Protein Design Quality Control

Critical Limitation

Individual metrics have weak predictive power for binding. Research shows:

• Individual metric ROC AUC: 0.64-0.66 (slightly better than random)
• Metrics are pre-screening filters, not affinity predictors
• Composite scoring is essential for meaningful ranking

These thresholds filter out poor designs but do NOT predict binding affinity.

QC Organization

QC is organized by purpose and level:

Each category has two levels:

• Metric-level: Calculated values with thresholds (pLDDT > 0.85)
• Design-level: Pattern/motif detection (odd cysteines, NG sites)

Quick Reference: All Thresholds

Design-Level Checks (Expression)

See: references/binding-qc.md, references/expression-qc.md, references/structural-qc.md

Interface metrics (PyRosetta)

Beyond shape complementarity and interface_dG, two interface metrics catch common de novo failure modes:

• Buried unsatisfied H-bonds (BUNS): buried polar atoms making no hydrogen bond.

This is a dominant energetic failure mode and is orthogonal to dG and dSASA. Keep interface BUNS at or below 4 (standard) or 2 (stringent).

• ContactMolecularSurface: shape-complementarity-weighted contact area that, unlike

dSASA, is not fooled by gappy or holey interfaces. Higher is better.

Both are in the Cao 2022, AlphaProteo, and BindCraft filter sets.

For structure-quality ranking, biomodals also provides modal_af2rank.py (AF2Rank), which scores how well a design re-predicts from its own structure as a template.

Binder ranking (benchmark-backed)

A meta-analysis of 3,766 experimentally tested binders across 15 targets (Overath et al., bioRxiv 2025, doi:10.1101/2025.08.14.670059) found that AF3 ipSAE_min is the single best in-silico predictor of binding, and that a simple linear model of three features generalizes best across targets. Complexity did not help: gradient-boosted and many-feature models did not beat the linear one.

Recommended filtering strategies from that work:

1. Threshold on one of: AF3 ipSAE_min > 0.61, or `ipSAE_min × interface_ΔG/ΔSASA <

-1.5, or LIS × shape_complementarity > 0.42`.

1. Pre-filter on shape_complementarity > 0.62 and RMSD_binder < 3.73 (input vs

re-predicted), then take the top-K by ipSAE_min.

ipSAE_min is the minimum of the two asymmetric ipSAE directions (binder→target and target→binder), not the average or max. Use the ipsae skill to compute it. Note the RMSD_binder filter can be over-restrictive on some targets, so prefer it as a soft pre-filter rather than a hard cutoff.

Stability prediction (small domains)

For small domains (roughly 60 to 80 residues, the minibinder range), absolute folding stability can be predicted directly. SaProtΔG (Cho et al., bioRxiv 2026, doi:10.64898/2026.05.19.726285) predicts absolute folding ΔG at about 0.8 kcal/mol RMSE and improves discrimination of stable versus unstable designed proteins. Use the SaProt variant rather than the ESM3 variant for commercial work, since the ESM3 weights are non-commercial. Filter for more negative (more stable) ΔG.

Sequence-liability scan

Implement liability checks directly as motif rules rather than taking an antibody-specific dependency. Severity rises with solvent exposure (gate by SASA when a structure is available).

import re

LIABILITIES = {
    "deamidation":   (r"N[GSNTH]",   "NG/NS high, NN/NT moderate"),
    "isomerization": (r"D[GSTDH]",   "Asp isomerization"),
    "N-glycosylation": (r"N[^P][ST]", "NxS/T sequon"),
    "polybasic":     (r"[KR]{3,}",   "proteolysis / charge patch"),
    "hydrophobic_run": (r"[AILMFWVY]{6,}", "aggregation"),
}

def scan_liabilities(seq):
    hits = {}
    for name, (pattern, note) in LIABILITIES.items():
        positions = [m.start() for m in re.finditer(pattern, seq)]
        if positions:
            hits[name] = (positions, note)
    # Unpaired cysteine check
    if seq.count("C") % 2 == 1:
        hits["unpaired_cysteine"] = ([seq.index("C")], "odd cysteine count")
    return hits

Met and Trp oxidation are also liabilities but should be flagged only when the residue is solvent-exposed.

Sequential Filtering Pipeline

import pandas as pd

designs = pd.read_csv('designs.csv')

# Stage 1: Structural confidence
designs = designs[designs['pLDDT'] > 0.85]

# Stage 2: Self-consistency
designs = designs[designs['scRMSD'] < 2.0]

# Stage 3: Binding quality
designs = designs[(designs['ipTM'] > 0.5) & (designs['PAE_interaction'] < 10)]

# Stage 4: Sequence plausibility
designs = designs[designs['esm2_pll_normalized'] > 0.0]

# Stage 5: Expression checks (design-level)
designs = designs[designs['cysteine_count'] % 2 == 0]  # Even cysteines
designs = designs[designs['instability_index'] < 40]

Composite Scoring (Required for Ranking)

Individual metrics alone are too weak. Use composite scoring:

def composite_score(row):
    return (
        0.30 * row['pLDDT'] +
        0.20 * row['ipTM'] +
        0.20 * (1 - row['PAE_interaction'] / 20) +
        0.15 * row['shape_complementarity'] +
        0.15 * row['esm2_pll_normalized']
    )

designs['score'] = designs.apply(composite_score, axis=1)
top_designs = designs.nlargest(100, 'score')

For advanced composite scoring, see references/composite-scoring.md.

