)
+
+    Input: DataFrame with columns: study, effect_size, variance
+    """
+    es = studies_df["effect_size"].values
+    var = studies_df["variance"].values
+    w = 1 / var
+    k = len(es)
+
+    # 固定効果
+    theta_fixed = np.sum(w * es) / np.sum(w)
+    se_fixed = 1 / np.sqrt(np.sum(w))
+
+    # 異質性
+    Q = np.sum(w * (es - theta_fixed)**2)
+    df_Q = k - 1
+    p_Q = 1 - __import__("scipy").stats.chi2.cdf(Q, df_Q)
+    I2 = max(0, (Q - df_Q) / Q * 100) if Q > 0 else 0
+
+    if model == "random":
+        # DerSimonian-Laird τ² 推定
+        C = np.sum(w) - np.sum(w**2) / np.sum(w)
+        tau2 = max(0, (Q - df_Q) / C)
+
+        # ランダム効果重み
+        w_re = 1 / (var + tau2)
+        theta_random = np.sum(w_re * es) / np.sum(w_re)
+        se_random = 1 / np.sqrt(np.sum(w_re))
+
+        summary_effect = theta_random
+        summary_se = se_random
+    else:
+        tau2 = 0
+        summary_effect = theta_fixed
+        summary_se = se_fixed
+
+    z = summary_effect / summary_se
+    p_val = 2 * (1 - norm.cdf(abs(z)))
+
+    return {
+        "model": model,
+        "summary_effect": summary_effect,
+        "se": summary_se,
+        "ci_lower": summary_effect - 1.96 * summary_se,
+        "ci_upper": summary_effect + 1.96 * summary_se,
+        "z_value": z,
+        "p_value": p_val,
+        "Q_statistic": Q,
+        "Q_p_value": p_Q,
+        "I_squared": I2,
+        "tau_squared": tau2,
+        "k_studies": k,
+    }
+```
+
+## 3. Forest プロット
+
+```python
+import matplotlib.pyplot as plt
+
+def forest_plot(studies_df, meta_result, effect_label="SMD",
+                figsize=(10, None)):
+    """
+    Forest プロットを描画する。
+
+    studies_df: study, effect_size, ci_lower, ci_upper, weight
+    meta_result: meta_analysis() の出力
+    """
+    k = len(studies_df)
+    if figsize[1] is None:
+        figsize = (figsize[0], max(4, k * 0.4 + 2))
+
+    fig, ax = plt.subplots(figsize=figsize)
+
+    y_positions = range(k, 0, -1)
+
+    # 個別研究
+    for i, (_, row) in enumerate(studies_df.iterrows()):
+        y = list(y_positions)[i]
+        ax.plot([row["ci_lower"], row["ci_upper"]], [y, y],
+               "b-", linewidth=1.5)
+        size = row.get("weight", 1) / studies_df["weight"].max() * 200 + 20
+        ax.plot(row["effect_size"], y, "bs", markersize=np.sqrt(size),
+               markerfacecolor="steelblue")
+        ax.text(-0.05, y, row.get("study", f"Study {i+1}"),
+               ha="right", va="center", fontsize=9,
+               transform=ax.get_yaxis_transform())
+
+    # サマリーダイヤモンド
+    y_summary = 0
+    diamond_x = [meta_result["ci_lower"], meta_result["summary_effect"],
+                 meta_result["ci_upper"], meta_result["summary_effect"]]
+    diamond_y = [y_summary, y_summary + 0.3, y_summary, y_summary - 0.3]
+    ax.fill(diamond_x, diamond_y, color="red", alpha=0.7)
+    ax.text(-0.05, y_summary, "Summary",
+           ha="right", va="center", fontsize=9, fontweight="bold",
+           transform=ax.get_yaxis_transform())
+
+    # 参照線
+    ax.axvline(0, color="black", linestyle="-", linewidth=0.5)
+
+    ax.set_xlabel(effect_label)
+    ax.set_yticks([])
+    ax.set_title(f"Forest Plot (I²={meta_result['I_squared']:.1f}%, "
+                f"p={meta_result['p_value']:.4f})", fontweight="bold")
+
+    # 右側に数値
+    for i, (_, row) in enumerate(studies_df.iterrows()):
+        y = list(y_positions)[i]
+        ax.text(1.02, y, f"{row['effect_size']:.2f} [{row['ci_lower']:.2f}, "
