🔬 The Scientific Defensibility Dashboard¶

Duration: ~30 minutes | Level: ⭐⭐⭐ Advanced

What makes a protein structure scientifically defensible?¶

This question sits at the heart of every structural biology publication. When researchers deposit a structure in the Protein Data Bank (PDB), it must pass rigorous validation checks before acceptance. The same standards apply to computationally generated structures.

synth-pdb's get_quality_report() applies five independent criteria, each rooted in a different branch of physical chemistry:

#	Criterion	Scientific standard	Threshold
1	Bond Z-score	Engh & Huber (1991) geometry dictionary	mean < 3.0
2	Rotamer quality	Dunbrack backbone-dependent library (1997)	favored > 80%
3	Ramachandran	MolProbity / Top8000 dataset	outliers < 5%
4	Potential energy	OpenMM / AMBER14 force field	< 1 × 10⁵ kJ/mol
5	Hydrophobic burial	Kauzmann Oil Drop model (1959)	burial_ratio ≥ 0.8

A structure must pass all five to be is_overall_scientifically_defensible = True.

In this tutorial we generate three structures, each designed to fail a different subset of criteria, demonstrating that the checks are genuinely independent.

In [ ]:

Copied!





import os
import sys

IN_COLAB = 'google.colab' in sys.modules

if IN_COLAB:
    # @title Environment Setup
    !pip install -q synth-pdb biotite matplotlib numpy scipy py3Dmol openmm
else:
    sys.path.append(os.path.abspath('../../'))

print("✅ Environment configured!")
import os
import sys

IN_COLAB = 'google.colab' in sys.modules

if IN_COLAB:
    # @title Environment Setup
    !pip install -q synth-pdb biotite matplotlib numpy scipy py3Dmol openmm
else:
    sys.path.append(os.path.abspath('../../'))

print("✅ Environment configured!")

🔧 Setup¶

In [ ]:

Copied!





import os
import sys

try:
    import google.colab  # noqa: F401
    IN_COLAB = True
except ImportError:
    IN_COLAB = False

if IN_COLAB:
    print('Running in Google Colab — installing dependencies...')
    import subprocess
    subprocess.run(['pip', 'install', '-q', 'synth-pdb', 'biotite'], check=True)
else:
    repo_root = os.path.abspath(os.path.join(os.path.dirname('__file__'), '..', '..'))
    if repo_root not in sys.path:
        sys.path.insert(0, repo_root)
    print(f'Running locally — {repo_root}')

import matplotlib.pyplot as plt
import numpy as np

from synth_pdb.generator import generate_pdb_content
from synth_pdb.validator import PDBValidator

print('All imports successful.')
import os
import sys

try:
    import google.colab  # noqa: F401
    IN_COLAB = True
except ImportError:
    IN_COLAB = False

if IN_COLAB:
    print('Running in Google Colab — installing dependencies...')
    import subprocess
    subprocess.run(['pip', 'install', '-q', 'synth-pdb', 'biotite'], check=True)
else:
    repo_root = os.path.abspath(os.path.join(os.path.dirname('__file__'), '..', '..'))
    if repo_root not in sys.path:
        sys.path.insert(0, repo_root)
    print(f'Running locally — {repo_root}')

import matplotlib.pyplot as plt
import numpy as np

from synth_pdb.generator import generate_pdb_content
from synth_pdb.validator import PDBValidator

print('All imports successful.')

📐 Step 1: The Five Criteria — Scientific Background¶

Criterion 1 — Bond Z-score (Engh & Huber, 1991)¶

Rainer Engh and Robert Huber analysed 985 high-resolution X-ray structures to establish ideal bond lengths and angles. The Z-score measures how many standard deviations a bond length deviates from this ideal:

$$Z = \frac{\text{observed} - \text{ideal}}{\sigma_{\text{ideal}}}$$

A mean Z-score < 3.0 means no bonds are dramatically wrong.

Criterion 2 — Rotamer Quality (Dunbrack, 1997)¶

Side-chain conformations are not random. Roland Dunbrack built a backbone-dependent rotamer library from thousands of PDB structures — a probability map of which Chi angles are physically preferred. Structures with > 80% favored rotamers are adopting low-energy, experimentally observed conformations.

Criterion 3 — Ramachandran Outliers (MolProbity)¶

The Ramachandran plot (phi/psi dihedral space) has forbidden regions where backbone atoms sterically clash. MolProbity's Top8000 dataset defines outlier boundaries from ultra-high-resolution structures. < 5% outliers is the wwPDB deposition standard.

