박준호 (서강대학교)

박준호 (서강대학교)

윤서휘 (서강대학교)

오한빈 (서강대학교)

The assessment of skin corrosion and irritation potential of chemicals is crucial for human health and chemical regulation. However, due to ethical and cost constraints on animal testing, the need for accurate <i style="font-family: "Times New Roman", serif; font-size: 12pt;">in silico</i> prediction models is growing. Existing models have relied on a single approach, using either deep learning based on molecular graphs or machine learning descriptors defined by experts, which has shown limitations in improving predictive performance.

The hybrid model proposed in this study combines Attentive FP (a Graph Attention Network) that learns the structural relationships of molecules with expert-defined physicochemical machine learning descriptors. An integrated feature vector, created by concatenating the graph embeddings from Attentive FP and the descriptor vectors, is utilized for the final prediction. This approach maximizes predictive performance through a synergistic effect where each model compensates for the other's weaknesses, capturing both the domain knowledge difficult for GAT to grasp alone and the structural context unknowable from descriptors alone.

Consequently, this hybrid model is expected to demonstrate significantly improved prediction accuracy compared to single-approach models. Moreover, by allowing for a multi-faceted interpretation of the model's predictive basis through GAT's attention weights and the SHAP (SHapley Additive exPlanations) values of the descriptors, it can provide a more reliable and explainable tool for chemical safety assessment.