Mining significant substructure-substructure pairs in structural associations

Biological and chemical objects in bioinformatics are often linked to structures: proteins and DNAs to their sequences, small-molecule compounds to their chemical structures, and pathways to their network structures. Associations among these objects, such as interactions between proteins and compounds, are of particular interest in the field, and these associations are often given as a graph: the objects are represented as nodes, and the associated pairs are connected by edges. To analyze these associations through relationships between the local substructures of individual objects, we propose a general framework for exactly enumerating substructure-substructure pairs significantly overrepresented in the “associated” pairs with edges rather than in all other “control” pairs with no edges. Analyzing relationships between local information have been successfully applied to sequences, table-form data, and texts. For example, co-evolution is a key biological concept for understanding protein interactions in terms of their sequences [1], co-modules which describe coherent patterns across paired data sets are developed to analyze the gene-expression and drug-response data [2], and co-occurrence of the words is one of fundamental tools for mining biomedical texts [3]. Thus, using our framework, applying this concept to the associations between more complicated structures such as trees and graphs would also give insights to understand underlying biological problems. The proposed framework can be applied to various types of association datasets with structures (sequences, trees, and graphs): interaction between compounds and proteins, and side-effect associations between compounds. In this work, we analyze the interactions between drugs and targets, i.e., the structural associations between chemical structures of drug compounds and amino-acid sequences of target proteins.