Criterion 4 — Potential Energy (AMBER14 / OpenMM)¶

The AMBER14 force field computes potential energy as a sum of bonded and non-bonded terms. Negative energies are typical for well-packed structures. Extremely positive energies (> 1 × 10⁵ kJ/mol) indicate severe steric clashes.

Criterion 5 — Hydrophobic Burial (Kauzmann, 1959)¶

The Oil Drop model predicts that hydrophobic residues bury in the core (low SASA) while polar residues remain surface-exposed (high SASA). burial_ratio ≥ 0.8 confirms this pattern. See the companion tutorial The Oil Drop Model for a deep dive.

🧬 Step 2: Three Structures, Three Failure Modes¶

We design each structure to have a specific defect:

Structure	Design intent	Expected failure
S1 — Minimized helix	Amphipathic helix, energy minimized	None — should PASS all 5
S2 — Drifted decoy	Same sequence, torsion noise `drift=120`	Ramachandran (backbone scrambled)
S3 — Hydrophobic strand	All-hydrophobic beta strand, unminimized	Burial ratio + energy

In [ ]:

Copied!





print('Generating 3 structures (seed=42 for reproducibility)...')

# S1: energy-minimized amphipathic alpha helix
pdb_s1 = generate_pdb_content(
    sequence_str='AAKLLLAAKLLLAAK',
    structure='1-15:alpha',
    minimize_energy=True,
    seed=42,
)

# S2: same sequence, large torsion perturbation scrambles Ramachandran
pdb_s2 = generate_pdb_content(
    sequence_str='AAKLLLAAKLLLAAK',
    structure='1-15:alpha',
    minimize_energy=False,
    drift=120.0,
    seed=42,
)

# S3: all-hydrophobic beta strand — no polar surface, no minimization
pdb_s3 = generate_pdb_content(
    sequence_str='VIVVIVVIVVI',
    structure='1-11:beta',
    minimize_energy=False,
    seed=42,
)

print('S1 — Minimized helix:      ready')
print('S2 — Drifted decoy:        ready')
print('S3 — Hydrophobic strand:   ready')
print('Generating 3 structures (seed=42 for reproducibility)...')

# S1: energy-minimized amphipathic alpha helix
pdb_s1 = generate_pdb_content(
    sequence_str='AAKLLLAAKLLLAAK',
    structure='1-15:alpha',
    minimize_energy=True,
    seed=42,
)

# S2: same sequence, large torsion perturbation scrambles Ramachandran
pdb_s2 = generate_pdb_content(
    sequence_str='AAKLLLAAKLLLAAK',
    structure='1-15:alpha',
    minimize_energy=False,
    drift=120.0,
    seed=42,
)

# S3: all-hydrophobic beta strand — no polar surface, no minimization
pdb_s3 = generate_pdb_content(
    sequence_str='VIVVIVVIVVI',
    structure='1-11:beta',
    minimize_energy=False,
    seed=42,
)

print('S1 — Minimized helix:      ready')
print('S2 — Drifted decoy:        ready')
print('S3 — Hydrophobic strand:   ready')

📊 Step 3: Run `get_quality_report()` on Each Structure¶

One call, five criteria, a single composite verdict:

In [ ]:

Copied!





structures = [
    ('S1  Minimized helix',    pdb_s1),
    ('S2  Drifted decoy',      pdb_s2),
    ('S3  Hydrophobic strand', pdb_s3),
]

reports = []
for label, pdb in structures:
    r = PDBValidator(pdb).get_quality_report()
    reports.append((label, r))
    verdict = 'DEFENSIBLE' if r['is_overall_scientifically_defensible'] else 'NOT defensible'
    print(f'{label}: {verdict}')
    print(f'  Bond Z-score  : {r["geometric_z_scores"]["mean_bond_zscore"]:.3f}')
    print(f'  Rotamers %%   : {r["rotamer_stats"]["favored_rotamers_pct"]:.1f}')
    print(f'  Rama outliers%%: {r["ramachandran_stats"]["outlier_pct"]:.1f}')
    print(f'  Energy kJ/mol : {r["potential_energy_kj_mol"]:.0f}')
    print(f'  Burial ratio  : {r["hydrophobic_burial_ratio"]:.3f}')
    print()
structures = [
    ('S1  Minimized helix',    pdb_s1),
    ('S2  Drifted decoy',      pdb_s2),
    ('S3  Hydrophobic strand', pdb_s3),
]

