Investigation and identification of protein γ-glutamyl carboxylation sites

Tzong-Yi Lee, Cheng-Tsung Lu, Shu-An Chen, Neil Arvin Bretaña, Tzu-Hsiu Cheng, Min-Gang Su, Kai-Yao HuangBMC Bioinformatics 2011 · DOI

TL;DR

This paper tackles the challenge of identifying specific sites on proteins that undergo γ-glutamyl carboxylation, a crucial post-translational modification for blood clotting and bone health. The researchers developed a computational model using machine learning that combines amino acid sequence, solvent accessibility, and secondary structure information to accurately predict these sites, significantly reducing the need for laborious experimental validation.

What Problem Does This Paper Address?

γ-glutamyl carboxylation is a vital post-translational modification (PTM) where glutamate (Glu) residues are converted to γ-carboxyglutamate (Gla) residues. This process is essential for the function of proteins involved in blood clotting (like factors II, VII, IX, and X), bone growth (osteocalcin, matrix Gla protein), and extraosseous calcification. Vitamin K acts as a critical cofactor in this reaction, which occurs in the endoplasmic reticulum. While the importance of Gla residues is well-established, experimentally identifying which specific Glu residues are carboxylated has been a significant bottleneck due to the difficulty and time commitment involved. This limitation hinders a deeper understanding of substrate site specificity for the γ-glutamyl carboxylase enzyme. This paper aims to bridge this gap by developing an in silico method to predict these carboxylation sites, thereby facilitating preliminary analysis and reducing the experimental burden.

Methods

The core of this study is the development of a predictive model using a Support Vector Machine (SVM). The researchers explored several features to characterize potential carboxylation sites:

  • Amino Acid Composition (AAC): This captures the relative frequencies of the 20 amino acids within a defined window around a Glu residue.
  • Positional Weighted Matrix (PWM): This goes a step further than AAC by considering the specific frequency of each amino acid at each position within the window, capturing position-specific conservation.
  • Solvent-Accessible Surface Area (ASA): This feature reflects how exposed an amino acid residue is to the solvent, as PTMs often occur on surface-exposed residues. Since experimental tertiary structures were often unavailable, the RVP-Net tool was used to predict ASA values from amino acid sequences.
  • Secondary Structure (SS): This feature describes whether an amino acid is part of a helix, sheet, or coil. The PSIPRED tool was employed for predicting secondary structures.

To train and evaluate the SVM model, a window of amino acids (2n+1 residues) centered on the Glu residue was used. The optimal window size was determined to be 15 residues. To prevent overestimation of performance due to homologous sequences, a rigorous process of removing similar sequences was applied to both training and testing datasets. The model's performance was assessed using k-fold cross-validation (specifically, 5-fold) and then validated on an independent test set that was not used during cross-validation. Performance metrics included precision, sensitivity, specificity, accuracy, and the Matthews Correlation Coefficient (MCC).

Datasets

The study utilized a comprehensive PTM resource, dbPTM, which aggregates data from UniProtKB (release 15.0) and HPRD (release 8.0).

  • Initial Dataset: Contained 1123 Gla residues from 182 protein entries across multiple organisms.
  • Filtered Dataset: Non-experimental sites (annotated as "by similarity," "potential," or "probable") were removed, resulting in 463 experimental carboxylation sites from 134 carboxylated proteins.
  • Positive Set: All experimentally confirmed carboxylated Glu residues.
  • Negative Set: Glu residues within the same experimental carboxylated proteins that were not annotated as carboxylation sites. This resulted in 854 non-carboxylated Glu sites.
  • Training Data (non-homologous): 302 carboxylated Glu residues and 567 non-carboxylated Glu residues from 79 proteins.
  • Independent Testing Data (non-homologous): 60 carboxylated Glu residues and 99 non-carboxylated Glu residues from 14 proteins. This set was strictly held out from the training and cross-validation process.

Key Insights and Ideas

  • Amino Acid Composition is Key: The study strongly suggests that the local amino acid environment around a Glu residue is a primary determinant of whether it will be carboxylated. The high conservation of glutamate itself and the presence of arginine were notable findings.
  • Structural Context Matters: Beyond just the sequence, the accessibility of the residue to the solvent (ASA) and its local secondary structure (helix, coil) provide crucial information for prediction. Carboxylated sites tend to be more exposed and are often found in coil or helix regions.
  • Feature Combination is Powerful: While individual features showed some predictive power, combining positional weighted matrix (AA_PWM), amino acid composition (AAC), and accessible surface area (ASA) yielded the highest accuracy, demonstrating the synergistic benefit of integrating diverse information.
  • Predictive Model Reduces Experimental Burden: The developed SVM model can accurately differentiate between carboxylation and non-carboxylation sites, significantly narrowing down the candidates for experimental validation and saving considerable time and resources.

Key Results

  • Optimal Window Size: A 15-mer window (7 residues on each side of the Glu) was found to be optimal for predictive performance.
  • Best Performing Features: The combination of AA_PWM, AAC, and ASA achieved the highest accuracy of 0.892 during 5-fold cross-validation.
  • Independent Test Performance: The best model (AA_PWM + AAC + ASA) achieved an accuracy of 0.823 on the independent test set, demonstrating its robustness and ability to generalize to unseen data. Precision was 0.735, sensitivity 0.833, and specificity 0.818.
  • Functional Group Categorization: Analysis of functional domains in carboxylated proteins revealed three main groups: Gla domain, peptidase, and EGF-like proteins.

Limitations and Future Work

  • Homology and Evolutionary Conservation: While homologous sequences were removed from training/testing, the authors suggest further investigation into the evolutionary conservation of carboxylation sites, especially in homologous proteins.
  • Structural Details: The study acknowledges that while ASA and secondary structure were used, a deeper exploration of other structural features (e.g., B-factor, protein linker regions, intrinsic disorder) in PDB entries could be beneficial.
  • Biological Function: The authors propose investigating the biological functions of carboxylated proteins more deeply, potentially through gene ontology analysis, identifying other PTMs, and analyzing protein-protein interaction networks.
  • Data Limitations: The study relies on experimentally verified sites, and the availability of such data can be a limiting factor. The authors also note that some carboxylated proteins lack corresponding PDB tertiary structures, necessitating the use of prediction tools for ASA and SS.

Summary updated

BibTeX
@article{lee2011investigation,
  author = {Tzong-Yi Lee and Cheng-Tsung Lu and Shu-An Chen and Neil Arvin Bretaña and Tzu-Hsiu Cheng and Min-Gang Su and Kai-Yao Huang},
  title = {{Investigation and identification of protein γ-glutamyl carboxylation sites}},
  journal = {BMC Bioinformatics},
  year = {2011},
  doi = {10.1186/1471-2105-12-s13-s10},
}
Figure