Every candidate design generated by Foldexa is validated using AlphaFold2 structure prediction. This article explains our scoring metrics — pLDDT, iptm, PAE — the Foldexa Score formula, and how we rank candidate outputs from both pipelines.
Azamat Armanuly
CEO & Bioengineer, KAIST
Every design generated by Foldexa — whether from DiffAb or RFdiffusion — must pass structural validation before being shown to you. Here is exactly how that validation works.
AlphaFold2 (and AlphaFold2 Multimer for complexes) produces four key numerical outputs beyond the predicted 3D structure itself. Foldexa uses three of them in its validation pipeline.
| Metric | Range | Foldexa Use | Notes |
|---|---|---|---|
| pLDDT | 0–100 | Primary filter | Per-residue confidence; > 85 is high confidence |
| iptm | 0–1 | Complex confidence | Interface predicted TM-score; > 0.7 is good |
| PAE | 0–35 Å | Interface quality | Predicted Aligned Error; lower = more confident |
| ptm | 0–1 | Internal check | Global chain confidence; not exposed in UI |
pLDDT (predicted Local Distance Difference Test) measures how confident AlphaFold2 is in each residue's position relative to its neighbors within 15 Å. It is not a measure of how well the structure will match an experimental structure — it is a measure of internal consistency. A protein that is predicted to be intrinsically disordered will have low pLDDT even if the prediction is correct.
Common Misconception
High pLDDT does not mean the protein folds in reality — it means AlphaFold2 is internally consistent about the predicted structure. A well-packed, globular design will naturally score higher than a flexible loop region. Always consider the context: CDR loops legitimately score lower than framework regions.
| pLDDT Range | Tier | Foldexa Interpretation |
|---|---|---|
| ≥ 90 | Very High | Backbone highly reliable; suitable for structure-guided downstream design |
| 70 – 89 | High | Good confidence; acceptable for most applications |
| 50 – 69 | Low | Possibly disordered or low-confidence region; review manually |
| < 50 | Very Low | Potentially unstructured; design likely does not fold as intended |
Individual metrics are combined into a single Foldexa Score for ranking. We use a weighted composite that emphasizes interface confidence for complex designs:
def foldexa_score(plddt_mean: float, iptm: float, pae_interface_mean: float) -> float:
"""
Compute Foldexa Score (0–1).
Args:
plddt_mean: Mean pLDDT of CDR or scaffold residues (0–100)
iptm: AlphaFold2 Multimer iptm score (0–1)
pae_interface_mean: Mean PAE across interface residue pairs (Å)
Returns:
Foldexa Score between 0 and 1
"""
plddt_norm = plddt_mean / 100.0
pae_norm = max(0.0, 1.0 - pae_interface_mean / 50.0)
return 0.40 * plddt_norm + 0.40 * iptm + 0.20 * pae_norm| Tier | Foldexa Score | Approx. Percentile | Recommendation |
|---|---|---|---|
| S | ≥ 0.85 | Top 5% | Prioritize for wet-lab synthesis |
| A | 0.75 – 0.84 | Top 20% | Strong candidates; validate before synthesis |
| B | 0.65 – 0.74 | Top 40% | Acceptable; use for backup or diversity |
| C | < 0.65 | Bottom 60% | Flagged; review manually or discard |
Practical Guidance
For a typical 50-design campaign, expect 4–7 S-tier designs, 8–12 A-tier, and the rest in B or C. Ordering 3–5 S-tier designs for expression is a cost-effective starting point — enough diversity to hedge against expression failure.
{
"job_id": "job_9f3a2b",
"status": "completed",
"designs": [
{
"design_id": "d001",
"rank": 1,
"tier": "S",
"foldexa_score": 0.912,
"plddt_mean": 92.4,
"iptm": 0.91,
"pae_interface_mean": 4.2,
"sequence": "EVQLVESGGGLVQPGGSLRL...",
"pdb_url": "https://results.foldexa.com/job_9f3a2b/d001.pdb",
"passed_self_consistency": true,
"self_consistency_rmsd": 0.74
}
],
"summary": {
"total_designs": 50,
"passed_plddt_85": 44,
"tier_s": 5,
"tier_a": 11,
"tier_b": 18,
"tier_c": 16
}
}Written by
Azamat Armanuly
CEO & Bioengineer, KAIST