visualize_v7.py

"""
visualize_v7.py — Final comparison visualization for Hatch v1–v7

4-panel figure:
  1. Full performance progression (v1–v7): Mean F1, F1 Folded, F1 Disordered
  2. CD threshold shifts (v7 vs midpoints): how much each feature moved
  3. Feature importance: global AUC vs window AUC comparison
  4. Confusion matrices: v5 (best sliding window) vs v7 (global)
"""

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
from matplotlib.gridspec import GridSpec

# ---------------------------------------------------------------------------
# Data
# ---------------------------------------------------------------------------

VERSIONS = ['v1', 'v2', 'v3', 'v4', 'v5', 'v7']
MEAN_F1  = [35.2, 46.9, 65.2, 70.0, 75.4, 89.4]
F1_FOLD  = [34.9, 31.8, 64.3, 73.9, 77.0, 90.2]
F1_DIS   = [35.5, 61.9, 66.2, 66.1, 73.8, 88.6]

FEATURE_NAMES_SHORT = [
    "Hydrophobicity",
    "Flexibility",
    "H-Bond Potential",
    "Net Charge",
    "Shannon Entropy",
    "Proline Freq",
    "Bulky Hydrophobic",
]

MIDPOINTS = [0.4353, 0.8244, 1.1851, 0.0505, 3.7639, 0.0499, 0.2607]
CD_OPT    = [0.4030, 0.8345, 1.1277, 0.0465, 3.8419, 0.0532, 0.2607]
SHIFTS    = [cd - mp for cd, mp in zip(CD_OPT, MIDPOINTS)]

# Window AUC (from feature_importance_scan.py)
WINDOW_AUC = [0.764, 0.778, 0.598, 0.599, 0.785, 0.617, 0.838]

# v5 confusion matrix (from grid_search_v5 best result, 400-seq test set)
# Approximate from 75.4% mean F1, ~200 per class
CM_V5 = np.array([[148, 52], [46, 154]])   # [[TN, FP], [FN, TP]]

# v7 confusion matrix (from global_cd_optimizer.py, 5680-seq test set)
CM_V7 = np.array([[2341, 339], [433, 2567]])

# ---------------------------------------------------------------------------
# Plot
# ---------------------------------------------------------------------------

fig = plt.figure(figsize=(16, 12))
fig.patch.set_facecolor('#0d1117')
gs = GridSpec(2, 2, figure=fig, hspace=0.38, wspace=0.32)

DARK_BG  = '#0d1117'
CARD_BG  = '#161b22'
GRID_COL = '#21262d'
TEXT_COL = '#e6edf3'
MUTED    = '#8b949e'

C_MEAN   = '#58a6ff'
C_FOLD   = '#3fb950'
C_DIS    = '#f78166'
C_V7     = '#d2a8ff'

def style_ax(ax, title):
    ax.set_facecolor(CARD_BG)
    for spine in ax.spines.values():
        spine.set_color(GRID_COL)
    ax.tick_params(colors=MUTED, labelsize=9)
    ax.xaxis.label.set_color(MUTED)
    ax.yaxis.label.set_color(MUTED)
    ax.set_title(title, color=TEXT_COL, fontsize=11, fontweight='bold', pad=10)
    ax.grid(axis='y', color=GRID_COL, linewidth=0.5, alpha=0.7)

# ── Panel 1: Full progression ──────────────────────────────────────────────
ax1 = fig.add_subplot(gs[0, 0])
style_ax(ax1, "Performance Progression: v1 → v7")

x = np.arange(len(VERSIONS))
w = 0.26
ax1.bar(x - w, MEAN_F1, width=w, color=C_MEAN, alpha=0.85, label='Mean F1')
ax1.bar(x,     F1_FOLD, width=w, color=C_FOLD, alpha=0.85, label='F1 Folded')
ax1.bar(x + w, F1_DIS,  width=w, color=C_DIS,  alpha=0.85, label='F1 Disordered')

ax1.set_xticks(x)
ax1.set_xticklabels(VERSIONS, color=TEXT_COL)
ax1.set_ylim(0, 100)
ax1.set_ylabel("F1 Score (%)", color=MUTED)

