Ensuring the safety of novel chemicals requires proactive strategies that consider environmental and health hazards already at the molecular design stage [1]. Phenolic antioxidants are widely used and increasingly considered contaminants of emerging concern, exemplifying the need for molecular (re)design that maintains function while reducing risk [2]. This work addresses this challenge by developing an interactive generative modelling framework for the de novo design of safer phenolic antioxidants aligned with Safe and Sustainable by Design (SSbD) principles.
We employ a genetic algorithm to generate and evolve molecular structures while optimizing multiple criteria spanning synthesizability, antioxidant function, environmental fate, and (eco)toxicity. Because these criteria rely on imperfect predictors, a fully automated search can exploit their weaknesses and produce high-scoring but practically unsuitable structures. To counteract this we have built a human-in-the-loop interface that enables domain experts to supervise and guide the optimization. Users can dynamically adjust objective weights and components, remove undesirable candidates and manually add new structures, integrating intuition that is otherwise difficult to capture.
The genetic algorithm efficiently identifies a diverse set of promising phenolic structures with improved predicted profiles for performance and safety. Molecular similarity analysis confirms high structural diversity among the generated molecules, which supports a broad exploration of independent structural strategies. Comparison with a curated set of antioxidants in current use shows that early generations of molecules generated by the genetic algorithm rediscover known molecules, while later generations explore increasingly novel chemical space.
This work demonstrates how generative modelling, expert interaction, and predictive hazard assessment can be integrated into a practical SSbD-oriented design workflow. It is embedded in an interdisciplinary project in which our project partners are developing experimental high-throughput methods for assessing persistence, toxicity, and degradation-product hazards. The project lays the foundation for a closed-loop design approach coupling high-throughput experimentation and generative modelling via active learning to accelerate the development of safer functional chemicals.

[1] European Commission Recommendation of 8.12.2022 establishing a European assessment framework for ”safe and sustainable by design” chemicals and materials.
[2] Carolin Seller-Brison, Franziska Weissbach, Kjell Jorner, Martin Scheringer, Kathrin Fenner, Environmental Science & Technology Letters 2026, 13, 560–567.