Confidence Metrics

Understanding pLDDT, PAE, pTM, and ipTM confidence scores

Overview

AlphaFold 3 provides multiple confidence metrics to assess prediction quality. These metrics help you understand:

• Local accuracy: How confident is each atom/residue?
• Global accuracy: Is the overall fold correct?
• Interface quality: Are chain-chain interactions accurate?
• Relative positions: How accurate are token-token distances?

Similar to AlphaFold 2 and AlphaFold-Multimer, but with enhanced metrics for ligands, nucleic acids, and multi-chain complexes.

pLDDT: Per-Atom Confidence

predicted Local Distance Difference Test

Range: 0-100 (higher = more confident)

Granularity: Per-atom (more detailed than AlphaFold 2’s per-residue)

What it measures: Predicted accuracy of local structure

Understanding pLDDT Values

Confidence Ranges

pLDDT Score	Interpretation	Typical Usage
90-100	Very high confidence	Well-ordered regions, reliable for detailed analysis
70-90	Confident	Generally accurate, suitable for most applications
50-70	Low confidence	Use with caution, may be flexible/disordered
< 50	Very low confidence	Likely disordered, unreliable coordinates

Polymer-Specific

Proteins

• Similar to AlphaFold 2’s lDDT-Cα metric
• Per-atom granularity (not just Cα)
• Considers distances to other polymer atoms within 15Å

DNA/RNA

• Uses wider 30Å radius (vs. 15Å for proteins)
• Reflects nucleic acid structural characteristics
• Modified LDDT considers errors only to polymers

Ligands

• Modified LDDT considers only ligand-to-polymer distances
• Does NOT consider ligand-to-ligand atom distances
• Focuses on protein-ligand interface quality

Accessing pLDDT Values

In the confidences JSON:

{
  "atom_plddts": [95.2, 94.8, 93.1, ...],  // num_atoms array
  "atom_chain_ids": ["A", "A", "A", ...]   // maps atoms to chains
}

Visualization

Most structure viewers can color by B-factor. AlphaFold 3 stores pLDDT in the B-factor column:

PyMOL

# Load and color by pLDDT (stored in B-factor)
load model.cif
spectrum b, blue_white_red, minimum=50, maximum=90

PAE: Predicted Aligned Error

Positional Uncertainty

Range: 0-32 Ångströms (lower = more confident)

Granularity: Per-token pair

What it measures: Error in relative position/orientation between two tokens

How PAE Works

PAE(i, j) estimates the error in token j’s position when the structure is aligned using token i’s coordinate frame.

1. Alignment

   Align predicted structure to true structure using token i’s frame

2. Measurement

   Measure positional error of token j after alignment

3. Prediction

   PAE predicts this error in Ångströms

Interpreting PAE Plots

PAE is visualized as a 2D heatmap where both axes represent token indices.

Low PAE (dark blue): Confident in relative positions

• Diagonal blocks: Well-structured domains
• Off-diagonal blocks: Confident interfaces/contacts

High PAE (yellow/red): Uncertain relative positions

• Between domains: Flexible linkers
• Between chains: Uncertain/no interaction

Frame Construction

Proteins & Nucleic Acids

Frames constructed from backbone atoms:

• Proteins: Same as AlphaFold 2 (Cα, N, C)
• Nucleic acids: Backbone phosphate groups

Small Molecules

Frames constructed for each atom from:

• The atom itself
• Its closest neighbors in reference conformer
• Enables PAE for ligands and PTMs

Accessing PAE Values

In the confidences JSON:

{
  "pae": [[0.5, 1.2, 15.3, ...],    // num_tokens × num_tokens
          [1.3, 0.4, 14.8, ...],
          [15.1, 14.9, 0.6, ...]],
  "token_chain_ids": ["A", "A", "B", ...]  // maps tokens to chains
}

Use Cases

Domain Organization

Low within-domain PAE, high between-domain PAE suggests multi-domain protein with flexible linkers

Interaction Confidence

Low off-diagonal PAE between chains indicates confident protein-protein interface

Binding Assessment

High PAE between protein and ligand suggests uncertain binding pose

Relative Orientation

PAE reveals which regions have well-defined spatial relationships

pTM: Predicted Template Modeling Score

Global Structure Confidence

Range: 0-1 (higher = more confident)

Granularity: Global scalar

What it measures: Accuracy of entire predicted structure

Interpretation Guidelines

Overall Quality

pTM Score	Interpretation
> 0.8	Very confident, high-quality prediction
0.5-0.8	Decent confidence, likely correct overall fold
< 0.5	Low confidence, fold may be incorrect

Size Considerations

pTM is strict for small structures. Values < 0.05 for structures with fewer than 20 tokens.