# Annotate v7 bars
for i, (mf, ff, df) in enumerate(zip(MEAN_F1, F1_FOLD, F1_DIS)):
    if VERSIONS[i] == 'v7':
        for val, offset in [(mf, -w), (ff, 0), (df, w)]:
            ax1.text(i + offset, val + 1.5, f'{val:.1f}', ha='center',
                     va='bottom', fontsize=7.5, color=TEXT_COL, fontweight='bold')

# Annotate v5 mean F1
ax1.text(4 - w, MEAN_F1[4] + 1.5, f'{MEAN_F1[4]:.1f}', ha='center',
         va='bottom', fontsize=7.5, color=C_MEAN)

# Architecture change arrow
ax1.annotate('', xy=(4.6, 88), xytext=(4.1, 78),
             arrowprops=dict(arrowstyle='->', color=C_V7, lw=1.5))
ax1.text(4.65, 87, 'Global\nArchitecture', color=C_V7, fontsize=7.5, va='top')

ax1.legend(loc='upper left', framealpha=0.3, facecolor=CARD_BG,
           labelcolor=TEXT_COL, fontsize=8)

# ── Panel 2: CD threshold shifts ──────────────────────────────────────────
ax2 = fig.add_subplot(gs[0, 1])
style_ax(ax2, "Coordinate Descent Threshold Shifts (v7)")

colors = [C_DIS if s < 0 else C_FOLD if s > 0 else MUTED for s in SHIFTS]
y_pos = np.arange(len(FEATURE_NAMES_SHORT))
bars = ax2.barh(y_pos, SHIFTS, color=colors, alpha=0.85, height=0.6)
ax2.set_yticks(y_pos)
ax2.set_yticklabels(FEATURE_NAMES_SHORT, color=TEXT_COL, fontsize=9)
ax2.axvline(0, color=MUTED, linewidth=0.8, linestyle='--')
ax2.set_xlabel("Threshold Shift (CD − Midpoint)", color=MUTED)
ax2.grid(axis='x', color=GRID_COL, linewidth=0.5, alpha=0.7)
ax2.grid(axis='y', visible=False)

for i, (s, feat) in enumerate(zip(SHIFTS, FEATURE_NAMES_SHORT)):
    if abs(s) > 0.001:
        ax2.text(s + (0.002 if s > 0 else -0.002), i,
                 f'{s:+.4f}', va='center',
                 ha='left' if s > 0 else 'right',
                 fontsize=8, color=TEXT_COL)

up_patch   = mpatches.Patch(color=C_FOLD, alpha=0.85, label='Threshold ↑ (tighter for folded)')
down_patch = mpatches.Patch(color=C_DIS,  alpha=0.85, label='Threshold ↓ (tighter for folded)')
ax2.legend(handles=[up_patch, down_patch], loc='lower right',
           framealpha=0.3, facecolor=CARD_BG, labelcolor=TEXT_COL, fontsize=8)

# ── Panel 3: Window AUC (feature importance) ──────────────────────────────
ax3 = fig.add_subplot(gs[1, 0])
style_ax(ax3, "Feature Discriminative Power (Window AUC, W=30)")

y_pos = np.arange(len(FEATURE_NAMES_SHORT))
auc_colors = [C_FOLD if a >= 0.76 else C_DIS for a in WINDOW_AUC]
ax3.barh(y_pos, WINDOW_AUC, color=auc_colors, alpha=0.85, height=0.6)
ax3.set_yticks(y_pos)
ax3.set_yticklabels(FEATURE_NAMES_SHORT, color=TEXT_COL, fontsize=9)
ax3.set_xlim(0.5, 0.95)
ax3.axvline(0.76, color=MUTED, linewidth=0.8, linestyle='--', alpha=0.6)
ax3.text(0.762, 6.6, 'Tier 1\nboundary', color=MUTED, fontsize=7.5)
ax3.set_xlabel("AUC (794,870 training windows)", color=MUTED)
ax3.grid(axis='x', color=GRID_COL, linewidth=0.5, alpha=0.7)
ax3.grid(axis='y', visible=False)

