Transformer-based Approach for Predicting Chemical Compound Structures

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 154–162 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 154 Transformer-based Approach for Predicting Chemical Compound Structures Yutaro Omote Ehime University omote@ai.cs.ehime-u.ac.jp Kyoumoto Matsushita Fujitsu Laboratories, Ltd. m.kyoumoto@fujitsu.com Tomoya Iwakura Fujitsu Laboratories, Ltd. iwakura.tomoya@fujitsu.com Akihiro Tamura Doshisha University aktamura@mail.doshisha.ac.jp Takashi Ninomiya Ehime University ninomiya@cs.ehime-u.ac.jp Abstract By predicting chemical compound structures from their names, we can better comprehend chemical compounds written in text and iden- tify the same chemical compound given dif- ferent notations for database creation. Pre- vious methods have predicted the chemical compound structures from their names and represented them by Simplified Molecular In- put Line Entry System (SMILES) strings. However, these methods mainly apply hand- crafted rules, and cannot predict the struc- tures of chemical compound names not cov- ered by the rules. Instead of handcrafted rules, we propose Transformer-based models that predict SMILES strings from chemical compound names. We improve the conven- tional Transformer-based model by introduc- ing two features: (1) a loss function that con- strains the number of atoms of each element in the structure, and (2) a multi-task learning ap- proach that predicts both SMILES strings and InChI strings (another string representation of chemical compound structures). In evaluation experiments, our methods achieved higher F- measures than previous rule-based approaches (Open Parser for Systematic IUPAC Nomen- clature and two commercially used prod- ucts), and the conventional Transformer-based model. We release the dataset used in this pa- per as a benchmark for the future research1. 1 Introduction Knowledge of chemical substances is necessary for developing new materials and drugs, and for synthe- sizing products from new materials. To utilize such knowledge, researchers have created databases con- taining the physical property values of chemical substances and the interrelationships among chem- ical substances. It is thought that several billions of chemical compounds exist (Lahana, 1999; Hoffmann and 1http://aiweb.cs.ehime-u.ac.jp/ pred-chem-struct Gastreich, 2019), but only a portion of these are en- tered into chemical databases. Even PubChem2, one of the largest databases of chemical com- pounds, includes the information of only approx- imately 100 million chemical compounds. More- over, databases for chemical domains are manually maintained, which consumes much time and cost. One of the time consuming processes is the inte- gration of the same chemical compounds with dif- ferent notations. For instance, a chemical structure can be derived from partial structures which are given notational variants, or the notation can fluc- tuate for a given chemical compound (Watanabe et al., 2019). Therefore, a system that automati- cally predicts a chemical compound structure from its chemical compound names would improve the database creation procedure. Structures are most commonly predicted from their notations by rule-based conversion meth- ods (Lowe et al., 2011). Although rule-based con- version can accurately predict the structures of chemical compounds based on systematic nomen- clatures such as the International Union of Pure and Applied Chemistry (IUPAC) 3 nomenclature, it …

Content truncated to 500 words