reports = []
for label, pdb in structures:
    r = PDBValidator(pdb).get_quality_report()
    reports.append((label, r))
    verdict = 'DEFENSIBLE' if r['is_overall_scientifically_defensible'] else 'NOT defensible'
    print(f'{label}: {verdict}')
    print(f'  Bond Z-score  : {r["geometric_z_scores"]["mean_bond_zscore"]:.3f}')
    print(f'  Rotamers %%   : {r["rotamer_stats"]["favored_rotamers_pct"]:.1f}')
    print(f'  Rama outliers%%: {r["ramachandran_stats"]["outlier_pct"]:.1f}')
    print(f'  Energy kJ/mol : {r["potential_energy_kj_mol"]:.0f}')
    print(f'  Burial ratio  : {r["hydrophobic_burial_ratio"]:.3f}')
    print()

🎨 Step 4: The Dashboard Heatmap¶

Each cell shows the raw value and a green (pass) / red (fail) background. Reading down each column shows which criterion fails, and reading across each row shows which structures fail.

Key insight: S2 fails only Ramachandran; S3 fails only burial and energy. The criteria are truly independent — a structure can have perfect geometry but wrong chemistry, or correct secondary structure but bad energetics.

In [ ]:

Copied!





# ── Criterion definitions ────────────────────────────────────────────────
CRITERIA = [
    # (display_name, report_key_fn, threshold_fn, format_str, invert_color)
    # invert_color=True means LOWER is better
    ('Bond Z-score\n(< 3.0)',
     lambda r: r['geometric_z_scores']['mean_bond_zscore'],
     lambda v: v < 3.0, '.3f', False),
    ('Rotamers\n(> 80 %)',
     lambda r: r['rotamer_stats']['favored_rotamers_pct'],
     lambda v: v > 80.0, '.0f', False),
    ('Ramachandran\noutliers (< 5 %)',
     lambda r: r['ramachandran_stats']['outlier_pct'],
     lambda v: v < 5.0, '.0f', True),
    ('Potential energy\n(< 1e5 kJ/mol)',
     lambda r: r['potential_energy_kj_mol'],
     lambda v: v < 1e5, '.0f', True),
    ('Burial ratio\n(>= 0.8)',
     lambda r: r['hydrophobic_burial_ratio'],
     lambda v: v >= 0.8, '.3f', False),
]

struct_labels = [label for label, _ in reports]
n_structs     = len(reports)
n_criteria    = len(CRITERIA)

fig, ax = plt.subplots(figsize=(11, 4.5))
ax.set_xlim(-0.5, n_criteria - 0.5)
ax.set_ylim(-0.5, n_structs - 0.5)
ax.invert_yaxis()

PASS_COLOR = '#d5f5e3'   # soft green
FAIL_COLOR = '#fadbd8'   # soft red
PASS_EDGE  = '#1e8449'
FAIL_EDGE  = '#922b21'

for col, (_crit_name, value_fn, pass_fn, fmt, _) in enumerate(CRITERIA):
    for row, (struct_label, report) in enumerate(reports):
        val     = value_fn(report)
        passes  = pass_fn(val)
        bg      = PASS_COLOR if passes else FAIL_COLOR
        edge    = PASS_EDGE  if passes else FAIL_EDGE
        verdict = 'PASS' if passes else 'FAIL'

        rect = plt.Rectangle(
            (col - 0.48, row - 0.48), 0.96, 0.96,
            facecolor=bg, edgecolor=edge, linewidth=1.5, zorder=1
        )
        ax.add_patch(rect)

        # Value
        display_val = format(val, fmt)
        if abs(val) > 9999:
            display_val = f'{val:.2e}'
        ax.text(col, row - 0.13, display_val,
                ha='center', va='center', fontsize=10,
                fontweight='bold', zorder=2,
                color=PASS_EDGE if passes else FAIL_EDGE)
        ax.text(col, row + 0.22, verdict,
                ha='center', va='center', fontsize=7.5, zorder=2,
                color=PASS_EDGE if passes else FAIL_EDGE)

