Collecting of chemical information a with the data on biological activity is essential for further preparation of computational models and mixing/matching biological and physical-chemical data of chemical compounds. Along with databases of chemical compounds text and data mining approaches can be used for creating collections of chemical named entities (CNEs) and specific biological activity types. This information can be used for identification of associations between chemical structures of CNEs and biological activity data.<br/>In our study, we aim at development and testing consensus models for extraction of chemical named entities associated with biological activity data based on Naïve Bayes (NB) approach, conditional random fields (CRF) and deep neural networks (DNN) combined with dictionary-based approach. Both NB and CRF approaches require annotated corpus of relevant texts. CHEMDNER, DrugProt, CDR corpora were used for extracting CNEs, names of proteins and pathology. Naïve Bayes approach is based on the representation of CNEs as a set of text fragments (FoT) split into short multi-n-grams (1-5 symbols) followed by further calculation of n-gram frequency resulting in probabilities for certain FoT to belong to CNE. Mean invariant accuracy obtained using five-fold cross validation of CNEs recognition using NB approach reaches 0.96. CRF algorithm was used for context-based extraction of CNEs and allows recognition of CNEs with balanced accuracy over 0.88 (five-fold cross validation procedure). We evaluated capabilities of improving accuracy through application of consensus CRF-NB models along with automatically updated dictionaries of chemical named entities. Preliminary results demonstrated that usage of dictionaries with either NB or CRF is preferable for building recognition models with high prediction accuracy comparing to consensus models. Also, we investigated the possibilities of extracting data on potential biological activity of chemical compound based on the associations between chemical named entities and names of proteins, genes and pathologies, including diseases. The associations are extracted based on the dictionaries containing templates indicating the particular association with usage of DNN; the accuracy calculated using test set of 0.87-0.89. These results demonstrate the applicability of machine learning based data and text mining approaches for building chemical libraries with biological activity data associated with chemicals.

Text and data mining approach for creating large chemical libraries with biological activity data

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.