안전한 화학제품 설계를 위한 딥러닝 기반 화학물질 기능 및 용도 예측모델 개발

• 영문명: Development of a Deep Learning-Based Model for Predicting Functions and Applications of Chemicals to Support Safe-by-Design Chemical Products
• 발행기관: 한국환경보건학회
• 저자명: 박종서(Jongseo Park) 이수진(Sujin Lee) 김선미(Sunmi Kim) 강경희(Kyounghee Kang) 한상문(Sangmoon Han) 서명원(Myungwon Seo)
• 간행물 정보: 『한국환경보건학회지』제51권 제2호, 62~75쪽, 전체 14쪽
• 주제분류: 공학 > 환경공학
• 파일형태: PDF
• 발행일자: 2025.04.30

국문 초록

Background: The variety of chemicals used in chemical products and industries and their applications have expanded with the advancement of science and technology. Since chemicals have different human exposure routes depending on their applications, understanding the function and application of chemicals is essential for safety and risk assessments. However, identifying the diverse functions of chemicals through experiments can be challenging due to the large number of chemicals involved. Objectives: This study aims to develop a deep learning-based predictive model for estimating the functions and applications of chemicals in products. Methods: We trained the model using the US EPA CPDat database, which includes data on various chemicals and their uses. The model’s performance was evaluated using both data-based and product-based validation. The model was trained and tested on curated datasets in data-based validation. For the product-based validation, a real-world chemical product was selected and the predicted functional uses of its constituent chemicals were compared to their actual applications. Results: In the data-based validation, the model achieved an AUC-ROC of 0.95 and an average F1-score of 0.68 across all functional uses. In a product-based validation focusing on a bleaching product, the model accurately predicted the functional uses of 87.5% of its constituent chemicals. The predicted functional uses were also mapped to chemical product categories, demonstrating the model’s applicability in chemical product development processes. Conclusions: The model showed strong predictive performance in both data-based and product-based validation, highlighting its potential to predict the functions and applications of chemicals. The developed model will help with the identification and design of chemicals with specific functions and applications and can be used in the Safe-by-Design (SbD) approach to developing chemical products.

키워드

Chemical function chemical product deep learning graph convolution network in silico