💻 bioinformatics

Predicting single mutation effects on binding affinity at protein-protein interfaces based on MMPBSA calculations

This paper presents a user-friendly, physics-based workflow using MMPBSA calculations to accurately predict the effects of single amino acid substitutions on protein-protein and protein-peptide binding affinity, demonstrating performance that matches or exceeds established methods across diverse targets including SARS-CoV-2 spike protein complexes.

Original authors: Noske, J., Janzen, M., Lepoivre, T., Höcker, B.

Published 2026-09-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Noske, J., Janzen, M., Lepoivre, T., Höcker, B.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Proteins are the workhorses of life, tiny molecular machines that fold into specific shapes to perform tasks, from carrying oxygen in the blood to fighting off infections. Often, these proteins must find and lock onto other proteins to function, a process driven by the precise fit of their surfaces. When a single building block, or amino acid, in one of these proteins changes, it can be like swapping a single brick in a wall; sometimes the structure holds firm, but other times the entire connection weakens or fails. Predicting exactly how a single change will alter the strength of this bond is a major challenge for scientists. If they could do this reliably, they could design better medicines, engineer enzymes to clean up pollution, or understand why certain genetic mutations cause disease. For years, researchers have tried to build computer programs that can calculate these changes, but the results have often been hit or miss, struggling to balance speed with accuracy.

A team of researchers at the University of Bayreuth in Germany has developed a new way to make these predictions, offering a tool that is both fast and surprisingly accurate. Their approach relies on a method called MMPBSA, which stands for Molecular Mechanics Poisson-Boltzmann Surface Area. In plain terms, this is a physics-based calculation that estimates how much energy is required to hold two proteins together. The team realized that while existing computer programs could guess the energy of a protein, they often missed the subtle shifts caused by a single mutation. To fix this, the researchers took a massive collection of experimental data—over 1,600 recorded instances where scientists had measured how a single amino acid change affected protein binding—and used it to train their new scoring system. They did not rely on complex artificial intelligence that learns patterns from data alone; instead, they built a straightforward mathematical model that weighs different physical forces, such as how atoms attract or repel each other and how water molecules interact with the protein surface.

The researchers tested their new method against two established competitors: a widely used software called FoldX, which relies on empirical rules, and a cutting-edge deep learning tool called DDMut-PPI, which uses artificial intelligence. They first ran their calculations on the same dataset used to train the artificial intelligence model. Here, the deep learning tool performed exceptionally well, correctly predicting the outcome for most mutations. However, this high score was expected, as the artificial intelligence had been trained specifically on those numbers. The true test came when the team applied all three methods to entirely new scenarios where the computer models had never seen the data before. In one test, they looked at mutations in the spike protein of the virus that causes COVID-19, specifically how it binds to human cells. In another, they examined engineered proteins designed to grab onto specific peptide chains.

In these new, unseen situations, the results shifted dramatically. The deep learning tool, which had excelled on its training data, struggled to rank the mutations correctly, showing a weak ability to distinguish between beneficial and harmful changes. The traditional software, FoldX, produced many large errors, often predicting that a mutation would have a massive effect when it had none, or vice versa. In contrast, the new MMPBSA-based method maintained a steady performance. It successfully predicted the strength of the bonds in both the viral spike protein and the engineered proteins, matching or exceeding the accuracy of the other tools. The researchers found that their method was particularly good at handling the complex physics of charged particles and water interactions, which are often the source of errors in other programs. They also discovered that including a term for the stability of the individual protein pieces, rather than just the bond between them, was crucial for getting the right answer.

The study suggests that while artificial intelligence is powerful, it can sometimes become too specialized on the data it learns from, failing to adapt when faced with new biological puzzles. The physics-based approach, by relying on fundamental laws of nature rather than memorized patterns, proved more robust when applied to different types of proteins and structures. The team made their workflow available to the public, allowing other scientists to use this tool to prioritize which mutations to test in the lab. By filtering out the most promising candidates before running expensive experiments, this method could speed up the development of new therapies and the engineering of better biological tools. The work demonstrates that sometimes, returning to first principles and carefully weighing the physical forces at play can yield more reliable results than the most complex algorithms, especially when the goal is to understand how a tiny change in a molecule can ripple out to change the behavior of a living system.

Problem Statement
Accurately predicting the effects of single amino acid mutations on protein-protein and protein-peptide binding affinity remains a central challenge in computational biophysics and protein engineering. While deep learning-based scoring functions have achieved impressive performance on their training sets, they often suffer from poor generalizability to unseen systems. Conversely, established physics-based protein design tools (e.g., Rosetta, FoldX) often lack the accuracy required to resolve small energetic differences at the single-residue level, frequently due to simplified energy terms that fail to fully capture mutation-dependent changes in electrostatics and solvation. There is a continued need for interpretable, physics-based methods that can robustly predict mutational effects across diverse protein systems without relying on extensive training data.

