Identifying glycan motifs using a novel subtree mining approach

Lachlan Coff, Jeffrey Chan, Paul A. Ramsland, Andrew J. GuyBMC Bioinformatics 2020 · DOI

TL;DR

This paper introduces CCARL, a new computational method for identifying specific sugar structures (glycan motifs) that bind to proteins. By enhancing a common "subtree mining" technique with a novel way to represent "missing" connections in glycans, CCARL can pinpoint terminal motifs more effectively, leading to better predictions of protein-glycan interactions and aiding in the discovery of new biological insights.

What Problem Does This Paper Address?

Glycans, complex sugar chains, are fundamental to many biological processes, including how cells interact with each other and with pathogens. Understanding which specific glycan structures (motifs) bind to proteins like lectins and antibodies is crucial for deciphering these interactions and has implications for understanding diseases like cancer and infections. While previous methods have used "frequent subtree mining" to find common glycan patterns, they often struggle to identify motifs that are important specifically at the "ends" (non-reducing termini) of glycans, which are frequently the key binding sites. This limitation hinders the precise characterization of glycan-binding specificities and the discovery of novel binding determinants.

Methods

The core of this work is an enhanced frequent subtree mining approach called Carbohydrate Classification Accounting for Restricted Linkages (CCARL). The key innovation lies in how glycans are represented. Instead of just showing existing connections, CCARL introduces "restricted linkage nodes" to explicitly mark where a connection could exist but doesn't. This allows the method to distinguish between similar glycan structures where the difference lies in a missing linkage at a specific position, particularly at the terminal end.

The workflow involves:

  1. Representing Glycans: Glycans are converted into directed graphs, with added "restricted linkage nodes" to denote absent connections.
  2. Frequent Subtree Mining: A modified gSpan algorithm (gBolt) is used to find common subtrees (potential motifs) that appear frequently in the dataset, with specific thresholds for "support" (frequency).
  3. Feature Selection (mRMR): To reduce the large number of potential motifs to a manageable set of informative ones, the Minimum-Redundancy, Maximum-Relevance (mRMR) algorithm is employed. This selects motifs that are highly relevant to binding and not redundant with each other.
  4. Motif Augmentation: Potential motifs are further refined by adding information about the type of glycosidic linkage at their reducing end (anomeric descriptor).
  5. Further Feature Selection (L1 Regularization): Logistic regression with L1 regularization is used to further select the most predictive motifs and encourage sparsity.
  6. Classification: A final logistic regression model is trained to predict glycan binding.

Datasets

The study primarily uses glycan microarray data from the Consortium for Functional Glycomics (CFG). This publicly available dataset contains extensive information on how various proteins (lectins) bind to a wide range of glycans. The data was downloaded and pre-processed for analysis. No new benchmark datasets were introduced; instead, the paper leverages existing, well-characterized glycan microarray experiments.

Key Insights and Ideas

  • The Importance of "Missing" Information: The most significant insight is that explicitly representing absent connections (restricted linkages) in glycan structures is crucial for identifying motifs that are specific to terminal positions. This allows for finer discrimination between glycans that might otherwise appear similar.
  • Terminal Motifs are Key: The paper strongly supports the idea that many glycan-binding proteins recognize motifs specifically at the non-reducing terminus of a glycan. CCARL's ability to identify these terminal motifs is a major advantage.
  • MAD for Robust Outlier Detection: Using Median Absolute Deviation (MAD) for identifying positive binding glycans is shown to be more robust than traditional z-score methods, especially when dealing with datasets that have a large number of non-binders or a few strong binders. This leads to a more accurate initial classification of binding vs. non-binding glycans.
  • Motif Specificity vs. Size: CCARL tends to identify smaller, more specific motifs compared to other tools like Glycan Miner Tool, which often return larger, more generalized patterns. This specificity can be more biologically relevant.

Key Results

  • High Classification Performance: CCARL achieved excellent classification performance across a range of lectins, with a median Area Under the Curve (AUC) of 0.887. For many lectins, AUC values were above 0.9, indicating high predictive accuracy.
  • Identification of Known Motifs: The identified motifs were largely consistent with experimentally determined binding specificities for well-characterized lectins like Peanut Agglutinin (PNA), Concanavalin A (Con A), and Ricinus communis agglutinin I (RCA I).
  • Comparison to Other Tools: CCARL outperformed existing tools like GLYMMR, Glycan Miner Tool, and MotifFinder in terms of median AUC across multiple datasets, demonstrating its superior predictive power and motif discovery capabilities.
  • Novel Motif Discovery Potential: The method successfully identified motifs that were not captured by other tools, suggesting its potential for discovering novel binding determinants.

Limitations and Future Work

  • Steric Hindrance and Linker Effects: The authors acknowledge that subtree mining, even with restricted linkages, has limitations in capturing steric hindrance effects or the impact of linker molecules used in microarray fabrication. These factors can influence glycan presentation and binding, which CCARL doesn't fully model.
  • Intermediate Binders: The study excluded "intermediate binders" to avoid Type I errors. The authors suggest that these could be a valuable source for identifying secondary motifs, but caution is needed.
  • Spatial Environment: Future work could explore methods that account for the spatial environment of potential motifs, which might improve predictions for complex glycans where steric factors are significant.
  • Organism-Specific Glycomes: The CFG dataset is biased towards mammalian glycans. For analyzing glycans from other organisms, training data specific to that organism's glycome would be beneficial.
  • Experimental Validation: While CCARL aids in identifying candidate motifs, experimental validation (e.g., using affinity chromatography or crystallography) remains essential to confirm their biological significance.

Summary updated

BibTeX
@article{coff2020identifying,
  author = {Lachlan Coff and Jeffrey Chan and Paul A. Ramsland and Andrew J. Guy},
  title = {{Identifying glycan motifs using a novel subtree mining approach}},
  journal = {BMC Bioinformatics},
  year = {2020},
  doi = {10.1186/s12859-020-3374-4},
}
Figure