For small structures or short chains:

• Use PAE or pLDDT instead for quality assessment
• pTM becomes more reliable with >20 tokens
• ipTM more informative for multi-chain systems

TM-Score Background

The Template Modeling score measures structural similarity:

$$\text{TM-score} = \max \left[ \frac{1}{L} \sum_{i=1}^{L} \frac{1}{1 + (d_i / d_0)^2} \right]$$

Where:

• L: Length of target sequence
• d_i: Distance between aligned residues
• d_0: Normalization factor (length-dependent)

References: Zhang & Skolnick, 2004; Xu & Zhang, 2010

Accessing pTM

In summary confidences JSON:

{
  "ptm": 0.84,
  "chain_ptm": [0.89, 0.76, 0.92]  // per-chain pTM
}

ipTM: Interface Predicted TM Score

Interface Quality

Range: 0-1 (higher = more confident)

Granularity: Global or per-chain-pair

What it measures: Accuracy of predicted interfaces between chains

Interpretation Guidelines

ipTM Score	Interpretation	Action
> 0.8	Confident, high-quality interface	Trust binding mode
0.6-0.8	Uncertain region	Verify with experiments
< 0.6	Low confidence	Likely incorrect interface

ipTM only considers inter-chain interactions. For single-chain predictions, use pTM instead.

Types of ipTM Metrics

Global ipTM

Overall confidence in all interfaces:

{
  "iptm": 0.72  // All chain-chain interfaces
}

Use for ranking predictions of the full complex.

Per-Chain ipTM

Average interface confidence per chain:

{
  "chain_iptm": [0.85, 0.68, 0.91]
}

Element i = average ipTM between chain i and all other chains.

Use case: Ranking predictions when interested in specific chain (e.g., ligand binding)

Pair-wise ipTM

Interface confidence for each chain pair:

{
  "chain_pair_iptm": [
    [0.89, 0.72, 0.15],  // Chain A to A,B,C
    [0.72, 0.84, 0.20],  // Chain B to A,B,C
    [0.15, 0.20, 0.78]   // Chain C to A,B,C
  ]
}

• Diagonal (i,i): pTM for chain i
• Off-diagonal (i,j): ipTM for i-j interface

Use case: Antibody-antigen, known interacting pairs

Ranking Strategies

1. Full Complex

   Use global iptm or composite ranking_score

2. Specific Interface

   Use chain_pair_iptm[i][j] for known interacting chains i and j

3. Specific Chain

   Use chain_iptm[i] when you care about chain i’s interactions but don’t know partners

4. Chain Structure

   Use chain_ptm[i] when only the structure of chain i matters

Ranking Score

A composite score combining multiple metrics:

ranking_score = 0.8 × ipTM + 0.2 × pTM + 0.5 × disorder - 100 × has_clash

Use ranking_score ONLY for ranking predictions, not for absolute quality assessment.

Components

ipTM · number

Interface confidence (0-1) - weighted 0.8

pTM · number

Overall structure confidence (0-1) - weighted 0.2

disorder · number

Fraction disordered (0-1) - penalizes spurious helices (+0.5)

has_clash · boolean

Significant clashing atoms - strong penalty (-100)

When to Use

Use Ranking Score For

• Selecting best prediction from multiple samples
• Comparing predictions with same input
• Automated prediction pipelines
• Initial filtering before manual inspection

Don't Use Ranking Score For

• Absolute quality assessment
• Comparing different proteins
• Scientific reporting (use pTM/ipTM)
• Deciding if prediction is correct

Additional Metrics

Contact Probabilities

{
  "contact_probs": [[0.95, 0.12, ...],  // num_tokens × num_tokens
                     [0.12, 0.98, ...]]
}

Probability that two tokens are within 8Å (representative atoms).

Use cases:

• Identifying likely contacts before structure determination
• Assessing binding likelihood
• Filtering potential interfaces

Chain-Pair PAE Min

{
  "chain_pair_pae_min": [
    [0.5, 15.2, 18.7],
    [15.1, 0.4, 19.2],
    [18.9, 19.5, 0.6]
  ]
}

Lowest PAE value from chain i to chain j.

Correlates with binding: Low chain_pair_pae_min between chains suggests they interact. Can distinguish binders from non-binders.

Disorder Fraction

{
  "fraction_disordered": 0.23
}

Fraction of structure classified as disordered (0-1), measured by accessible surface area.

Interpretation:

• High disorder: Flexible regions, loops, or prediction artifacts
• Used in ranking score to penalize spurious helices

Clash Detection

{
  "has_clash": false
}

Boolean indicating significant clashing atoms:

• More than 50% of any chain has clashes, OR
• More than 100 clashing atoms in any chain

Action: If true, consider re-running with different seeds or inspect manually.

Practical Guidelines

1. Initial Assessment

   Check ranking_score, ptm, and iptm in summary JSON

   • pTM > 0.5: Likely correct fold
   • ipTM > 0.8: Confident interface
   • has_clash = false: No major issues
2. Visual Inspection

   Load structure colored by pLDDT (B-factor)

   • Blue regions (>90): High confidence
   • Red regions (<50): Disordered/uncertain
3. Interface Analysis

   Examine PAE plot

   • Low PAE blocks between chains: Good interface
   • High PAE: Uncertain interaction
4. Chain-Specific Quality

   Check per-chain metrics

   • chain_ptm[i]: Structure of chain i
   • chain_iptm[i]: Chain i’s interfaces
   • chain_pair_pae_min[i][j]: Likely interaction between i and j
5. Domain Analysis

   For multi-domain proteins:

   • Low within-domain PAE: Well-structured domains
   • High between-domain PAE: Flexible linkers

Additional Resources

AlphaFold 3 Paper

Detailed description of confidence metrics

EBI Training

Practical guide to confidence scores (AlphaFold 2 focused, still applicable)

Next Steps

Output Format

Learn about output file structure

Performance

Optimize predictions for better results