Absorption Distribution Metabolism Excretion and Toxicity Property Prediction Utilizing a Pre-Trained Natural...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are crucial in drug discovery for interpreting drug structures.
• Machine learning (ML) techniques, particularly natural language processing (NLP), are increasingly utilized in this domain.

Purpose of the Study:

• To evaluate the performance of a pre-trained NLP model, ChemBERTa, in drug discovery.
• To compare four different model architectures (DNN, encoder, concatenation, pipe) using physicochemical properties and Simplified Molecular Input Line Entry System (SMILES) data.

Main Methods:

• Utilized ChemBERTa, a pre-trained NLP model, for drug discovery tasks.
• Developed and compared four model architectures: DNN (physicochemical properties), encoder (SMILES + NLP), concatenation (SMILES + properties, parallel), and pipe (SMILES + properties, sequential).
• Collected 5238 entries from DrugBank with physicochemical and ADMET features.

Main Results:

• The encoder model achieved the highest AUROC (76.0%) on the initial dataset, outperforming DNN (62.4%), concatenation (74.9%), and pipe (68.2%).
• On an external microsomal stability dataset, the DNN model showed superior predictive capability with an AUROC of 78%, compared to encoder (44%) and concatenation (50%).
• These results suggest that models relying solely on structural information (SMILES) may need further optimization.

Conclusions:

• The DNN model, utilizing physicochemical properties, demonstrated better generalization for new, unseen data.
• The study highlights the promise of NLP in pharmaceutical applications but also points to the need for more extensive datasets to improve model generalization.
• Further research into alternative tokenization strategies for structural data may enhance the performance of NLP-based drug discovery models.