scripts/08_visualization/make_paper_figures.py

#!/usr/bin/env python3
import argparse
import json
from pathlib import Path

import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns


def load_json(path: Path):
    if not path.exists():
        return None
    with open(path, "r", encoding="utf-8") as f:
        return json.load(f)


def savefig(path: Path):
    path.parent.mkdir(parents=True, exist_ok=True)
    plt.tight_layout()
    plt.savefig(path, dpi=260, bbox_inches="tight")
    plt.close()


def plot_generator_comparison(rows_csv: Path, output_dir: Path):
    df = pd.read_csv(rows_csv)
    methods = [m for m in ["reference", "ar_unconditional", "ar_conditioned", "diffusion_conditioned"] if m in set(df["method"])]
    df = df[df["method"].isin(methods)].copy()

    plt.figure(figsize=(8.5, 4.2))
    sns.violinplot(data=df, x="method", y="gc_content", order=methods, inner="quartile", cut=0)
    plt.xticks(rotation=18, ha="right")
    plt.title("GC Content Distribution")
    savefig(output_dir / "fig_generation_gc_content.png")

    plt.figure(figsize=(8.5, 4.2))
    sns.boxplot(data=df, x="method", y="max_homopolymer", order=methods, showfliers=False)
    plt.xticks(rotation=18, ha="right")
    plt.title("Homopolymer Length Distribution")
    savefig(output_dir / "fig_generation_homopolymer.png")

    plt.figure(figsize=(8.5, 4.2))
    sns.boxplot(data=df, x="method", y="nearest_reference_hamming", order=methods, showfliers=False)
    plt.xticks(rotation=18, ha="right")
    plt.title("Nearest Reference Hamming Distance")
    savefig(output_dir / "fig_nearest_reference_distance.png")

    scored = df[df["prediction_sum"].notna()].copy()
    if not scored.empty:
        plt.figure(figsize=(8.5, 4.2))
        sns.violinplot(data=scored, x="method", y="prediction_sum", order=[m for m in methods if m in set(scored["method"])], inner="quartile", cut=0)
        plt.xticks(rotation=18, ha="right")
        plt.title("Predicted Enhancer Activity by Method")
        savefig(output_dir / "fig_predicted_activity_by_method.png")

    two_dim = df[df["prediction_label_0"].notna() & df["prediction_label_1"].notna()].copy()
    if not two_dim.empty:
        plt.figure(figsize=(6.2, 5.2))
        sns.scatterplot(
            data=two_dim,
            x="prediction_label_0",
            y="prediction_label_1",
            hue="method",
            style="activity_bucket" if "activity_bucket" in two_dim else None,
            s=18,
            alpha=0.65,
        )
        plt.title("Predicted Activity Space")
        savefig(output_dir / "fig_activity_2d_scatter.png")

    pll = df[df["diffusion_pll"].notna()].copy()
    if not pll.empty:
        plt.figure(figsize=(7.2, 4.2))
        sns.violinplot(data=pll, x="source", y="diffusion_pll", hue="activity_bucket", inner="quartile", cut=0)
        plt.title("Diffusion PLL by Source and Bucket")
        savefig(output_dir / "fig_diffusion_pll.png")


def plot_summary_table(summary_json: Path, output_dir: Path):
    summary = load_json(summary_json)
    if not summary:
        return
    rows = []
    for method, values in summary.get("methods", {}).items():
        row = {"method": method}
        for key in [
            "valid_dna_rate",
            "unique_rate",
            "mean_gc_content",
            "mean_max_homopolymer",
            "mean_nearest_reference_hamming",
            "kmer3_js_to_reference",
            "kmer4_js_to_reference",
            "mean_prediction_sum",
            "mean_diffusion_pll",
        ]:
            if key in values:
                row[key] = values[key]
        rows.append(row)
    if not rows:
        return
    table = pd.DataFrame(rows)
    table.to_csv(output_dir / "table_generator_comparison.csv", index=False)

    plot_cols = [c for c in ["valid_dna_rate", "unique_rate", "kmer4_js_to_reference", "mean_nearest_reference_hamming"] if c in table]
    long = table.melt(id_vars="method", value_vars=plot_cols, var_name="metric", value_name="value")
    plt.figure(figsize=(9.2, 4.4))
    sns.barplot(data=long, x="metric", y="value", hue="method")
    plt.xticks(rotation=18, ha="right")
    plt.title("Generator Quality Summary")
    savefig(output_dir / "fig_generator_quality_summary.png")


def plot_motif(motif_csv: Path, output_dir: Path):
    if not motif_csv.exists():
        return
    df = pd.read_csv(motif_csv)
    if df.empty:
        return
    top = (
        df.groupby("motif_name")["hit_rate"]
        .max()
        .sort_values(ascending=False)
        .head(20)
        .index
    )
    sub = df[df["motif_name"].isin(top)]
    pivot = sub.pivot_table(index="motif_name", columns="method", values="hit_rate", fill_value=0)
    plt.figure(figsize=(8.5, 6.5))
    sns.heatmap(pivot, cmap="viridis")
    plt.title("Top Motif Hit Rates")
    savefig(output_dir / "fig_motif_hit_rate_heatmap.png")


def main():
    parser = argparse.ArgumentParser(description="Build paper-ready figures from experiment outputs.")
    parser.add_argument("--result_root", required=True)
    parser.add_argument("--output_dir", required=True)
    args = parser.parse_args()

    result_root = Path(args.result_root)
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    sns.set_theme(style="whitegrid")

    plot_generator_comparison(result_root / "sequence_metrics" / "sequence_metrics_rows.csv", output_dir)
    plot_summary_table(result_root / "sequence_metrics" / "sequence_metrics_summary.json", output_dir)
    plot_motif(result_root / "motif_analysis" / "motif_scan_summary.csv", output_dir)
    print(output_dir)


if __name__ == "__main__":
    main()