Beyond pLDDT, AlphaFold2 provides global confidence metrics—pTM and ipTM—that assess overall structural quality and interface reliability. Master these metrics for comprehensive quality assessment.
#Understanding pTM (predicted TM-score)
pTM estimates the overall quality of the predicted structure by predicting the TM-score against the true structure.
What is TM-score?
TM-score Background
TM-score measures structural similarity between two proteins, ranging from 0 to 1:
- > 0.5: Same fold (statistically significant)
- > 0.6: Topology-level similarity
- > 0.8: Very similar structures
AlphaFold2's pTM predicts what the TM-score would be if compared to the true experimental structure.
Interpreting pTM Scores
> 0.8
High Confidence
0.6-0.8
Moderate
< 0.6
Low
pTM > 0.8
Strong indicator that the overall fold is correct. Individual regions may still have issues—check pLDDT!
pTM < 0.6
Suggests significant structural uncertainty. Use with caution and validate extensively.
#Understanding ipTM (interface pTM)
For multimer predictions, ipTM specifically measures confidence in the protein-protein interface.
How ipTM Works
ipTM focuses only on inter-chain residue pairs (residues from different chains that are close in 3D space):
- Considers only residues at chain-chain interfaces
- Assesses relative positioning between chains
- Independent of individual chain quality
Interpreting ipTM Scores
> 0.7
High Confidence
0.5-0.7
Moderate
< 0.5
Low
ipTM Thresholds
Note that ipTM thresholds are generally lower than pTM thresholds because interface prediction is inherently more challenging.
#Combining Metrics for Complete Picture
Understanding Metric Combinations
Different metric combinations tell different stories:
High pTM + High ipTM
Best case: Both individual structures and interface are well-predicted. High confidence overall.
High pTM + Low ipTM
Common scenario: Individual proteins are well-predicted, but their relative positioning is uncertain. May indicate:
- Weak or transient interaction
- Multiple possible binding modes
- Missing biological context (ligand, membrane, etc.)
Low pTM + High ipTM
Unusual case: Interface confidence without overall structure confidence. Suggests:
- Large disordered regions away from interface
- Multi-domain proteins with flexible linkers
- Check PAE matrix carefully!
#pTM vs. pLDDT: When to Use Which
Complementary Information
- pLDDT: Local, per-residue confidence
- pTM: Global, overall fold confidence
- ipTM: Interface-specific confidence (multimers only)
Decision Matrix
Use this guide to interpret combined metrics:
bash
High pLDDT + High pTM:
→ Excellent prediction, proceed with confidence
High pLDDT + Low pTM:
→ Rare, check for domain arrangement issues
Low pLDDT + High pTM:
→ Disordered regions present, but fold is correct
→ Check which regions have low pLDDT
Low pLDDT + Low pTM:
→ Unreliable prediction overall
→ Consider alternative methods (ESMFold, experimental)#Calculating and Extracting Metrics
From AlphaFold2 Output
Most AlphaFold2 implementations report these metrics automatically:
json
{
"model_1": {
"ptm": 0.873,
"iptm": 0.756,
"ranking_confidence": 0.834
}
}Manual Calculation
If metrics aren't provided, you can calculate from PAE matrix:
python
import numpy as np
import json
# Load PAE matrix
with open('pae_matrix.json') as f:
data = json.load(f)
pae = np.array(data['predicted_aligned_error'])
# Calculate pTM (simplified)
def calculate_ptm(pae_matrix, threshold=8.0):
d0 = 1.24 * (len(pae_matrix) - 15) ** (1.0/3.0) - 1.8
scores = 1.0 / (1.0 + (pae_matrix / d0) ** 2)
return scores.mean()
ptm = calculate_ptm(pae)
print(f"pTM: {ptm:.3f}")#Using Metrics for Model Ranking
AlphaFold2 generates 5 models per prediction. They're ranked by a confidence score that combines metrics:
Ranking Confidence
For single chains:
python
ranking_confidence = 0.8 * pTM + 0.2 * mean(pLDDT)For multimers:
python
ranking_confidence = 0.8 * ipTM + 0.2 * pTMModel Selection
While rank_1 model usually has highest confidence, checking all 5 models for consistency is best practice!
#Case Studies
Case 1: High-Quality Monomer
bash
Protein: 250 residues, well-studied enzyme
pTM: 0.91
Mean pLDDT: 92.3
Assessment: Excellent prediction, suitable for all applicationsCase 2: Heterodimer Complex
bash
Complex: A (180 res) + B (220 res)
pTM: 0.87 (both chains well-predicted)
ipTM: 0.62 (moderate interface confidence)
Mean pLDDT: 88.5
Assessment:
- Individual structures reliable
- Interface geometry uncertain
- Validate interface experimentally
- Check for alternative binding modesCase 3: Protein with Disorder
bash
Protein: 340 residues, signaling protein
pTM: 0.78
Mean pLDDT: 68.2 (N-term 45, Core 91, C-term 52)
Assessment:
- Core domain well-predicted (high pLDDT)
- Termini disordered (low pLDDT, expected)
- Overall pTM acceptable given disorder
- Use core for structural analysis#Advanced Analysis Techniques
PAE Matrix Decomposition
Extract more information from the PAE matrix:
- Domain identification: Cluster low-PAE regions
- Confidence profiles: Row/column averages show relative confidence
- Interface mapping: Off-diagonal blocks reveal inter-chain confidence
Model Ensemble Analysis
Compare metrics across all 5 models:
python
# Calculate metric spread across models
ptm_values = [0.87, 0.86, 0.85, 0.83, 0.81]
ptm_std = np.std(ptm_values) # 0.023
if ptm_std < 0.05:
print("Consistent prediction across models")
else:
print("High model variability - examine differences")#Limitations and Caveats
Important Limitations
- Metrics are predictions, not ground truth
- High confidence doesn't guarantee correctness
- Low confidence doesn't always mean wrong
- Experimental validation remains gold standard
Analyze Your Predictions
Use our confidence metric analyzer tools
#Best Practices Summary
Confidence Metric Checklist
- ✓ Always check pLDDT, pTM, and ipTM (if multimer)
- ✓ Use pLDDT for local confidence, pTM for global
- ✓ Compare all 5 models for consistency
- ✓ Interpret metrics in biological context
- ✓ Validate predictions experimentally when possible
- ✓ Document all confidence scores in publications