+               f"{row['ci_upper']:.2f}]",
+               ha="left", va="center", fontsize=8,
+               transform=ax.get_yaxis_transform())
+
+    plt.tight_layout()
+    plt.savefig("figures/forest_plot.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+## 4. 出版バイアス検出
+
+### 4.1 Funnel プロット
+
+```python
+def funnel_plot(studies_df, meta_result, figsize=(8, 6)):
+    """
+    Funnel プロットを描画する。
+    非対称なら出版バイアスの存在を示唆。
+    """
+    fig, ax = plt.subplots(figsize=figsize)
+
+    ax.scatter(studies_df["effect_size"], studies_df["se"],
+              c="steelblue", s=50, edgecolors="black", zorder=5)
+
+    # 参照線
+    ax.axvline(meta_result["summary_effect"], color="red", linestyle="--")
+
+    # 95% 擬似信頼区間
+    se_range = np.linspace(0, studies_df["se"].max() * 1.1, 100)
+    ax.plot(meta_result["summary_effect"] - 1.96 * se_range, se_range,
+           "gray", linestyle="--", alpha=0.5)
+    ax.plot(meta_result["summary_effect"] + 1.96 * se_range, se_range,
+           "gray", linestyle="--", alpha=0.5)
+
+    ax.set_xlabel("Effect Size")
+    ax.set_ylabel("Standard Error")
+    ax.set_title("Funnel Plot", fontweight="bold")
+    ax.invert_yaxis()
+    plt.tight_layout()
+    plt.savefig("figures/funnel_plot.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+### 4.2 Egger 検定
+
+```python
+import statsmodels.api as sm
+
+def egger_test(studies_df):
+    """
+    Egger 回帰検定 — Funnel プロットの非対称性を統計的に検定。
+
+    y = effect_size / se
+    x = 1 / se
+    切片 ≠ 0 → 出版バイアスあり
+    """
+    precision = 1 / studies_df["se"]
+    z_score = studies_df["effect_size"] / studies_df["se"]
+
+    X = sm.add_constant(precision)
+    model = sm.OLS(z_score, X).fit()
+
+    return {
+        "intercept": model.params["const"],
+        "intercept_se": model.bse["const"],
+        "intercept_p": model.pvalues["const"],
+        "publication_bias": model.pvalues["const"] < 0.05,
+    }
+```
+
+## 5. サブグループ解析
+
+```python
+def subgroup_analysis(studies_df, subgroup_col, model="random"):
+    """サブグループごとにメタ解析を行い、グループ間差を検定する。"""
+    subgroups = studies_df[subgroup_col].unique()
+    results = []
+
+    for sg in subgroups:
+        subset = studies_df[studies_df[subgroup_col] == sg]
+        if len(subset) >= 2:
+            ma = meta_analysis(subset, model=model)
+            ma["subgroup"] = sg
+            ma["k"] = len(subset)
+            results.append(ma)
+
+    results_df = pd.DataFrame(results)
+
+    # グループ間検定 (Q_between)
+    overall = meta_analysis(studies_df, model=model)
+    Q_within = sum(r["Q_statistic"] for r in results)
+    Q_between = overall["Q_statistic"] - Q_within
+    df_between = len(results) - 1
+    p_between = 1 - __import__("scipy").stats.chi2.cdf(Q_between, df_between)
+
+    return {
+        "subgroup_results": results_df,
+        "Q_between": Q_between,
+        "df_between": df_between,
+        "p_between": p_between,
+    }
+```
+
+## 6. 累積メタ解析
+
+```python
+def cumulative_meta_analysis(studies_df, sort_by="year", model="random"):
+    """
+    研究を順に追加しながらメタ解析を実行する。
+    エビデンスの蓄積過程を可視化。
+    """
+    sorted_df = studies_df.sort_values(sort_by)
+    cumulative_results = []
+
+    for i in range(2, len(sorted_df) + 1):
+        subset = sorted_df.iloc[:i]
+        ma = meta_analysis(subset, model=model)
+        ma["n_studies"] = i
+        ma["last_added"] = sorted_df.iloc[i-1].get("study", f"Study {i}")
+        cumulative_results.append(ma)
+
