CS 229 Project: A Machine Learning Framework for Biochemical Reaction Matching

Biochemical reaction databases capture the sum of human knowledge of biochemical reactions and chemical compounds. As the amount of data available on metabolic reactions and chemical substrates increases, the necessity of central repositories increases as well. Unfortunately, there is no established algorithm for being able to calculate the degree of similarity between reactions in different databases. A set of features have been defined and were calculated for all pairs of reactions between the Kegg and MetaCyc reaction databases. Features include reaction name match, Tanimoto coefficient of the reactions in stoichiometric vector form, enzyme identifier matching, and Enzyme Commission classification differences. Logistic regression, non-linear SVM, näıve Bayes, and decision tree learning methods were implemented, and feature selection, cross-validation, k-means clustering and other debugging methods were applied to determine how to improve the algorithms and data. We conclude that decision trees and logistic regression provide the most accurate methods, and the Tanimoto coefficient is a key feature for the performance of both learning methods.