Machine learning driven LD₅₀ prediction for cancer risk assessment using modern molecular language models.

Area of Science:

• Computational toxicology
• cheminformatics
• Machine learning for drug discovery

Background:

• Accurate chemical toxicity assessment is crucial for cancer research, guiding carcinogenicity testing, safety evaluations, and regulatory decisions.
• Early identification of hazardous compounds minimizes risks in drug development and chemical safety.

Purpose of the Study:

• To develop and evaluate ChemModernBERT, a novel molecular language model for predicting chemical toxicity.
• To compare the performance of ChemModernBERT against other molecular representation learning methods for toxicity prediction.

Main Methods:

• Developed ChemModernBERT, a ModernBERT-based model pretrained on over 1.8 million SMILES strings using curriculum learning.
• Compared ChemModernBERT with ChemBERT, ChemProp (a message-passing neural network), and ensemble learning on a dataset of 8,898 compounds.
• Evaluated model performance using internal and external test sets for predicting median lethal dose (LD50) values.

Main Results:

• ChemModernBERT achieved the lowest mean absolute error (MAE) and highest coefficient of determination (R2) on both internal and external test sets.
• Outperformed existing methods like ChemBERT, ChemProp, and ensemble models in predicting LD50 values.
• Demonstrated strong transferability across diverse chemical compounds with a minimal generalization gap.

Conclusions:

• Curriculum-pretrained transformer architectures offer a scalable and accurate framework for large-scale toxicity prediction.
• ChemModernBERT can significantly support computational pipelines for carcinogenicity assessment, dose selection, and early chemical safety evaluations.
• This study highlights the potential of advanced language models in advancing chemical safety and drug development research.