Pinc: a simple probabilistic AlphaFold interaction score

The paper introduces Pinc, a simple probabilistic scoring metric that converts AlphaFold predicted aligned errors into calibrated contact probabilities to more sensitively identify protein interactions, particularly those with smaller interfaces, and provides tools for its implementation.

The Gist

Imagine you are trying to build a complex Lego castle with a friend, but you only have a blurry, computer-generated sketch of what the finished castle should look like. You both have a hunch about how the pieces fit together, but you need a way to know: "Is this sketch actually right, or did the computer just guess?"

For years, scientists have used a super-smart AI called AlphaFold to predict how proteins (the tiny machines inside our cells) stick together. But AlphaFold gives its answers in a confusing language of numbers and scores that are hard to interpret. It's like the AI saying, "I'm 85% sure," but not telling you what that 85% actually means in the real world.

This paper introduces a new tool called Pinc (Probability of Interface Native Contacts) to translate those confusing numbers into plain English.

The Problem: The "Confidence Score" Riddle

When AlphaFold predicts two proteins interacting, it gives a score (like ipTM). Think of this score like a weather forecast that says, "There is a 0.5 chance of rain."

• Is that good? Is it bad?
• Does it mean it will drizzle or pour?
• Does it depend on how big the city is?

The old scores were tricky. They changed based on how long the proteins were, making it hard to compare a tiny interaction with a huge one. Scientists needed a way to say, "This prediction is 80% likely to be correct," in a way that makes immediate sense.

The Solution: Pinc (The "Truth-O-Meter")

The authors, Mihaly Badonyi and Agnes Toth-Petroczy, created Pinc. Here is how it works, using a simple analogy:

The Analogy: The Foggy Room
Imagine two people trying to hold hands in a dark, foggy room.

1. The Fog (Uncertainty): AlphaFold knows roughly where the people are, but there is "fog" around them. The thicker the fog, the less sure the AI is about their exact position.
2. The Handshake (The Contact): To shake hands, their hands need to be within a certain distance (say, 12 inches).
3. The Calculation: Pinc looks at the "fog" around every pair of atoms (the tiny building blocks of the proteins) that could touch. It calculates the probability that they are actually touching, given the fog.

Instead of a vague score, Pinc gives you a percentage.

• If Pinc is 0.8, it means: "Based on the AI's prediction, there is an 80% chance that the proteins are actually holding hands in the way nature intended."
• It's like a Truth-O-Meter that tells you exactly how much of the predicted "handshake" is real.

Why is Pinc Special?

1. It's Honest About Small Things
Some old scores were like a magnifying glass that only worked well on huge objects. If two proteins had a tiny, delicate connection (like a specific key fitting into a tiny lock), the old scores often ignored them or said they were wrong.
Pinc is like a high-powered microscope. It is very sensitive to small, precise interactions. In the paper, they tested it on a virus protein (HIV Nef) that uses a tiny "hook" to latch onto a human protein. Pinc correctly identified this tiny, crucial connection when other scores missed it.

2. It Helps Find the "Hot Spots"
Imagine the protein interaction is a campfire. The whole fire is hot, but some logs are burning brighter than others.

• Old scores told you the whole fire was hot.
• Pinc tells you exactly which logs are the hottest.
  This helps scientists know exactly which parts of the protein to test in the lab. If you want to stop a virus from infecting a cell, you don't need to destroy the whole protein; you just need to block the "hot spot" Pinc identified.

3. It's Simple and Free
The authors wrote a simple computer script (an R script) that anyone can use. You feed it the AlphaFold prediction, and it spits out the Pinc score. No complex math degrees required.

The Bottom Line

Before this paper, using AlphaFold to find new protein interactions was like playing a game of "Guess the Score." You had to guess if a high number meant a good interaction or just a long protein.

With Pinc, the game changes. Now, scientists have a clear, intuitive probability.

• Score of 0.5? "Maybe, but it's a coin toss."
• Score of 0.9? "Very likely real; let's test this in the lab!"

This tool helps biologists stop wasting time on false leads and focus their experiments on the interactions that are most likely to be true, speeding up the discovery of new medicines and biological secrets.

Technical Summary

1. Problem Statement

While AlphaFold (specifically AlphaFold-Multimer and AlphaFold3) has become the gold standard for predicting protein-protein interaction structures, interpreting the reliability of these predictions remains challenging.

• Limitations of Existing Scores: The primary confidence metric, ipTM (interface predicted Template Modeling score), is derived from the TM-score. While useful, the TM-score has a non-trivial relationship with interaction probability (e.g., a score of 0.5 is already statistically significant under random models) and is sensitive to total chain length, particularly when non-interface regions are unstructured.
• Interpretability Gap: Other metrics like pDockQ, LIS, actifpTM, and ipSAE attempt to address specific limitations but often lack a direct, intuitive probabilistic interpretation. Setting biologically meaningful consensus thresholds for these scores is difficult.
• Need: There is a need for a scoring metric that converts structural predictions into a simple, interpretable probability of native interface contacts, particularly for small or motif-mediated interfaces where current scores may be less sensitive.