+    return pd.DataFrame(cumulative_results)
+```
+
+## References
+
+### Output Files
+
+| ファイル | 形式 |
+|---|---|
+| `results/meta_analysis_summary.csv` | CSV |
+| `results/effect_sizes.csv` | CSV |
+| `results/publication_bias_tests.csv` | CSV |
+| `results/subgroup_analysis.csv` | CSV |
+| `figures/forest_plot.png` | PNG |
+| `figures/funnel_plot.png` | PNG |
+| `figures/cumulative_meta.png` | PNG |
+
+#### 依存パッケージ
+
+```
+scipy>=1.10
+statsmodels>=0.14
+```

package/src/.github/skills/scientific-metabolomics/SKILL.md

@@ -0,0 +1,326 @@
+---
+name: scientific-metabolomics
+description: |
+  メタボロミクス解析スキル。Pareto スケーリング、PLS-DA + VIP スコア、置換検定（Q²）、
+  代謝パスウェイ濃縮解析（Fisher exact test）、代謝物相関ネットワーク、
+  Volcano プロット/箱ひげ図による差次代謝物同定パイプライン。
+  Scientific Skills Exp-07 で確立したパターン。
+---
+
+# Scientific Metabolomics Analysis
+
+LC-MS / GC-MS / NMR ベースのメタボロミクスデータを対象に、品質管理→前処理→
+単変量解析→多変量解析→パスウェイ解析の標準パイプラインを提供する。
+メタボロミクス固有の統計手法（PLS-DA、VIP スコア）に特化している。
+
+## When to Use
+
+- メタボロミクスデータの統計解析パイプラインが必要なとき
+- PLS-DA による群間判別＋VIP スコアによるバイオマーカー候補を算出するとき
+- 代謝パスウェイ濃縮解析が必要なとき
+- Pareto スケーリングや代謝物相関ネットワークが必要なとき
+
+---
+
+## Quick Start
+
+## 1. メタボロミクス前処理
+
+```python
+import numpy as np
+import pandas as pd
+from sklearn.impute import KNNImputer
+
+def metabolomics_preprocessing(df, sample_col="Sample_ID", group_col="Group",
+                                 min_detect_pct=0.5):
+    """
+    メタボロミクス標準前処理パイプライン。
+    1. 低検出率代謝物の除去
+    2. KNN 欠損値補完
+    3. log2 変換
+    4. Pareto スケーリング
+    """
+    metabolite_cols = [c for c in df.columns if c not in [sample_col, group_col]]
+
+    # Step 1: 低検出率フィルタリング
+    detect_rate = df[metabolite_cols].notna().mean()
+    keep = detect_rate[detect_rate >= min_detect_pct].index.tolist()
+    removed = len(metabolite_cols) - len(keep)
+    print(f"  Removed {removed} metabolites with <{min_detect_pct*100:.0f}% detection rate")
+    metabolite_cols = keep
+
+    # Step 2: KNN 補完
+    imputer = KNNImputer(n_neighbors=5)
+    df[metabolite_cols] = imputer.fit_transform(df[metabolite_cols])
+
+    # Step 3: log2 変換
+    df[metabolite_cols] = np.log2(df[metabolite_cols].clip(lower=1e-10) + 1)
+
+    # Step 4: Pareto スケーリング
+    for col in metabolite_cols:
+        mean = df[col].mean()
+        std = df[col].std()
+        df[col] = (df[col] - mean) / np.sqrt(std + 1e-10)
+
+    return df, metabolite_cols
+```
+
+## 2. 単変量解析 — 差次代謝物同定
+
+```python
+from scipy.stats import mannwhitneyu, ttest_ind
+from statsmodels.stats.multitest import multipletests
+
+def univariate_analysis(df, metabolite_cols, group_col, group1, group2,
+                         test="mannwhitneyu", correction="fdr_bh"):
+    """
+    2 群間の差次代謝物を同定する。
+
+    Returns:
+        DataFrame with columns: metabolite, log2FC, pvalue, padj, significant
+    """
+    g1 = df[df[group_col] == group1]
+    g2 = df[df[group_col] == group2]
+
+    results = []
+    for met in metabolite_cols:
+        v1 = g1[met].dropna()
+        v2 = g2[met].dropna()
+
+        if test == "mannwhitneyu":
+            stat, pval = mannwhitneyu(v1, v2, alternative="two-sided")
+        else:
+            stat, pval = ttest_ind(v1, v2)
+
+        log2fc = v2.mean() - v1.mean()
+        results.append({"metabolite": met, "log2FC": log2fc, "pvalue": pval})
+
+    results_df = pd.DataFrame(results)
+
+    # 多重検定補正
+    reject, padj, _, _ = multipletests(results_df["pvalue"], method=correction)
+    results_df["padj"] = padj
+    results_df["significant"] = reject
