Advancing Adverse Drug Reaction Prediction with Deep Chemical Language Model for Drug Safety Evaluation.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Accurate prediction of adverse drug reactions (ADRs) is critical for drug safety.
• Pre-trained deep chemical language models show potential for predicting molecular properties.
• The efficacy of these models for predicting idiosyncratic ADRs is underexplored.

Purpose of the Study:

• To develop and evaluate MoLFormer-XL, a novel pre-trained model for ADR prediction.
• To assess the model's performance in predicting drug-induced QT interval prolongation (DIQT), teratogenicity (DIT), and rhabdomyolysis (DIR).
• To identify molecular substructures associated with specific ADRs.

Main Methods:

• Utilized MoLFormer-XL, a pre-trained model encoding molecular features from canonical SMILES.
• Integrated MoLFormer-XL with a CNN-based model for predicting DIQT, DIT, and DIR.
• Analyzed learned linear attention maps to identify ADR-associated chemical substructures.

Main Results:

• MoLFormer-XL significantly outperformed conventional models in predicting DIQT, DIT, and DIR.
• The model demonstrated robust performance in identifying specific ADRs.
• Learned attention maps identified amines, alcohol, ethers, and aromatic halogen compounds as key substructures linked to the studied ADRs.

Conclusions:

• MoLFormer-XL offers a powerful approach for predicting specific ADRs.
• The findings can enhance drug discovery pipelines by identifying potential safety liabilities early.
• This research contributes to reducing drug attrition rates due to safety concerns.