Tool-Specific Filtering

BindCraft Filter Levels

BoltzGen Filtering

boltzgen run ... \
  --budget 60 \
  --alpha 0.01 \
  --filter_biased true \
  --refolding_rmsd_threshold 2.0 \
  --additional_filters 'ALA_fraction<0.3'

• alpha=0.0: Quality-only ranking
• alpha=0.01: Default (slight diversity)
• alpha=1.0: Diversity-only

Design-Level Severity Scoring

For pattern-based checks, use severity scoring:

Experimental Correlation

Key insight: Metrics work as filters (eliminating failures) not predictors (ranking successes).

Campaign Health Assessment

Quick assessment of your design campaign:

Failure Recovery Trees

Too Few Pass pLDDT Filter (< 5% with pLDDT > 0.85)

Low pLDDT across campaign
├── Check scRMSD distribution
│   ├── High scRMSD (>2.5Å): Backbone issue
│   │   └── Fix: Regenerate backbones with lower noise_scale (0.5-0.8)
│   └── Low scRMSD but low pLDDT: Disordered regions
│       └── Fix: Check design length, simplify topology
├── Try more sequences per backbone
│   └── modal run modal_ligandmpnn.py --input-pdb bb.pdb --params-str "--number_of_batches 32 --temperature 0.1"
├── Use SolubleMPNN instead of ProteinMPNN
│   └── Better for expression-optimized sequences
└── Consider different design tool
    └── BindCraft (integrated design) may work better

Too Few Pass ipTM Filter (< 5% with ipTM > 0.5)

Low ipTM across campaign
├── Review hotspot selection
│   ├── Are hotspots surface-exposed? (SASA > 20Ų)
│   ├── Are hotspots conserved? (check MSA)
│   └── Try 3-6 different hotspot combinations
├── Increase binder length (more contact area)
│   └── Try 80-100 AA instead of 60-80 AA
├── Check interface geometry
│   ├── Is target flat? → Try helical binders
│   └── Is target concave? → Try smaller binders
└── Try all-atom design tool
    └── BoltzGen (all-atom, better packing)

High scRMSD (> 50% with scRMSD > 2.0Å)

Sequences don't specify intended structure
├── ProteinMPNN issue
│   ├── Lower temperature: --sampling_temp "0.1"
│   ├── Increase sequences: --num_seq_per_target 32
│   └── Check fixed_positions aren't over-constraining
├── Backbone geometry issue
│   ├── Backbones may be unusual/strained
│   ├── Regenerate with lower noise_scale (0.5-0.8)
│   └── Reduce diffuser.T to 30-40
└── Try different sequence design
    └── ColabDesign (AF2 gradient-based) may work better

Everything Passes But No Experimental Hits

In silico metrics don't predict affinity
├── Generate MORE designs (10x current)
│   └── Computational metrics have high false positive rate
├── Increase diversity
│   ├── Higher ProteinMPNN temperature (0.2-0.3)
│   ├── Different backbone topologies
│   └── Different hotspot combinations
├── Try different design approach
│   ├── BindCraft (different algorithm)
│   ├── ColabDesign (AF2 hallucination)
│   └── BoltzGen (all-atom diffusion)
└── Check if target is druggable
    └── Some targets are inherently difficult

Too Many Designs Pass (> 50%)

Suspiciously high pass rate
├── Check if thresholds are too lenient
│   └── Use stringent thresholds: pLDDT > 0.90, ipTM > 0.60
├── Verify prediction quality
│   ├── Are predictions actually running? Check output files
│   └── Are complexes being predicted, not just monomers?
├── Check for data issues
│   ├── Same sequence being predicted multiple times?
│   └── Wrong FASTA format (missing chain separator)?
└── Apply diversity filter
    └── Cluster at 70% identity, take top per cluster

Diagnostic Commands

Quick Campaign Assessment

import pandas as pd

df = pd.read_csv('designs.csv')

# Pass rates at each stage
print(f"Total designs: {len(df)}")
print(f"pLDDT > 0.85: {(df['pLDDT'] > 0.85).mean():.1%}")
print(f"ipTM > 0.50: {(df['ipTM'] > 0.50).mean():.1%}")
print(f"scRMSD < 2.0: {(df['scRMSD'] < 2.0).mean():.1%}")
print(f"All filters: {((df['pLDDT'] > 0.85) & (df['ipTM'] > 0.5) & (df['scRMSD'] < 2.0)).mean():.1%}")

# Identify top issue
if (df['pLDDT'] > 0.85).mean() < 0.1:
    print("ISSUE: Low pLDDT - check backbone or sequence quality")
elif (df['ipTM'] > 0.50).mean() < 0.1:
    print("ISSUE: Low ipTM - check hotspots or interface geometry")
elif (df['scRMSD'] < 2.0).mean() < 0.5:
    print("ISSUE: High scRMSD - sequences don't specify backbone")