+    results_df["neg_log10p"] = -np.log10(results_df["pvalue"] + 1e-300)
+
+    results_df = results_df.sort_values("pvalue")
+    return results_df
+```
+
+## 3. PLS-DA + VIP スコア
+
+```python
+from sklearn.cross_decomposition import PLSRegression
+from sklearn.preprocessing import LabelEncoder
+
+def plsda_analysis(X, y, n_components=2):
+    """
+    PLS-DA を実行し、VIP スコアを算出する。
+
+    VIP (Variable Importance in Projection):
+        VIP_j = sqrt(p * Σ(q²_a * w²_ja) / Σ(q²_a))
+        VIP > 1 の変数がバイオマーカー候補
+
+    Parameters:
+        X: 代謝物データ行列 (n_samples, n_metabolites)
+        y: グループラベル
+
+    Returns:
+        pls_model, scores, vip_scores
+    """
+    le = LabelEncoder()
+    y_encoded = le.fit_transform(y).astype(float)
+
+    pls = PLSRegression(n_components=n_components, scale=True)
+    pls.fit(X, y_encoded)
+
+    # スコア（潜在変数）
+    scores = pls.transform(X)
+
+    # VIP スコア算出
+    T = pls.x_scores_       # (n, n_comp)
+    W = pls.x_weights_      # (p, n_comp)
+    Q = pls.y_loadings_     # (1, n_comp)
+
+    p = X.shape[1]
+    vip = np.zeros(p)
+
+    ss_total = np.sum(Q**2 * np.sum(T**2, axis=0))
+    for j in range(p):
+        ss_j = np.sum(Q**2 * np.sum(T**2, axis=0) * W[j, :]**2)
+        vip[j] = np.sqrt(p * ss_j / ss_total)
+
+    return pls, scores, vip
+
+
+def plot_plsda_scores(scores, y, group_names=None, figsize=(8, 6)):
+    """PLS-DA スコアプロットを描画する。"""
+    import matplotlib.pyplot as plt
+
+    fig, ax = plt.subplots(figsize=figsize)
+    unique = np.unique(y)
+    colors = plt.cm.Set1(np.linspace(0, 0.5, len(unique)))
+
+    for color, group in zip(colors, unique):
+        mask = y == group
+        label = group_names[group] if group_names else str(group)
+        ax.scatter(scores[mask, 0], scores[mask, 1],
+                  c=[color], label=label, s=60, alpha=0.7, edgecolors="black")
+        # 95% 信頼楕円
+        from matplotlib.patches import Ellipse
+        cov = np.cov(scores[mask, 0], scores[mask, 1])
+        vals, vecs = np.linalg.eigh(cov)
+        angle = np.degrees(np.arctan2(vecs[1, 1], vecs[0, 1]))
+        w, h = 2 * np.sqrt(vals * 5.991)  # chi2(2, 0.95) = 5.991
+        ell = Ellipse(xy=(scores[mask, 0].mean(), scores[mask, 1].mean()),
+                     width=w, height=h, angle=angle,
+                     fill=False, color=color, linewidth=2, linestyle="--")
+        ax.add_patch(ell)
+
+    ax.set_xlabel("PLS Component 1")
+    ax.set_ylabel("PLS Component 2")
+    ax.set_title("PLS-DA Score Plot", fontweight="bold")
+    ax.legend()
+    plt.tight_layout()
+    plt.savefig("figures/plsda_scores.png", dpi=300, bbox_inches="tight")
+    plt.close()
+```
+
+## 4. 置換検定（PLS-DA バリデーション）
+
+```python
+def permutation_test_plsda(X, y, n_components=2, n_permutations=100):
+    """
+    PLS-DA モデルの置換検定。
+    Q² と R²Y の分布を生成し、真のモデルの有意性を評価する。
+    """
+    from sklearn.model_selection import cross_val_predict
+    le = LabelEncoder()
+    y_enc = le.fit_transform(y).astype(float)
+
+    # 真のモデル
+    pls_true = PLSRegression(n_components=n_components, scale=True)
+    y_pred = cross_val_predict(pls_true, X, y_enc, cv=5)
+    ss_res = np.sum((y_enc - y_pred.ravel())**2)
+    ss_tot = np.sum((y_enc - y_enc.mean())**2)
+    q2_true = 1 - ss_res / ss_tot
+
+    # 置換
+    q2_perm = []
+    for _ in range(n_permutations):
+        y_perm = np.random.permutation(y_enc)
+        pls_p = PLSRegression(n_components=n_components, scale=True)
