Physics-Informed Machine Learning Predicts Mutation Effects in Protein-RNA Complexes

Protein-RNA interactions are fundamental to gene regulation, where proteins bind to RNA molecules to control transcription, splicing, and translation. Despite their biological importance, accurately predicting how mutations in these interactions affect binding strength has remained a significant challenge. This difficulty stems from the scarcity of experimental data, the heterogeneity of measurement methods, and the complex interplay between sequence, structure, and thermodynamics.

What Happened: A New Framework for Predicting Binding Free Energy Changes

Researchers from Central China Normal University have developed Pred-MutPRI, a physics-informed machine learning framework designed to predict changes in binding free energy (ΔΔG) in protein–RNA complexes resulting from amino acid or nucleotide mutations. The study, published in Communications Biology, introduces a unified dataset of experimentally measured ΔΔG values, standardized under consistent thermodynamic conventions to ensure comparability across studies.

Central to the framework is the use of thermodynamic permutation (TP), a method that generates cycle-consistent training pairs of mutations. This approach helps balance the representation of different mutation types—such as substitutions, insertions, and deletions—while reducing bias toward common or easily measurable mutations. By creating synthetic, logically consistent pairs, the model learns more robust patterns of energy change without overfitting to rare or outlier cases.

Pred-MutPRI integrates multiple types of molecular descriptors: conventional structure-derived features such as residue contacts and secondary structure, and novel atom-level interaction networks that capture the local binding microenvironment. These networks emphasize the spatial and energetic context of each residue pair at the interface between protein and RNA.

Two key features enhance the model’s predictive power:

• A masked ESM-2 entropy term that estimates sequence-level constraints, reflecting how mutations deviate from evolutionary conservation patterns.
• An AlphaFold3-derived local effective strain descriptor that quantifies structural deformation caused by mutations across multiple predicted structures, capturing conformational flexibility and stability changes.

The model uses an XGBoost regressor, a machine learning algorithm known for its robustness in handling structured data and non-linear relationships. On a blind test set—comprising sequences and structures with no overlap with the training data—Pred-MutPRI achieves a Pearson Correlation Coefficient (PCC) of 0.705, demonstrating strong generalization performance. This result outperforms existing tools for predicting ΔΔG in protein–RNA systems.

Background: How the Model Works

Binding free energy (ΔG) is a thermodynamic quantity that reflects the stability of a molecular interaction. A negative ΔG indicates a favorable interaction, while a positive value suggests instability. The change in this energy due to a mutation—denoted as ΔΔG—is a key metric for understanding functional consequences.

Traditional methods for predicting ΔΔG rely on physical models such as molecular dynamics simulations or empirical scoring functions. However, these are computationally expensive and often fail to generalize to novel or rare mutations. Machine learning offers a faster alternative, but most models lack integration with fundamental physical principles, leading to poor performance on unseen cases.

Pred-MutPRI addresses this gap by embedding physical constraints directly into the learning process. For instance, the atom-level interaction networks are designed to reflect known physical forces—such as van der Waals interactions, hydrogen bonding, and electrostatics—within the binding interface. The model does not simply learn from data; it learns with the guidance of known physical laws.

Additionally, the use of AlphaFold3-derived structural predictions allows the model to assess how a mutation might perturb the 3D structure of the complex. Since AlphaFold3 has demonstrated high accuracy in predicting protein and RNA structures, this integration provides a reliable source of structural context for mutation analysis.

Why It Matters: Implications for Biotechnology and Medicine

Accurate prediction of mutation effects on protein–RNA binding enables researchers to anticipate functional outcomes without costly and time-consuming experiments. This capability is particularly valuable in fields such as synthetic biology, where engineered RNA elements are designed to regulate gene expression, and in disease research, where mutations in RNA-binding proteins are linked to disorders like neurodegenerative diseases and cancer.

For example, in drug development, understanding how a mutation alters binding affinity could guide the design of RNA-targeted therapeutics. Similarly, in genetic screening, such tools can prioritize mutations most likely to disrupt regulatory functions, improving diagnostic accuracy.

Moreover, Pred-MutPRI’s open-source nature and publicly available dataset make it accessible to a broad scientific community. This democratization of predictive tools supports reproducibility and collaborative research across institutions.

Limitations and Open Questions

While Pred-MutPRI represents a significant advance, several limitations remain. First, the model is trained on a dataset derived from experimental measurements, which are still sparse and unevenly distributed across mutation types. This may limit generalizability to rare or non-canonical mutations.

Second, the integration of AlphaFold3 predictions introduces dependencies on the quality of the underlying structural models. While AlphaFold3 is highly accurate, it is not perfect, and errors in predicted conformations could propagate into the final energy estimates.

Third, the model does not yet account for dynamic effects—such as conformational changes over time or environmental conditions (e.g., pH, ionic strength)—which may influence binding behavior in real biological settings.

Future work will likely focus on incorporating more dynamic and environmental factors, as well as expanding the training data to include diverse RNA types and protein families. Additionally, validation against real-world biological outcomes—such as gene expression changes or phenotypic effects—will be essential to confirm functional relevance.

What to Watch Next

As machine learning continues to evolve in biophysics, the integration of physical principles with data-driven models is expected to grow. The success of Pred-MutPRI suggests a broader trend: models that combine deep learning with known physical laws will outperform purely data-driven approaches in complex biological systems.

Researchers may soon apply similar frameworks to other biomolecular interactions—such as protein–protein or protein–DNA binding—where thermodynamic and structural complexity is equally high. The methodology could also inspire tools for predicting the effects of mutations in viral RNA, aiding pandemic preparedness and antiviral design.

For those interested in the broader implications of AI in molecular biology, readers may explore spatial tissue patterns in disease outcomes or rapid cancer biomarker detection, which illustrate how AI-driven models are transforming medical diagnostics and research.

For a deeper dive into the physics of molecular interactions, see nanoparticle catalysis in environmental chemistry, which demonstrates how physical principles guide material design.

Original source: https://www.nature.com/articles/s42003-026-10948-9

Sources & further reading

• Original source

Featured image: SEM image of a quantum chip composed of an array of 24 artificial atoms (aluminum qubits). Here, you can consider the key architectural elements of superconducting devices: transmon qubits based on sub-100nm Josephson Junctions, microwave coplanar resonators, and superconducting airbridges to equalize the electric potential of the ground over the entire area of the quantum circuit. Atoms interact with each other through the resonator and their spontaneous emission can merge into one short powerful pulse (Dicke model).
In order to create such quantum circuits, FMN Laboratory has developed unique multilayer technologies including hundreds of technological operations and bringing the parameters of these devices at the level of the best world analogues. Today, the results of this work allow dozens of leading Russian scientists to carry out complex experiments aimed at developing a Russian quantum computer. by FMNLab, CC BY 4.0, via Wikimedia Commons. Image source · License