Integrated Skin Sensitization Assessment Using Consensus Modelling and Tiered New Approach Methodology: A SaferSkin Case Study

Abstract The transition toward animal-free safety assessment of chemicals has accelerated the development of New Approach Methodologies (NAMs) for predicting skin sensitization. However, individual in silico models and experimental NAM assays frequently produce inconsistent or contradictory results, limiting their reliability when used in isolation. To address this challenge, we present a tiered integrated assessment framework implemented through the open source SaferSkin application, which enables systematic comparison and integration of multiple predictive models and experimental data within a transparent weight-of-evidence workflow. In this case study, a diverse set of 21 reference compounds was evaluated using a battery of in silico approaches, including the OECD QSAR Toolbox, VEGA, CASE Ultra and additional machine-learning models implemented within SaferSkin. The platform enables side-by-side comparison of predictions and integration of experimental data through Bayesian network models, allowing probabilistic updating of predictions as new evidence becomes available. Our results demonstrate that reliance on any single predictive model is insufficient for robust hazard identification due to frequent disagreement between models. In contrast, consensus interpretation across multiple modelling approaches combined with targeted experimental evidence substantially improves predictive confidence. The integrated weight-of-evidence framework showed strong concordance with reference classifications and was further supported by independent validation using the Pred-Skin Bayesian model. Importantly, the tiered workflow enables resolution of ambiguous cases. For example, lower-tier predictions for ethyl (2E,4Z)-deca-dienoate were inconsistent across models, whereas targeted third-tier testing using the SENS-IS assay identified the compound as a strong sensitiser (GHS Category 1A). Overall, this study demonstrates how integrated modelling, Bayesian evidence updating and targeted NAM testing can reduce uncertainty in skin sensitization assessment. The SaferSkin framework provides a transparent and reproducible approach for implementing Next Generation Risk Assessment (NGRA) strategies and supports the development of animal-free regulatory toxicology and Safe-and-Sustainable-by-Design chemical innovation. Competing Interest Statement BH is CEO of Edelweiss Connect GmbH. The work of Tomaz Mohoric, Shaheena Parween, Csaba Boglari, Amanda Y. Poon, Daniel C. Ukaegbu, and Barry Hardy was carried out as employees at Edelweiss Connect GmbH which also supported the development of the SaferSkin application. The other authors declare that they have no conflict of interest. List of Abbreviations - ACD - Allergic Contact Dermatitis - AEL - Acceptable Exposure Level - AI - Artificial Intelligence - ANN - Artificial Neural Network - AOP - Adverse Outcome Pathway - BR - Borderline range - CAS - Chemical Abstracts Service - CP - Crossing Point - CV - Cell Viability - Cys - Cysteine - DA - Defined Approach - DE - Differential Expression - DPRA - Direct Peptide Reactivity Assay - EC - Estimated Concentration - h-CLAT - Human Cell Line Activation Test - HPLC - High Pressure Liquid Chromatography - HSP - Heat Shock Protein - IC - Inconclusive - ITS - Integrated Testing Strategy - KE - Key Event - LLNA - Local Lymph Node Assay - LTT - Lymphocyte Transformation Test - Lys - Lysine - MIT - Minimal Induction Threshold - MLR - Multi Linear Regression - NC - Not Classified - NESIL - No Expected Sensitization Induction Levels - OECD - Organisation for Economic Cooperation and Development - QMRF - QSAR Model Reporting Format - QPRF - QSAR Prediction Reporting Format - QSAR - Quantitative Structure Activity Relationship - RF - Random Forest - RFI - Relative Fluorescence Intensity - SAF - Sensitization Assessment Factor - SMILES - Simplified Molecular Input Line Entry System - UN - GHS United Nations Globally Harmonised System