# Axes labels
ax.set_xticks(range(n_criteria))
ax.set_xticklabels([c[0] for c in CRITERIA], fontsize=9)
ax.set_yticks(range(n_structs))
ax.set_yticklabels(struct_labels, fontsize=9.5, fontweight='bold')
ax.tick_params(length=0)
ax.set_title(
    'Scientific Defensibility Dashboard\n'
    'Green = criterion passes  |  Red = criterion fails',
    fontsize=12, fontweight='bold', pad=12
)

# Verdict column on the right
for row, (_, report) in enumerate(reports):
    ok = report['is_overall_scientifically_defensible']
    ax.text(n_criteria - 0.5 + 0.65, row,
            'ALL PASS' if ok else 'FAILED',
            ha='center', va='center', fontsize=9, fontweight='bold',
            color=PASS_EDGE if ok else FAIL_EDGE)
ax.text(n_criteria - 0.5 + 0.65, -0.48 - 0.12, 'Verdict',
        ha='center', va='center', fontsize=9, style='italic', color='#555')

ax.set_xlim(-0.5, n_criteria + 0.4)
ax.spines[:].set_visible(False)
plt.tight_layout()
plt.show()
# ── Criterion definitions ────────────────────────────────────────────────
CRITERIA = [
    # (display_name, report_key_fn, threshold_fn, format_str, invert_color)
    # invert_color=True means LOWER is better
    ('Bond Z-score\n(< 3.0)',
     lambda r: r['geometric_z_scores']['mean_bond_zscore'],
     lambda v: v < 3.0, '.3f', False),
    ('Rotamers\n(> 80 %)',
     lambda r: r['rotamer_stats']['favored_rotamers_pct'],
     lambda v: v > 80.0, '.0f', False),
    ('Ramachandran\noutliers (< 5 %)',
     lambda r: r['ramachandran_stats']['outlier_pct'],
     lambda v: v < 5.0, '.0f', True),
    ('Potential energy\n(< 1e5 kJ/mol)',
     lambda r: r['potential_energy_kj_mol'],
     lambda v: v < 1e5, '.0f', True),
    ('Burial ratio\n(>= 0.8)',
     lambda r: r['hydrophobic_burial_ratio'],
     lambda v: v >= 0.8, '.3f', False),
]

struct_labels = [label for label, _ in reports]
n_structs     = len(reports)
n_criteria    = len(CRITERIA)

fig, ax = plt.subplots(figsize=(11, 4.5))
ax.set_xlim(-0.5, n_criteria - 0.5)
ax.set_ylim(-0.5, n_structs - 0.5)
ax.invert_yaxis()

PASS_COLOR = '#d5f5e3'   # soft green
FAIL_COLOR = '#fadbd8'   # soft red
PASS_EDGE  = '#1e8449'
FAIL_EDGE  = '#922b21'

for col, (_crit_name, value_fn, pass_fn, fmt, _) in enumerate(CRITERIA):
    for row, (struct_label, report) in enumerate(reports):
        val     = value_fn(report)
        passes  = pass_fn(val)
        bg      = PASS_COLOR if passes else FAIL_COLOR
        edge    = PASS_EDGE  if passes else FAIL_EDGE
        verdict = 'PASS' if passes else 'FAIL'

        rect = plt.Rectangle(
            (col - 0.48, row - 0.48), 0.96, 0.96,
            facecolor=bg, edgecolor=edge, linewidth=1.5, zorder=1
        )
        ax.add_patch(rect)

        # Value
        display_val = format(val, fmt)
        if abs(val) > 9999:
            display_val = f'{val:.2e}'
        ax.text(col, row - 0.13, display_val,
                ha='center', va='center', fontsize=10,
                fontweight='bold', zorder=2,
                color=PASS_EDGE if passes else FAIL_EDGE)
        ax.text(col, row + 0.22, verdict,
                ha='center', va='center', fontsize=7.5, zorder=2,
                color=PASS_EDGE if passes else FAIL_EDGE)

# Axes labels
ax.set_xticks(range(n_criteria))
ax.set_xticklabels([c[0] for c in CRITERIA], fontsize=9)
ax.set_yticks(range(n_structs))
ax.set_yticklabels(struct_labels, fontsize=9.5, fontweight='bold')
ax.tick_params(length=0)
ax.set_title(
    'Scientific Defensibility Dashboard\n'
    'Green = criterion passes  |  Red = criterion fails',
    fontsize=12, fontweight='bold', pad=12
)

