IPSAE

What is IPSAE?

ipSAE (interaction prediction Score from Aligned Errors) is a replacement for AlphaFold's ipTM score, developed by Roland Dunbrack Jr. at the Fox Chase Cancer Center. It addresses a fundamental flaw in how AlphaFold evaluates protein-protein interactions: ipTM averages predicted alignment quality across all residue pairs between chains, including disordered regions and non-interacting domains that drag the score down even when the actual interface is well-predicted.

The problem becomes acute with full-length protein sequences. A kinase dimer predicted with high confidence at its interface can receive a mediocre ipTM simply because both chains contain long disordered tails. Trimming those tails raises the ipTM score despite the structure prediction being identical — a clear sign the metric is measuring the wrong thing.

ipSAE fixes this by filtering out residue pairs with high predicted aligned error (PAE) before scoring, and adjusting the length normalization to reflect only high-confidence residues. In benchmarks on 40 heterodimers with paired decoys, ipSAE cleanly separates true from false interactions even with full-length UniProt sequences, where ipTM struggles.

Beyond ipSAE itself, the tool also computes ipTM, pDockQ, pDockQ2, and LIS (Local Interaction Score) — giving a comprehensive view of interface quality from a single run.

How does ipSAE work?

AlphaFold's ipTM is derived from the TM-score formula:

$\text{TM} = \frac{1}{L_{\text{target}}} \sum_j \frac{1}{1 + (d_j / d_0)^2}$

where $d_0 = 1.24 \sqrt{L_{\text{target}} - 15} - 1.8$ scales with protein length. The ipTM applies this across chains, averaging over all interchain residue pairs regardless of prediction quality.

ipSAE modifies this in three ways:

1. PAE filtering: Only residue pairs where the predicted aligned error falls below a cutoff (default 10 Å) contribute to the score. Disordered regions and non-interacting domains are excluded automatically.

2. Adjusted normalization: The $d_0$ parameter is recalculated using only the count of residues that pass the PAE filter, not the total chain length. Short effective lengths ($L < 27$) use $d_0 = 1.0$.

The score is computed asymmetrically — ipSAE(A→B) and ipSAE(B→A) — and the final value is the maximum of the two directions.

Normalization variants

The tool reports three ipSAE variants differing in how $d_0$ is calculated:

How to use IPSAE online

ProteinIQ runs the ipSAE scoring pipeline in the cloud — upload prediction files and get scores back without installing Python dependencies or writing command-line arguments.

Inputs

Settings

Output columns

Interpreting results

ipSAE scores

ipSAE ranges from 0 to 1. A high score means AlphaFold confidently predicts the interface with low alignment error. A score near zero means AlphaFold found no confident interchain contacts — strong evidence against a physical interaction.

The key advantage over ipTM is visible in borderline cases. For the RAF1-KRAS kinase interaction with full-length sequences, ipTM reports 0.59 (ambiguous), while ipSAE reports 0.80 (clearly positive). For a false pairing like RAF1-RIPK1, ipTM gives 0.28 while ipSAE returns 0.0.

pDockQ and pDockQ2

pDockQ estimates the DockQ quality of a predicted complex based on interface pLDDT and contact count:

pDockQ2 refines this with per-residue analysis and tends to be more reliable for complexes with multiple interfaces.

Limitations

• ipSAE scores are only meaningful for predictions from AlphaFold2, AlphaFold3, or Boltz — it requires a PAE matrix, which experimental structures do not produce.
• Like all confidence metrics, ipSAE reflects model certainty, not ground truth. A high score means AlphaFold is confident, not that the interaction necessarily occurs in vivo.
• Optimal PAE cutoff values may vary by application. The default of 10 Å works well in published benchmarks, but screening peptide binders or highly flexible interfaces may benefit from a slightly higher cutoff (12–15 Å).