Information Extraction of Compound-Protein Interaction from Scientific Paper using Machine Learning

Abstract

Drug Target Interaction (DTI) is an important process in drug discovery that aims to identify useful compounds in treatment. DTI research is mostly found in databases and literature or papers. To obtain DTI information, another method such as information extraction is required to retrieve information related to DTI interactions. The information in the abstract of the research paper contains many compound sentences. This study performs regular expressions to identify compound sentences, text mining for information extraction, and classification using Bernoulli Naive Bayes. The research uses a collection of abstract documents, where 3.000 abstract documents will be arranged into 29.363 sentences. Sentences that the regular expression has parsed are matched using pattern matching and conducted by text pre-processing. Sentences resulting from text pre-processing stages are used as training datasets. We use 10- fold cross-validation to evaluate the model. This research obtained the best average accuracy value of 0.72 for using naive Bayes without regular expression for compound sentences and 0.76 accuracies for naive Bayes with a regular expression for single sentences. Furthermore, by applying the feature selection process for compound sentence data, we obtained an accuracy of 0.731 for the model without regular expressions and an accuracy of 0.7644 for the model with feature selection using regular expressions.