+        y_pred_p = cross_val_predict(pls_p, X, y_perm, cv=5)
+        ss_res_p = np.sum((y_perm - y_pred_p.ravel())**2)
+        ss_tot_p = np.sum((y_perm - y_perm.mean())**2)
+        q2_perm.append(1 - ss_res_p / ss_tot_p)
+
+    p_value = np.mean(np.array(q2_perm) >= q2_true)
+
+    return {
+        "Q2_true": q2_true,
+        "Q2_perm_mean": np.mean(q2_perm),
+        "Q2_perm_std": np.std(q2_perm),
+        "p_value": p_value,
+        "significant": p_value < 0.05,
+    }
+```
+
+## 5. 代謝パスウェイ濃縮解析
+
+```python
+from scipy.stats import fisher_exact
+
+def pathway_enrichment(significant_metabolites, pathway_annotations,
+                        metabolite_col="Metabolite", pathway_col="Pathway",
+                        total_metabolites=None):
+    """
+    Fisher 正確検定による代謝パスウェイ濃縮解析。
+
+    Parameters:
+        significant_metabolites: list of significant metabolite names
+        pathway_annotations: DataFrame with metabolite-pathway mapping
+        total_metabolites: 解析対象の全代謝物数
+    """
+    sig_set = set(significant_metabolites)
+    all_annotated = set(pathway_annotations[metabolite_col])
+    if total_metabolites is None:
+        total_metabolites = len(all_annotated)
+
+    pathways = pathway_annotations[pathway_col].unique()
+    results = []
+
+    for pw in pathways:
+        pw_members = set(
+            pathway_annotations[pathway_annotations[pathway_col] == pw][metabolite_col]
+        )
+        k = len(sig_set & pw_members)       # hit
+        K = len(pw_members)                  # pathway size
+        n = len(sig_set)                     # significant
+        N = total_metabolites                # total
+
+        # 2x2 分割表
+        table = [[k, K - k], [n - k, N - K - n + k]]
+        odds_ratio, p_value = fisher_exact(table, alternative="greater")
+
+        results.append({
+            "Pathway": pw,
+            "Hits": k,
+            "Pathway_Size": K,
+            "Significant_Total": n,
+            "Odds_Ratio": odds_ratio,
+            "p_value": p_value,
+        })
+
+    results_df = pd.DataFrame(results).sort_values("p_value")
+    _, padj, _, _ = multipletests(results_df["p_value"], method="fdr_bh")
+    results_df["padj"] = padj
+
+    results_df.to_csv("results/pathway_enrichment.csv", index=False)
+    return results_df
+```
+
+## 6. 代謝物相関ネットワーク
+
+```python
+def metabolite_correlation_network(df, metabolite_cols, method="spearman",
+                                     threshold=0.7):
+    """
+    代謝物間の相関からネットワークを構築する。
+
+    Parameters:
+        threshold: |r| ≥ threshold のペアのみエッジとして採用
+    """
+    import networkx as nx
+
+    corr = df[metabolite_cols].corr(method=method)
+
+    G = nx.Graph()
+    for i, met_i in enumerate(metabolite_cols):
+        G.add_node(met_i)
+        for j, met_j in enumerate(metabolite_cols):
+            if i < j:
+                r = corr.iloc[i, j]
+                if abs(r) >= threshold:
+                    G.add_edge(met_i, met_j, weight=abs(r),
+                              sign="positive" if r > 0 else "negative")
+
+    return G, corr
+```
+
+## References
+
+### Output Files
+
+| ファイル | 形式 |
+|---|---|
+| `results/univariate_results.csv` | CSV |
+| `results/vip_scores.csv` | CSV |
+| `results/pathway_enrichment.csv` | CSV |
+| `figures/plsda_scores.png` | PNG |
+| `figures/vip_barplot.png` | PNG |
+| `figures/metabolite_network.png` | PNG |
+
+#### 参照実験
+
+- **Exp-07**: PLS-DA + VIP、Pareto スケーリング、パスウェイ濃縮、相関ネットワーク