Combining Explicit Molecular Features and Latent ChemBERTa Embeddings for Ecotoxicological Screening

Introduction: Evaluating the environmental impact of cosmetic ingredients requires robust alternatives to in vivo testing. This study proposes an in silico workflow to predict fish toxicity by integrating bioconcentration factor (BCF; toxicokinetics) and Aryl Hydrocarbon Receptor (AhR) activation (toxicodynamics) models.

Methods: Structures were standardized (cleanup, salt removal, charge neutralization and tautomeric canonicalization with RDkit), retaining canonical SMILES with >= 2 carbon atoms and specific organic elements. A hybrid descriptor pipeline combined six physicochemical properties, 166 MACCS keys, and 384 transformer ChemBERTa-77M-MLM embeddings. Dimensionality reduction incorporated low-variance filtering, principal component analysis (PCA) for ChemBERTa embeddings, and Random Forest (RF) feature selection. The BCF dataset (N=1,046, from three public sources) trained an RF regressor. Duplicated SMILES were excluded or merged by the mean of logBCF if standard deviation <= 0.25. The AhR dataset (PubChem AID 2845; N=948 actives after removing discordant duplicates) trained an XGBoost classifier using undersampling for class imbalance and a 70/20/10 split.

Results: The BCF model achieved a Leave-One-Cluster-Out Cross-Validation (LOCO-CV) R^2 of 0.48 (RMSE=0.88). On an external validation set (N=77), it yielded R^2=0.54 and RMSE=0.66, outperforming existing literature models for the same molecules with minimal overfitting (Y-scramble R^2=-0.05). The AhR activation model achieved a cross-validation Q^2 of 0.89(Y-scramble ROC-AUC=0.49). The validation set metrics were: accuracy 0.86, sensitivity 0.74, specificity 0.88, F1-score 0.56, and ROC-AUC 0.88.

Conclusions: Integrating explicit features with Transformer-based embeddings generates robust predictive models. Despite improving predictive performance, dense embeddings lack the mechanistic transparency of OECD QSAR Principle 5, hindering regulatory acceptance. Addressing this gap requires coupling explainable AI (e.g., SHAP and similarity maps) with established frameworks like AOP 150 (AhR Activation) to provide mechanistic weight-of-evidence.