2. Methodology

The authors developed Pinc (Probability of interface native contacts), a score that translates AlphaFold's Predicted Aligned Error (PAE) matrix into conditional contact probabilities.

• Geometric Model: The method builds on a geometric model (Pellegrini et al., 2025) that maps PAE values to residue-residue contact probabilities.
  • Mechanism: It treats the PAE value for a residue pair as the radius of an "uncertainty sphere" around the aligned residue. A "contact sphere" of a fixed radius is centered on the scored residue. The contact probability is calculated as the volume of the intersection between these two spheres, normalized by the volume of the uncertainty sphere.
  • Extension to Multi-chain: The authors adapted this for multi-chain systems by replacing the standard $C_\alpha$-based distance representation with a center-of-mass representation calculated over all heavy atoms within each residue. This better reflects side-chain mediated interactions and reduces sensitivity to local noise.
• Score Definition: Pinc is defined as the mean contact probability across all interchain residue pairs whose mass centers lie within a specific contact radius.
• Implementation:
  • Developed as an R script (with a Colab notebook interface) that requires only the standard AlphaFold output files (PDB/mmCIF and JSON PAE matrix).
  • Uses a fixed contact radius of 12 Å, determined empirically to be optimal for calibration.
  • The tool is computationally efficient, utilizing R's vectorization and a compiled C backend.

3. Key Contributions

1. Probabilistic Interpretation: Pinc provides a direct probabilistic meaning: a score of $X$ implies that approximately $X$ fraction of the native interfacial contacts are expected to be present in the model.
2. Minimal Parametrization: Unlike metrics requiring full PAE distributions or distograms (which are often discarded during inference), Pinc relies only on the mean PAE matrix and coordinates, making it easy to calculate post-hoc.
3. Sensitivity to Small Interfaces: The metric is specifically noted for its sensitivity to small, motif-mediated interfaces, where global structure-based scores often struggle.
4. Residue-Level Resolution: Beyond a global score, Pinc provides per-residue contact probabilities, enabling the identification of specific "hotspot" residues for experimental validation.

4. Results

The authors validated Pinc using three benchmark datasets of protein dimers and protein-nucleic acid complexes (Dunbrack Jr, 2025; Genz et al., 2025; Peng et al., 2025).

• Calibration: Across protein dimer datasets, Pinc showed excellent calibration. The predicted contact probabilities closely matched the observed fraction of native contacts ($F_{nat}$) in experimental structures. This confirms Pinc as a conservative estimate of the true interface quality.
• Correlation with Existing Metrics: Pinc showed strong positive correlations with established scores:
  • ipTM: $\rho = 0.96$
  • ipSAE: $\rho = 0.96$
  • LIS: $\rho = 0.93$
  • DockQ: $\rho = 0.87$
• Performance on Small Interfaces:
  • Pinc showed a weak dependence on interface size ($\rho = 0.18$), outperforming pDockQ ($\rho = 0.46$) and ipSAE ($\rho = 0.30$) in terms of size bias.
  • Case Study (HIV-1 Nef:HCK): In a complex mediated by a short linear motif, Pinc assigned a high confidence score (0.84) and correctly identified key residues (P72, P75, R77) with high contact probabilities, ranking the model higher than ipTM and LIS.
• Hotspot Identification: In the FAF1:HSPA1 complex, despite global modeling inaccuracies in domain orientation, Pinc correctly identified a compact set of high-probability residues forming the binding hotspot, demonstrating utility for mutational studies.
• Protein-Nucleic Acid Limitation: For protein-nucleic acid complexes, calibration was poorer unless the predicted topology was broadly correct ($F_{nat} > 0.5$). The authors advise caution when applying Pinc to these systems without complementary quality checks.

5. Significance

• Interpretability: Pinc bridges the gap between complex structural metrics and biological intuition, allowing researchers to interpret model confidence as a direct probability of native contact formation.
• Experimental Prioritization: By providing residue-level probabilities, Pinc serves as a powerful tool for prioritizing specific interface residues for mutagenesis or experimental validation, even when the global model is imperfect.
• Complementary Tool: The authors propose Pinc not as a replacement for scores like ipTM or pDockQ, but as a complementary metric. Its simplicity, lack of chain-length bias, and sensitivity to small interfaces make it particularly valuable for screening large-scale interaction datasets and studying motif-mediated interactions.
• Accessibility: The availability of a simple R script and Colab notebook lowers the barrier to entry for researchers to apply probabilistic scoring to their AlphaFold models.