CHEMICAL COMPOUND CLASSIFICATION WITH AUTOMATICALLY MINED STRUCTURE PATTERNS

Area of Science:

• Computational Chemistry
• Machine Learning
• Data Mining

Background:

• Accurate chemical structure classification is crucial for drug discovery and understanding molecular properties.
• Existing methods for molecular similarity and classification have limitations in capturing complex structural relationships.

Purpose of the Study:

• To develop novel methods for chemical structure classification by integrating graph database mining and graph kernel functions.
• To improve the accuracy and generalizability of molecular classification using pattern-based graph similarity.

Main Methods:

• Identification of general graph patterns within chemical structure data.
• Augmentation of graph kernel functions using identified patterns to compute pairwise molecular similarity.
• Classification of chemical compounds using a Support Vector Machine (SVM) with the generated similarity matrix.

Main Results:

• The pattern-based graph similarity approach achieved classification performance comparable to, and in some cases exceeding, state-of-the-art methods.
• Highly discriminative patterns were identified, enabling generalizations about a compound's function based on its structure.
• The methods demonstrated effectiveness in classifying molecular structures.

Conclusions:

• The proposed integration of graph mining and graph kernels offers a powerful new approach to chemical structure classification.
• The pattern-based method enhances molecular similarity computation and provides insights into structure-activity relationships.
• The methodology is adaptable for general graph data, with potential applications beyond bioinformatics.