for i, a in enumerate(WINDOW_AUC):
    ax3.text(a + 0.003, i, f'{a:.3f}', va='center', fontsize=8.5, color=TEXT_COL)

t1_patch = mpatches.Patch(color=C_FOLD, alpha=0.85, label='Tier 1 (AUC ≥ 0.76)')
t2_patch = mpatches.Patch(color=C_DIS,  alpha=0.85, label='Tier 2 (AUC < 0.76)')
ax3.legend(handles=[t1_patch, t2_patch], loc='lower right',
           framealpha=0.3, facecolor=CARD_BG, labelcolor=TEXT_COL, fontsize=8)

# ── Panel 4: Confusion matrices ────────────────────────────────────────────
ax4 = fig.add_subplot(gs[1, 1])
ax4.set_facecolor(CARD_BG)
for spine in ax4.spines.values():
    spine.set_color(GRID_COL)
ax4.set_title("Confusion Matrices: v5 (Sliding Window) vs v7 (Global)", 
              color=TEXT_COL, fontsize=11, fontweight='bold', pad=10)
ax4.axis('off')

def draw_cm(ax, cm, title, x_offset, y_offset, scale=1.0):
    total = cm.sum()
    labels = [['TN', 'FP'], ['FN', 'TP']]
    colors_cm = [['#1f6feb', '#da3633'], ['#da3633', '#238636']]
    cell_size = 0.18 * scale
    for r in range(2):
        for c in range(2):
            val = cm[r, c]
            pct = val / total * 100
            x = x_offset + c * cell_size
            y = y_offset - r * cell_size
            rect = plt.Rectangle((x, y - cell_size), cell_size, cell_size,
                                  facecolor=colors_cm[r][c], alpha=0.4,
                                  transform=ax.transAxes)
            ax.add_patch(rect)
            ax.text(x + cell_size/2, y - cell_size/2,
                    f'{labels[r][c]}\n{val:,}\n({pct:.1f}%)',
                    ha='center', va='center', fontsize=7.5 * scale,
                    color=TEXT_COL, fontweight='bold',
                    transform=ax.transAxes)
    ax.text(x_offset + cell_size, y_offset + 0.04,
            title, ha='center', va='bottom', fontsize=9 * scale,
            color=TEXT_COL, fontweight='bold', transform=ax.transAxes)
    # Axis labels
    for c, lbl in enumerate(['Pred\nDisordered', 'Pred\nFolded']):
        ax.text(x_offset + c * cell_size + cell_size/2, y_offset - 2*cell_size - 0.02,
                lbl, ha='center', va='top', fontsize=7 * scale, color=MUTED,
                transform=ax.transAxes)
    for r, lbl in enumerate(['Actual\nDisordered', 'Actual\nFolded']):
        ax.text(x_offset - 0.03, y_offset - r * cell_size - cell_size/2,
                lbl, ha='right', va='center', fontsize=7 * scale, color=MUTED,
                transform=ax.transAxes)

draw_cm(ax4, CM_V5, f'v5 (Sliding Window)\n75.4% Mean F1', 0.08, 0.82, scale=0.9)
draw_cm(ax4, CM_V7, f'v7 (Global Classifier)\n89.4% Mean F1', 0.57, 0.82, scale=0.9)

# ── Title & footer ─────────────────────────────────────────────────────────
fig.suptitle(
    "Hatch: Biophysical Protein Disorder Classifier  ·  v1 → v7 Experiment Summary",
    color=TEXT_COL, fontsize=14, fontweight='bold', y=0.98
)
fig.text(0.5, 0.01,
         "v7: Global Feature Classifier + Coordinate Descent  |  "
         "89.47% Accuracy  |  89.41% Mean F1  |  5,680-sequence test set",
         ha='center', color=MUTED, fontsize=9)

plt.savefig("hatch_v7_final.png", dpi=150, bbox_inches='tight',
            facecolor=DARK_BG, edgecolor='none')
print("Saved: hatch_v7_final.png")
plt.close()