Comparative Study of Molecular Descriptors and AI-Based Embeddings for Toxicity Prediction

Area of Science:

• Computational chemistry
• Toxicology
• Artificial intelligence in drug discovery

Background:

• Accurate toxicity prediction is crucial for pharmaceutical development and regulatory safety.
• Traditional methods rely on molecular descriptor-based models.
• Emerging AI language models offer new approaches for chemical data analysis.

Purpose of the Study:

• To compare the performance of descriptor-based features versus AI language model embeddings for toxicity prediction.
• To evaluate models across Tox21, ClinTox, and DILIst datasets.
• To assess the utility of different data inputs (SMILES, chemical names, descriptions) for AI models.

Main Methods:

• Utilized descriptor-based features from Mordred and RDKit.
• Applied ten AI language models to generate embeddings from SMILES strings, chemical names, and descriptions.
• Employed logistic regression classifiers for toxicity prediction tasks.
• Evaluated performance using ROC-AUC on Tox21, ClinTox, and DILIst datasets.

Main Results:

• Mordred (descriptor-based) achieved the highest average ROC-AUC (0.855) on the Tox21 dataset.
• AI language models outperformed descriptor models on ClinTox and DILIst datasets.
• GPT-3 demonstrated superior performance on ClinTox (ROC-AUC 0.996) and DILIst (ROC-AUC 0.806) using textual data.
• MolBERT showed competitive performance on Tox21 (ROC-AUC 0.801) using SMILES embeddings.

Conclusions:

• AI language models show significant promise for predictive toxicology, especially for focused endpoints.
• Descriptor-based models remain effective for multi-endpoint predictions.
• Integrating molecular descriptors with textual embeddings could enhance predictive accuracy and model adaptability.
• This study highlights the potential of AI in revolutionizing drug safety assessment.