잠재적인 피부과민성 유발물질 선별을 위한 딥러닝 기반 KeratinoSens™ 활성 예측모델 개발

• 영문명: Development of a Deep Learning-Based Model to Predict KeratinoSens™ Activity for Screening Potential Skin Sensitizers
• 발행기관: 한국환경보건학회
• 저자명: 이수진(Sujin Lee) 박종서(Jongseo Park) 강경희(Kyounghee Kang) 한상문(Sangmoon Han) 김선미(Sunmi Kim) 서명원(Myungwon Seo)
• 간행물 정보: 『한국환경보건학회지』제51권 제3호, 137~148쪽, 전체 12쪽
• 주제분류: 공학 > 환경공학
• 파일형태: PDF
• 발행일자: 2025.06.30

국문 초록

Background: Skin sensitization is a common dermatological condition, and it is crucial to effectively evaluate potential sensitizers. Comprising four key events, the adverse outcome pathway (AOP) framework used by the OECD systematically describes the mechanism of skin sensitization. With the global shift away from animal testing, alternative testing methods have gained increasing importance. Among them, KeratinoSensTM is considered a key in vitro assay for evaluating KE2, the activation of keratinocytes. However, it is impractical to assess every chemical substance experimentally, which highlights the need for efficient in silico prediction approaches. Objectives: This study aims to analyze methodologies for developing predictive models for KeratinoSensTM activation. We compare algorithms and molecular descriptors to propose a highly accurate prediction model. Methods: We collected and curated a dataset with KeratinoSensTM assay results through a literature review. Predictive models were developed by combining machine learning and deep learning algorithms with diverse molecular descriptors. Model performance was evaluated using five-fold cross-validation using multiple evaluation metrics, and PCA and UMAP were applied to data distribution analysis and identification of prediction error patterns. Results: The GCN and DNN models presented superior predictive performance compared to other machine learning-based models. The GCN model, in particular, achieved an average ROC-AUC of 0.80, indicating a strong learning capability from molecular structures. PCA and UMAP analyses confirmed that the GCN model effectively captured structural differences between datasets. However, prediction errors were observed for structurally novel compounds, likely due to the limited dataset size. Conclusions: This study highlights the GCN model's strong performance in predicting KE2 events within the AOP framework, underscoring its potential as a pre-screening tool for identifying skin sensitizers. Expanding the dataset could further improve its generalizability for efficient skin sensitization assessment.

목차

Ⅰ. 서 론

Ⅱ. 재료 및 방법

Ⅲ. 결과 및 고찰

Ⅳ. 결 론

References

키워드

Skin sensitization adverse outcome pathway keratinosense graph convolution network in silico