TL;DR
This paper uses machine learning to analyze the structural dynamics of the FGFR2 tyrosine kinase domain, revealing how mutations impact its function and lead to various diseases. By integrating experimental data with predicted structures, the study provides a more holistic understanding of FGFR2's role in health and disease.
What Problem Does This Paper Address?
Receptor tyrosine kinases (RTKs), like FGFR2, are crucial for cell signaling, and their dysregulation is linked to numerous diseases, including cancers and developmental disorders. While many mutations in FGFR2 are known to cause constitutive activation and disease, current analyses often rely on static protein structures. This overlooks the dynamic nature of proteins, which is critical for understanding how mutations alter function. The paper aims to bridge this gap by characterizing the intrinsic structural dynamics of FGFR2 and connecting these dynamics to functional regions, disease types, and specific mutations, even for variants lacking experimental structural data.
Methods
The study employs a multi-pronged approach:
- Principal Component Analysis (PCA): Used to capture linear structural dynamics by identifying the principal axes of variation in the protein's structure. This method helps to understand how different parts of the protein move in concert.
- Isomap: Utilized to capture nonlinear structural dynamics, focusing on preserving geodesic distances between data points. This method is valuable for uncovering complex, non-linear relationships in protein conformation.
- Structure Prediction: Three methods were used to predict tertiary structures for FGFR2 variants lacking experimental data:
- SWISS-MODEL: A homology-based modeling tool.
- AlphaFold2: A state-of-the-art deep learning model for protein structure prediction.
- AlphaMissense: A recent method for predicting the impact of missense mutations.
- Energetic Refinement: Predicted structures were refined using SCWRL4 and NAMD2 to improve their energetic profiles and physical realism.
Datasets
- Protein Data Bank (PDB): The primary dataset consists of 29 experimentally resolved tertiary structures of the FGFR2 tyrosine kinase domain (FGFR2-TKD), including wildtype and various mutant forms in active and inactive states.
- Databases for Disease Association: The Online Mendelian Inheritance in Man (OMIM), Catalog of Somatic Mutations in Cancer (COSMIC), and Human Gene Mutation Database (HGMD) were used to link FGFR2 variants to specific diseases.
Key Insights and Ideas
- Activation Loop is Key: The study consistently highlights the activation loop as a pivotal region for FGFR2 function and regulation, with its dynamics strongly influencing the protein's active/inactive states.
- Dynamics Connect Structure to Function: By mapping structural dynamics onto functional regions, the research demonstrates how subtle structural changes can translate into altered protein function.
- FGFR2 Dynamics are Dominated by Few Modes: PCA analysis reveals that a significant portion of FGFR2's structural variation can be explained by just a few principal components, suggesting a relatively conserved set of underlying dynamic motions.
- Predictive Models Complement Experimental Data: The integration of computationally predicted structures for variants lacking experimental data significantly enriches the analysis, allowing for a more comprehensive understanding of mutation impacts.
- Disease Segregation in Structure Space: The study shows that different disease classifications (genetic disorders vs. cancers) tend to cluster in the PCA and Isomap embeddings, suggesting that structural dynamics can serve as a signature for disease.
Key Results
- PCA and Isomap Effectiveness: Both PCA and Isomap were found to be effective in characterizing FGFR2-TKD structural dynamics, with the top three PCs capturing over 81% of the structural variation.
- SWISS-MODEL Superiority for Prediction: In evaluating structure prediction methods, SWISS-MODEL demonstrated the lowest RMSD for reproducing known structures, making it the preferred choice for modeling unknown variants.
- Disease Classification Correlation: The 2D embeddings of tertiary structures, when labeled with disease classifications, showed a clear segregation between genetic disorders and cancers, underscoring the link between structural dynamics and disease.
- AlphaMissense Accuracy: AlphaMissense correctly predicted the pathogenicity of FGFR2 variants in 14 out of 16 cases (87.5%), demonstrating its utility in assessing mutation impact.
Limitations and Future Work
- Limitations of Crystal Structures: The study acknowledges that relying on crystal structures from different laboratories, potentially with varying methods and space groups, could introduce batch effects that influence structural comparisons.
- Incomplete Clinical Factor Consideration: Protein structural methods cannot fully account for all factors contributing to disease manifestation, such as whether a mutation is germline or somatic.
- Limited Sample Size for Rare Disorders: The ambiguity in classifying certain rare disorders (e.g., UCS and LADD1) is attributed to the limited sample size of available structures for these conditions.
- Future Work: The authors suggest exploring other RTKs using similar approaches to characterize structure-function-disease relationships. Further research could also focus on understanding the relative frequency of different activated states and their intermediates, as well as developing methods to predict actual pathogenicity from sequence alone. The availability of more tertiary structures in various forms and clinical settings is also crucial for improving discriminative power.
Summary updated