TL;DR
This paper introduces a novel method for protein structure classification by representing protein structures as weighted networks and applying a deep neural network (DNN) to analyze these networks. This approach significantly outperforms existing methods, nearly halving the classification error and demonstrating the power of weighted graphlet measures combined with advanced deep learning architectures.
What Problem Does This Paper Address?
Predicting protein function is crucial for understanding biological mechanisms, and protein structure analysis is a key pathway to achieving this. However, computationally classifying protein structures into predefined groups (protein structure classification) remains a challenge. Existing methods often rely on either sequence-based or 3D-structure-based features. While graphlet-based measures have shown promise in capturing network topology for classification, they have traditionally been applied to unweighted networks. This paper addresses the limitation of existing graphlet-based methods by developing a novel approach that incorporates edge weights into protein structure networks and utilizes a deep neural network to effectively leverage this richer information for improved classification accuracy.
Methods
The authors propose a two-pronged approach:
- Weighted Protein Structure Networks (PSNs): They first construct an unweighted PSN where nodes represent amino acids and edges represent spatial proximity. Crucially, they then introduce edge weights. Instead of simply using spatial distance, their weight function prioritizes amino acids that are far apart in sequence but close in space, hypothesizing this indicates protein folding and is more informative for classification.
- Weighted Edge Graphlet Degree Matrix (wEGDVM): They develop a novel graphlet-based measure, wEGDVM, specifically designed for weighted networks. Unlike previous measures that summarized graphlet counts, wEGDVM is a matrix that captures the distribution of edge weights within graphlets around each edge. This matrix retains local network topology information and edge weight statistics.
- Deep Neural Network (DNN) Classifier: Because wEGDVM is a variable-size matrix, it cannot be directly used by standard classifiers. The authors propose a custom DNN architecture consisting of convolutional layers (CNNs) to capture local structural patterns and recurrent layers (bidirectional LSTMs) to integrate these patterns into a global representation. This DNN directly processes the wEGDVM matrix.
Datasets
The study evaluates their proposed method on 36 real protein domain datasets. These datasets are derived from the CATH database (9,440 protein domains) and the SCOP database (13,068 protein domains). The datasets vary in sample size (from 69 to 11,362) and the number of classes (from 2 to 33).
Key Insights and Ideas
- Weighting Matters: The core insight is that simply considering spatial proximity in protein structures is insufficient. By weighting edges based on the interplay between sequence and spatial distance, the authors capture more informative relationships, particularly those indicative of protein folding.
- Graphlets on Weighted Networks: Extending graphlet-based measures to weighted networks is a significant step. The wEGDVM measure effectively captures local network topology and the distribution of edge weights, providing a richer representation than unweighted graphlet counts.
- DNNs for Variable-Size Matrices: The development of a DNN that can directly process variable-size matrices like wEGDVM is crucial. This bypasses the information loss that can occur when converting such matrices into fixed-length vectors for traditional classifiers.
- CNNs + RNNs for Protein Structure: The combination of CNNs for local feature extraction and RNNs for sequential integration proves highly effective for protein structure classification, mirroring how local structural elements combine to form larger protein motifs.
Key Results
The proposed framework, combining weighted PSNs, wEGDVM, and the DNN classifier, demonstrates dramatic improvements in protein structure classification:
- State-of-the-Art Performance: The method significantly outperforms existing graphlet-based approaches on all 36 datasets.
- Error Reduction: Compared to the current state-of-the-art (Ordered-Graphlet-3-4), the proposed method almost halves the classification error. Specifically, the average CV error is reduced by 47% (from 0.096 to 0.051).
- wEGDVM-CC vs. Ordered-Graphlet-3-4: Even when wEGDVM is converted to a vector (wEGDVM-CC) and used with logistic regression, it outperforms EGDVM-CC by 39% on average, suggesting the weighted measure itself captures more information.
- DNN Power: The DNN classifier applied to the wEGDVM matrix shows a substantial performance boost over using logistic regression on the vector-converted wEGDVM-CC, highlighting the DNN's ability to effectively utilize the rich information in the matrix.
Limitations and Future Work
- Hyperparameter Sensitivity: While the authors found a common set of hyperparameters that worked well across all datasets, they acknowledge that dataset-specific optimization could potentially lead to further improvements. However, the consistent performance suggests their chosen hyperparameters are robust.
- Generalizability: The authors suggest that their wEGDVM measure and DNN classifier can be applied to other weighted networks in computational biology and beyond, indicating a broad potential for their framework.
- Alternative Statistics: The Cramér-von-Mises statistic used in wEGDVM could potentially be replaced by other statistics (e.g., Kolmogorov-Smirnov, Kuiper), though these showed inferior performance in initial tests.
- DNN Architecture: While the proposed DNN architecture is effective, further exploration of different CNN and RNN configurations could potentially yield even better results.
Summary updated