[3K4-GS-10] AI application: text analysis

[3K4-GS-10-02] Prediction of Chemical Toxicity on Bibliometric Data

〇Takahiro Miura¹, Kimitaka Asatani¹, Yumi Miyamoto², Aori Shimizu², Ichiro Sakata¹ (1. Graduate School of Engineering, The University of Tokyo, 2. Daikin Industries, Ltd.)

Keywords:Chemical Hazard, Network Embedding, Representation Learning

Based on the concept of Sustainable Development Goals, regulations on chemical substances have been developed due to their adverse effects on the environment and human health. Despite speeding up the assessment cycle from recognizing the chemical substance's toxicity to its regulation thanks to data-driven approaches like QSAR, companies and universities are required to quickly and comprehensively detect even the slightest toxicity risk to minimize the chemical threat on our lives. This study proposes a machine learning pipeline to predict chemical toxicity from bibliometric data of toxicity-related papers in academic communities. The proposed method is based on the assumption that researchers' assessment of chemical hazards provokes public concern on regulations. Results show that both citation and text embeddings can predict toxicity slightly better for citation. This method enables the comprehensive extraction of risky substances and can be utilized for regulatory prediction.