TL;DR
This paper introduces "reaction graph kernels," a novel method for automatically assigning Enzyme Commission (EC) numbers to unknown enzymatic reactions. By representing reactions as graphs and using graph kernels to measure similarity, the approach achieves high accuracy in predicting EC numbers, significantly improving upon existing methods for identifying enzyme functions in complex metabolic pathways.
What Problem Does This Paper Address?
Understanding plant secondary metabolism is crucial for discovering new drug candidates. However, many metabolic pathways are incomplete, with numerous enzymes whose functions are yet to be identified. Assigning EC numbers, a hierarchical classification system for enzymes and their catalyzed reactions, is key to uncovering these missing functions. The manual process of assigning EC numbers is slow and prone to errors, leading to a backlog of unannotated enzymes. Existing computational methods, like E-zyme, struggle with coverage and accurately identifying subtle differences in chemical transformations. This paper aims to bridge this gap by developing a more robust and comprehensive computational approach for EC number prediction.
Methods
The core innovation is the "reaction graph kernel." This method represents a chemical reaction as a graph where compounds are nodes and their relationships (substrate, product, cofactor, etc.) are edges. Crucially, each compound node itself can be represented as a graph, creating a "graph of graphs."
To measure the similarity between these reaction graphs, the authors employ a recursive approach using marginalized graph kernels, specifically random walk kernels. The process involves:
- Lower-level graph kernel: Computing similarity between individual chemical compounds (nodes in the reaction graph). This is done using random walks on the compound's own graph representation.
- Upper-level graph kernel: Using the compound similarities as "label matching probabilities" to compute the similarity between entire reaction graphs. This is also achieved through a recursive random walk process.
This hierarchical approach allows the kernel to capture both the similarity of individual molecules and their roles and relationships within the overall reaction. The authors also explore different "settings" for the reaction graph representation, including "full-edge" (all relationships considered), "RPAIR" (only reactant pairs), and "main-pair" (only main substrate/product pairs), to evaluate the impact of graph complexity.
Datasets
The primary dataset used for training and evaluation is the KEGG/REACTION database. This database contains information on chemical reactions and their associated EC numbers. For evaluation, the authors also utilized a subset of reactions from the KEGG "Biosynthesis of Secondary Metabolites - Reference pathway" data, specifically focusing on those with unassigned EC numbers. This subset was curated with the help of an expert from the KEGG team, resulting in 36 reactions for a blind test.
Key Insights and Ideas
- Representing reactions as graphs is powerful: The paper demonstrates that encoding chemical reactions as structured graphs, where compounds and their relationships are explicitly represented, is a highly effective way to capture the essence of a reaction for similarity comparison.
- Hierarchical graph kernels capture complex similarities: The recursive application of random walk kernels at both the compound and reaction graph levels allows for a nuanced understanding of similarity, going beyond simple molecular structure matching.
- More detailed graph representations improve prediction: Using the "full-edge" representation of reaction graphs, which includes all relationships between compounds, generally leads to better prediction accuracy compared to simplified representations like "RPAIR" or "main-pair."
- Approximate matching is key for coverage: Unlike rule-based systems that require exact matches, reaction graph kernels' ability to perform approximate matching significantly improves the coverage of predictions, especially for complex or novel reactions.
Key Results
- High accuracy on KEGG/REACTION data: The reaction graph kernel, combined with a kernel nearest neighbor approach, achieved 83% accuracy in predicting the first three digits of the EC number on the KEGG/REACTION database.
- Improved performance on unannotated reactions: In a blind test on 36 unannotated plant secondary metabolism reactions, the reaction graph kernels successfully assigned EC numbers to all 36 reactions, demonstrating superior coverage compared to E-zyme (which only answered 22 queries).
- More accurate EC number assignments: While E-zyme sometimes achieved slightly higher accuracy on the queries it could answer, the reaction graph kernels assigned more accurate EC numbers overall, especially when considering lower-level classifications (subclasses and subsubclasses). For instance, on the blind test, RGK assigned correct EC numbers for 22 EC classes, 14 EC subclasses, and 12 EC subsubclasses, whereas E-zyme assigned fewer across all levels.
Limitations and Future Work
- Handling multi-step reactions: The current method, like E-zyme, does not explicitly handle multi-step reactions or situations where a single reaction might have multiple plausible EC numbers. This is identified as an area for future research.
- Degenerated settings: While the paper explores simplified graph representations (RPAIR, main-pair), the full-edge representation generally yields better results. The authors note that additional edges in reaction graphs improve classification performance.
- Future directions: The authors suggest further research into more sophisticated graph kernel designs and exploring how to better handle the complexities of multi-step reactions and multiple EC assignments. They also highlight the potential for reaction graph kernels to be applied to other areas of bioinformatics that involve comparing structured data.
Summary updated