Methodology
The authors present a physics-based scoring function derived from Molecular Mechanics Poisson-Boltzmann Surface Area (MMPBSA) calculations to predict mutation-induced changes in binding affinity ( $\Delta\Delta G_{binding}$ ).

Training Data: The model was fitted using a filtered subset of the SKEMPI v2.0 database, comprising 1,662 data points from 175 wildtype crystal structures. Strict filtering criteria were applied, retaining only interface mutations (core, rim, support regions) in crystal structures with resolution better than 2.5 Å, while excluding glycine and proline mutations.
Feature Engineering: The scoring function utilizes a linear regression model combining several energy terms:
- Van der Waals (VdW) energy.
- Electrostatic energy (sum of Coulomb and polar solvation).
- Nonpolar solvation energy (estimated via buried solvent-accessible surface area).
- Folding stability change ( $\Delta\Delta G_{folding}$ ), calculated using FoldX.
Model Formulation: A least-squares linear regression was used to determine weights for these terms. The final equation is:
$\Delta\Delta G_{binding} = w_0 + w_1 \Delta\Delta G_{VdW} + w_2 (\Delta\Delta G_{Coulomb} + \Delta\Delta G_{PolarSolv}) + w_3 \Delta\Delta G_{NonpolarSolv} + w_4 \Delta\Delta G_{Folding}$
Notably, the folding stability term was explicitly included, as the analysis indicated it significantly contributes to binding affinity changes, a factor often omitted in standard MMPBSA calculations where intramolecular energies are assumed to cancel out.
Structural Modeling: For all calculations, structures were prepared by removing waters and non-protein molecules. Mutant models were generated using FoldX's BuildModel command, followed by energy minimization using OpenMM with the AMBER ff19SB force field and OBC2 implicit solvent. Protonation states were assigned based on hydrogen-bonding potential (histidine) or standard states (deprotonated carboxyls, protonated amines).
Evaluation: The method was benchmarked against FoldX (v5.1) and the deep learning-based DDMut-PPI. Performance was assessed using Pearson correlation coefficients and Root Mean Squared Error (RMSE) on three distinct datasets:
1. The training-related SKEMPI v2.0 subset.
2. An independent SARS-CoV-2 Spike RBD-ACE2 deep mutational scanning dataset (272 points).
3. A computationally predicted dataset of designed Armadillo Repeat Proteins (dArmRPs) binding to peptides (63 points), where structures were generated using Chai-1.

Key Results

SKEMPI v2.0 Benchmark: The MMPBSA-based method achieved a Pearson correlation of 0.59 and an RMSE of 1.69 kcal/mol, outperforming FoldX (0.47 correlation, 1.98 RMSE). While the deep learning method DDMut-PPI showed superior performance on this dataset (0.92 correlation, 0.82 RMSE), the authors note this is likely due to DDMut-PPI being trained on the same data.
SARS-CoV-2 Benchmark (Generalizability): On the independent SARS-CoV-2 dataset, the MMPBSA method maintained robust performance with a correlation of 0.51 and RMSE of 0.97 kcal/mol. In contrast, FoldX yielded a lower correlation (0.40) and higher RMSE (1.61). DDMut-PPI showed a significant drop in performance, with a correlation of only 0.22, indicating limited ability to rank variants outside its training domain.
dArmRP Benchmark: On the computationally predicted dArmRP dataset, all three methods achieved correlations around 0.6. The MMPBSA method and FoldX achieved ~0.65 correlation. However, the MMPBSA method produced substantially fewer large outliers than FoldX, resulting in a significantly lower RMSE (0.73 vs. 1.56 kcal/mol). DDMut-PPI performed better here than on the SARS-CoV-2 set (0.6 correlation) but still lagged behind the physics-based methods in terms of error magnitude.
Error Analysis: The model showed higher prediction errors for charged residues (lysine, aspartic acid, glutamic acid) and hydrogen-bonding sidechains, suggesting limitations in modeling sidechain conformations and protonation states.

Significance and Claims
The paper claims that the proposed MMPBSA-derived scoring function provides a computationally efficient and interpretable framework for prioritizing interface mutations. Its primary significance lies in its transferability: unlike the deep learning method which overfitted to the training data, the physics-based approach demonstrated consistent performance across diverse targets (viral proteins, designed enzymes) and structure types (experimental crystal structures and computationally predicted models).

The authors emphasize that the inclusion of the folding stability term is a critical contribution, highlighting that binding effects are heavily influenced by monomer stability changes. Furthermore, the method achieves this robustness without relying on computationally expensive molecular dynamics ensembles; instead, it utilizes minimized structures, offering a practical compromise between accuracy and efficiency for protein engineering and binding affinity optimization. The work suggests that while deep learning offers high accuracy on known data, physics-based approaches remain superior for generalizable predictions in protein design tasks involving unseen systems.

More like this