Physics-informed machine learning predicts binding free energy changes in protein-RNA complexes upon mutations

Introduction: Artificial intelligence for molecular biology

Protein–RNA interactions play a crucial role in gene regulation and numerous cellular processes. Understanding how mutations alter these interactions is important for biomedical research, disease investigation, and the identification of potential therapeutic targets.

Published in 2026 in Communications Biology, a study by Wang Xu, Houfang Zhang, Xinyi Zhao, Shankai Yan, Yunjie Zhao, and Yun Hui Peng introduces Pred-MutPRI, a machine-learning framework designed to predict mutation-induced changes in binding free energy in protein–RNA complexes.

The research addresses a major challenge in molecular modelling: reliably estimating how genetic changes affect the stability of biological interactions when experimental data remain limited, heterogeneous, and unevenly distributed across mutation categories.

Pred-MutPRI: Combining physics and machine learning

Predicting changes in binding free energy

Binding free energy characterizes the thermodynamic stability of an association between two molecules. Its change, denoted ΔΔG, helps quantify how a mutation affects an interaction between a protein and an RNA molecule.

This measure can help identify mutations that may strengthen or weaken molecular interactions. Accurate prediction remains difficult because experimental measurements are scarce, protocols vary, and some substitution categories are underrepresented.

Pred-MutPRI aims to overcome these limitations by integrating multiple sources of molecular information into a unified predictive framework.

Harmonizing experimental data and reducing training bias

The researchers assembled a unified dataset of experimentally measured ΔΔG values, normalized according to consistent thermodynamic conventions.

They also introduced thermodynamic permutation (PT), a procedure that generates training pairs consistent with thermodynamic relationships. This approach expands the coverage of substitution types while reducing biases caused by their uneven distribution.

The method addresses a recurring limitation in biological machine learning: models trained on imbalanced datasets may overemphasize common mutations and generalize less effectively to unfamiliar cases.

Integrating molecular structure, sequence information, and physical constraints

Pred-MutPRI incorporates several complementary feature categories:

• Conventional structural descriptors, which characterize the molecular environment of the complex.
• Weighted atomic interaction networks, which explicitly represent the local microenvironment surrounding the binding site.
• A masked entropy term derived from ESM-2, which captures context-dependent sequence constraints.
• A local effective-strain descriptor derived from AlphaFold 3, which summarizes mutation-associated structural deformation across multiple predicted mutant structures.

These features are processed by an XGBoost regressor, which estimates changes in binding free energy.

The main contribution lies in the coherent integration of experimental measurements, structural representations, and features derived from specialized predictive models rather than in the isolated use of a single algorithm.

Promising results on independent test data

Assessing generalization on a separate test set

According to the authors, Pred-MutPRI achieves a Pearson correlation coefficient of 0.705 on a blind test set that is sequence-disjoint from the training data and exhibits limited structural overlap.

This evaluation design helps assess whether the model can make predictions on examples that are less directly related to those encountered during training.

The authors also report that Pred-MutPRI outperforms existing predictors in their comparative evaluations.

The reported coefficient indicates a substantial positive association between predictions and experimental observations. However, it does not mean that 70.5% of predictions are correct. Correlation alone does not measure absolute prediction error, calibration, or the accuracy of individual estimates.

These findings are encouraging, although further validation across diverse biological systems is needed to establish the model’s domain of applicability.

AlphaFold 3: Structural information, not a substitute for thermodynamics

From three-dimensional structures to functional prediction

Advances in AlphaFold have significantly improved access to predicted three-dimensional protein structures. Associated resources allow researchers to explore extensive collections of structural models and develop hypotheses about molecular mechanisms.

AlphaFold 3 extends these capabilities to complexes involving different biomolecules, including proteins and nucleic acids. In Pred-MutPRI, structural predictions contribute to the construction of a local-strain feature associated with mutations.

A fundamental distinction remains essential: predicting a molecular structure is not equivalent to directly predicting a change in binding free energy.

A structure provides a spatial representation of a molecular system, but it does not independently determine all the thermodynamic contributions governing binding. Entropic effects, molecular rearrangements, and experimental conditions can also influence interaction stability.

Pred-MutPRI therefore demonstrates the value of incorporating structural models into a broader predictive framework rather than treating them as direct calculators of binding free energy.

Potential applications in biomedical research

Prioritizing mutations and investigating biological mechanisms

Predicting changes in binding free energy could support several areas of research:

• Gene regulation: identifying mutations that may alter protein–RNA interactions.
• Disease research: investigating molecular changes that could affect biological processes.
• Therapeutic development: prioritizing molecular interactions and targets for further experimental investigation.
• Experimental planning: ranking candidate mutations before laboratory testing and reducing some low-information exploratory experiments.

These applications remain prospective. A predicted change in binding free energy does not, by itself, demonstrate that a mutation causes disease, makes a pathogen more susceptible to the immune system, or improves treatment efficacy. Such conclusions require dedicated biological and experimental validation.

Scientific accessibility and reproducibility

The authors have made the associated dataset and Python package available, allowing the scientific community to examine the methodology and potentially reuse the resources.

• Project repository: Pred-MutPRI on GitHub.
• Scientific publication: article in Communications Biology.

The availability of these resources supports reproducibility. Reuse nevertheless requires careful consideration of training conditions, input-data quality, and the domain in which the predictions are valid.

Conclusion: Toward more robust molecular AI

Pred-MutPRI illustrates an important development in artificial intelligence for the life sciences: combining statistical learning with physical knowledge and structural information to improve predictions of complex molecular phenomena.

Its reported Pearson correlation coefficient of 0.705 on an independent test set, together with a strategy designed to reduce mutation-distribution bias, highlights the potential of this approach for investigating protein–RNA interactions.

The findings should nevertheless be interpreted carefully. Pred-MutPRI is a prediction and prioritization tool, not a replacement for thermodynamic measurements or biological validation. Its potential value lies in guiding experiments more efficiently, improving the characterization of mutation effects, and supporting biomedical research.

As experimental datasets, sequence models, and structural prediction tools continue to advance, hybrid approaches of this kind could help researchers better connect molecular changes with their functional consequences.