Machine Learning for Go/No-Go Screening of Acute Dermal Toxicity to Support Sustainable Chemical Design

Early identification of potentially hazardous chemicals is crucial in the Safe and Sustainable by Design (SSbD) framework. Developing digital screening tools prior to material design will enable proactive hazard identification to avoid costly redesign. This article presents an XGBoost Machine Learning (ML) classifier to support Go/No-Go decisions on the acute dermal toxicity of small organic molecules used in materials design. The molecules were characterized using selected RDKit descriptors, including structural, electronic, and surface-area features. The final model was evaluated as a screening tool on the locked-scaffold test dataset, which yielded an accuracy of 0.853, a toxic-class precision of 0.587, a toxic-class recall of 0.574, a toxic-class F1-score of 0.581, an ROC-AUC of 0.882, and a PR-AUC of 0.657. From a Go/No-Go perspective, the model gave a conditional Go classification for 398 out of 436 non-toxic chemicals and blocked 54 out of 94 toxic chemicals. As a result, the ratio of non-toxic chemicals allowed was 91.3%, while the ratio of toxic chemicals stopped was 57.4%. The above results show that it is possible to use a ML conditional model evaluation focused on the toxic class. The suggested methodology can be used in the early-stage SSbD process to assess conditional Go classification for human chemical safety by providing rapid toxicity rankings. This ML model indicates progression to a next screening stage, which is needed to prove that a chemical is sustainable, as the complete SSbD process considers multiple factors.