# Verdict column on the right
for row, (_, report) in enumerate(reports):
    ok = report['is_overall_scientifically_defensible']
    ax.text(n_criteria - 0.5 + 0.65, row,
            'ALL PASS' if ok else 'FAILED',
            ha='center', va='center', fontsize=9, fontweight='bold',
            color=PASS_EDGE if ok else FAIL_EDGE)
ax.text(n_criteria - 0.5 + 0.65, -0.48 - 0.12, 'Verdict',
        ha='center', va='center', fontsize=9, style='italic', color='#555')

ax.set_xlim(-0.5, n_criteria + 0.4)
ax.spines[:].set_visible(False)
plt.tight_layout()
plt.show()

🔍 Step 5: Why Each Failure Mode Is Distinct¶

S2 — Drifted decoy: only Ramachandran fails¶

Adding drift=120 perturbs backbone torsion angles far from their ideal values, scattering phi/psi angles into sterically forbidden regions. But the bond lengths (determined by Biotite templates, not torsions) remain correct — so the Z-score still passes. And since the sequence is the same as S1, the burial_ratio is identical.

Take-away: A Ramachandran failure catches backbone geometry errors that bond-length checks completely miss. Both are necessary.

S3 — Hydrophobic strand: burial and energy fail¶

The all-Val sequence has no polar residues, so mean_polar_sasa ≈ 0 and burial_ratio ≈ 0. The unminimized extended strand also has severe Van der Waals clashes (positive energy). Yet the Ramachandran angles for a beta strand are exactly in the allowed region, and bond lengths are template-correct.

Take-away: Burial ratio catches sequence-level design flaws that are invisible to geometry and backbone checks.

The composite flag is conservative by design¶

A structure passes is_overall_scientifically_defensible only if all five criteria pass simultaneously. This mirrors wwPDB deposition requirements, where a structure must clear geometry, rotamer, and Ramachandran checks independently.

📈 Step 6: Per-Criterion Score Comparison¶

A radar/bar view showing all five criterion values normalized to their thresholds:

In [ ]:

Copied!





# Normalise each criterion value to [0, 1] where 1.0 = exactly at threshold
# Values > 1 are above threshold (good for maximising criteria, bad for minimising)
NORM_FUNS = [
    ('Bond Z-score',      lambda r: r['geometric_z_scores']['mean_bond_zscore'] / 3.0),
    ('Rotamers %',        lambda r: r['rotamer_stats']['favored_rotamers_pct'] / 100.0),
    ('Rama outliers %',   lambda r: 1.0 - r['ramachandran_stats']['outlier_pct'] / 100.0),
    ('Energy / 1e5',      lambda r: max(0, 1.0 - r['potential_energy_kj_mol'] / 1e5)),
    ('Burial ratio',      lambda r: r['hydrophobic_burial_ratio'] / 0.8),
]

crit_names = [c[0] for c in NORM_FUNS]
colors = ['#2ecc71', '#e67e22', '#9b59b6']
x = np.arange(len(crit_names))
width = 0.25

fig, ax = plt.subplots(figsize=(11, 4.5))

for i, (struct_label, report) in enumerate(reports):
    vals = [norm_fn(report) for _, norm_fn in NORM_FUNS]
    bars = ax.bar(x + (i - 1) * width, vals, width,
                  label=struct_label.strip(), color=colors[i], alpha=0.82)

ax.axhline(1.0, color='#2c3e50', linestyle='--', linewidth=1.0,
           alpha=0.6, label='Pass threshold')
ax.set_xticks(x)
ax.set_xticklabels(crit_names, fontsize=10)
ax.set_ylabel('Normalised score  (> 1.0 = passes threshold)', fontsize=9)
ax.set_title('All Five Criteria — Normalised to Their Pass Thresholds',
             fontsize=12, fontweight='bold')
ax.legend(fontsize=8.5, loc='lower right')
ax.set_ylim(0, max(2.2, ax.get_ylim()[1]))

plt.tight_layout()
plt.show()

print('\nNormalised scores (> 1.0 means criterion passes):')
for struct_label, report in reports:
    vals = [norm_fn(report) for _, norm_fn in NORM_FUNS]
    print(f'{struct_label.strip()}: {", ".join(f"{n}={v:.2f}" for n, v in zip(crit_names, vals, strict=False))}')
# Normalise each criterion value to [0, 1] where 1.0 = exactly at threshold
# Values > 1 are above threshold (good for maximising criteria, bad for minimising)
NORM_FUNS = [
    ('Bond Z-score',      lambda r: r['geometric_z_scores']['mean_bond_zscore'] / 3.0),
    ('Rotamers %',        lambda r: r['rotamer_stats']['favored_rotamers_pct'] / 100.0),
    ('Rama outliers %',   lambda r: 1.0 - r['ramachandran_stats']['outlier_pct'] / 100.0),
    ('Energy / 1e5',      lambda r: max(0, 1.0 - r['potential_energy_kj_mol'] / 1e5)),
    ('Burial ratio',      lambda r: r['hydrophobic_burial_ratio'] / 0.8),
]

