Prediction of heme binding residues from protein sequences with integrative sequence profiles

Yi Xiong, Juan Liu, Wen Zhang, Tao ZengProteome 2012 · DOI

TL;DR

This paper introduces a novel method for predicting heme-binding residues in proteins by integrating evolutionary information (PSSM) with physicochemical properties. This "PSSMPP" approach significantly improves prediction accuracy compared to existing methods, offering a more efficient and effective way to identify crucial protein regions involved in heme interactions.

What Problem Does This Paper Address?

Heme-protein interactions are fundamental to a vast array of biological processes, including electron transfer, catalysis, and signal transduction. Identifying the specific amino acid residues within a protein that bind to heme is crucial for understanding these functions and for annotating protein roles. However, experimentally determining these binding sites is a time-consuming and labor-intensive process. While computational methods exist, such as the HemeBIND tool, they often rely on structural information, which is not available for all proteins. Existing sequence-based methods, primarily using Position Specific Scoring Matrices (PSSMs), have limitations. This paper aims to bridge this gap by developing a more accurate and efficient sequence-based prediction method that leverages both evolutionary and intrinsic physicochemical properties of amino acids.

Methods

The core of this work is the development of an "integrative sequence profile" called PSSMPP. This method combines two types of information:

  • Evolutionary Information (PSSM): This is derived from Position-Specific Iterated BLAST (PSI-BLAST) searches against a non-redundant protein database. PSSMs capture the evolutionary conservation of amino acids at each position in a protein sequence, indicating their importance.
  • Physicochemical Properties (PP): These are intrinsic properties of amino acids (e.g., hydrophobicity, charge, size) that influence protein structure and function. The authors utilized the AAindex database, which contains 544 such indices, and selected a subset of 531 after removing those with missing values.

Instead of simply concatenating PSSM and PP features (which can lead to very high-dimensional and potentially redundant feature vectors), the authors propose a "condensed integrative profile" (PSSMPP). This involves weighting the amino acid columns of the PSSM by the corresponding physicochemical property values and then summing them up. This creates a more compact feature representation.

To select the most informative physicochemical properties, a novel feature selection scheme was designed. This scheme combines a greedy strategy with correlation analysis. First, individual physicochemical properties were ranked by their predictive power using the Area Under the Curve (AUC) metric. Then, a greedy approach iteratively selected properties, ensuring that newly added properties were not highly correlated (Pearson correlation coefficient < 0.6) with already selected ones and that they improved the overall prediction performance. Support Vector Machines (SVMs) with a radial basis function kernel were used as the classification algorithm.

Datasets

The study utilized two datasets of heme proteins:

  • PHeme-75 (Training Set): Consists of 75 heme protein chains, derived from a larger set of 89 proteins. This set was constructed from proteins identified before March 2007.
  • PHeme-72 (Testing Set): A non-redundant set of 72 heme protein chains collected after March 2007. This set ensures that the model is evaluated on proteins not seen during training and shares no more than 30% sequence identity with the training set.

Heme binding and non-binding residues were assigned using the Ligand Protein Contact server.

Key Insights and Ideas

  • The Power of Integration: The study strongly suggests that combining evolutionary information (PSSM) with intrinsic physicochemical properties (PP) is more powerful than using either alone for predicting heme-binding residues.
  • Compact Representation Matters: Simply concatenating features leads to high dimensionality and redundancy. The proposed PSSMPP method, by integrating and condensing information, achieves better performance with a significantly smaller feature set. This is crucial for computational efficiency and potentially better generalization.
  • Physicochemical Properties are Informative: While PSSMs capture evolutionary constraints, physicochemical properties provide a different, complementary perspective on why certain residues are involved in binding. The selected properties (related to helix propensity and linker regions) offer insights into the local structural context of heme-binding sites.
  • Feature Selection is Key: Not all physicochemical properties are equally useful. The developed feature selection strategy effectively identifies a small, informative subset of properties, reducing noise and improving model performance.

Key Results

  • The proposed PSSMPP method consistently outperformed conventional methods (PSSM alone, and PSSM + PP concatenation) on both 5-fold cross-validation and independent tests.
  • On the PHeme-75 dataset, PSSMPP achieved an AUC of 0.780, compared to 0.762 for PSSM+PP and 0.767 for PSSM.
  • On the independent test set (PHeme-72), PSSMPP achieved an AUC of 0.790, outperforming binary classification (0.718) and PSSM-based methods (0.768).
  • The optimal window size for feature extraction was found to be 17.
  • The PSSMPP method uses a significantly reduced feature vector size (85 dimensions for a window size of 17) compared to PSSM (357 dimensions) and PSSM+PP (425 dimensions), demonstrating improved efficiency.

Limitations and Future Work

  • Window Size Optimization: While a window size of 17 was found to be optimal for the PSSMPP method, the paper notes that different window sizes might be optimal for other types of functional residues.
  • Generalizability: The authors suggest that the integrative sequence profile feature developed in this study could potentially be expanded to predict other functional residues on proteins, implying that the approach is not limited to heme-binding sites.
  • Comparison with HemeBIND: While the paper mentions HemeBIND, a direct quantitative comparison on identical datasets and metrics is not fully detailed, though the PSSMPP method is presented as superior to HemeBIND's sequence-based classifier. Further validation against HemeBIND's structure-based method would be valuable.

Summary updated

Abstract

The novel feature of an integrative sequence profile achieves good performance using a reduced set of feature vector elements.
BibTeX
@inproceedings{xiong2012prediction,
  author = {Yi Xiong and Juan Liu and Wen Zhang and Tao Zeng},
  title = {{Prediction of heme binding residues from protein sequences with integrative sequence profiles}},
  booktitle = {Proteome},
  year = {2012},
  doi = {10.1186/1477-5956-10-s1-s20},
}
Figure