crit_names = [c[0] for c in NORM_FUNS]
colors = ['#2ecc71', '#e67e22', '#9b59b6']
x = np.arange(len(crit_names))
width = 0.25

fig, ax = plt.subplots(figsize=(11, 4.5))

for i, (struct_label, report) in enumerate(reports):
    vals = [norm_fn(report) for _, norm_fn in NORM_FUNS]
    bars = ax.bar(x + (i - 1) * width, vals, width,
                  label=struct_label.strip(), color=colors[i], alpha=0.82)

ax.axhline(1.0, color='#2c3e50', linestyle='--', linewidth=1.0,
           alpha=0.6, label='Pass threshold')
ax.set_xticks(x)
ax.set_xticklabels(crit_names, fontsize=10)
ax.set_ylabel('Normalised score  (> 1.0 = passes threshold)', fontsize=9)
ax.set_title('All Five Criteria — Normalised to Their Pass Thresholds',
             fontsize=12, fontweight='bold')
ax.legend(fontsize=8.5, loc='lower right')
ax.set_ylim(0, max(2.2, ax.get_ylim()[1]))

plt.tight_layout()
plt.show()

print('\nNormalised scores (> 1.0 means criterion passes):')
for struct_label, report in reports:
    vals = [norm_fn(report) for _, norm_fn in NORM_FUNS]
    print(f'{struct_label.strip()}: {", ".join(f"{n}={v:.2f}" for n, v in zip(crit_names, vals, strict=False))}')

🎓 Key Takeaways¶

The five criteria are orthogonal¶

Each measures a genuinely different aspect of structural quality:

Geometry (bonds) ← scale: ångström
Conformation (rotamers) ← rotational degrees of freedom
Backbone topology (Ramachandran) ← steric clash in phi/psi space
Thermodynamics (energy) ← kJ/mol force field
Composition (burial) ← Å² surface area per residue type

A structure that cheats on one dimension will be caught by a different check.

When to use `get_quality_report()` in practice¶

Workflow	What to look for
Before MD simulation	All criteria should pass; high energy = run minimization first
After homology modeling	Check Ramachandran and rotamers in modeled loops
Evaluating AlphaFold outputs	Burial_ratio failures in low-pLDDT regions = likely disordered
Synthetic dataset generation	Use `is_overall_scientifically_defensible` as a quality filter

What to explore next¶

The Oil Drop Model tutorial — deep dive into criterion 5 (burial ratio)
The Hard Decoy Challenge — use all five criteria to score a real decoy set
Co-evolutionary Fitness Landscape (incubator) — scan all single-point mutants through this scorecard to build a structural fitness map

🔬 The Scientific Defensibility Dashboard¶

What makes a protein structure scientifically defensible?¶

🔧 Setup¶

📐 Step 1: The Five Criteria — Scientific Background¶

Criterion 1 — Bond Z-score (Engh & Huber, 1991)¶

Criterion 2 — Rotamer Quality (Dunbrack, 1997)¶

Criterion 3 — Ramachandran Outliers (MolProbity)¶

Criterion 4 — Potential Energy (AMBER14 / OpenMM)¶

Criterion 5 — Hydrophobic Burial (Kauzmann, 1959)¶

🧬 Step 2: Three Structures, Three Failure Modes¶

📊 Step 3: Run get_quality_report() on Each Structure¶

🎨 Step 4: The Dashboard Heatmap¶

🔍 Step 5: Why Each Failure Mode Is Distinct¶

S2 — Drifted decoy: only Ramachandran fails¶

S3 — Hydrophobic strand: burial and energy fail¶

The composite flag is conservative by design¶

📈 Step 6: Per-Criterion Score Comparison¶

🎓 Key Takeaways¶

The five criteria are orthogonal¶

When to use get_quality_report() in practice¶

What to explore next¶

📊 Step 3: Run `get_quality_report()` on Each Structure¶

When to use `get_quality_report()` in practice¶