Comprehensive Drug-Likeness Prediction Using a Pretrained Transformer Model and Multitask Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Drug-likeness is crucial for identifying viable drug candidates.
• Current methods struggle with feature engineering, generalizability, and adaptability in drug development.
• Limitations hinder the efficiency and scope of traditional drug-likeness prediction.

Purpose of the Study:

• To introduce an innovative framework for enhanced drug-likeness prediction.
• To overcome limitations of existing rule-based and machine learning approaches.
• To improve the accuracy and generalizability of computational drug discovery tools.

Main Methods:

• Integration of molecular pretrained transformer models with multitask learning.
• Development of two models: SpecDL for specialized tasks and GeneralDL for broad evaluation.
• Utilizing attention weight analysis for interpretable model outputs.

Main Results:

• SpecDL achieved an average ROC-AUC of 0.836 across four drug-likeness tasks.
• GeneralDL attained an average ROC-AUC of 0.781 on six diverse test sets, outperforming existing methods.
• GeneralDL demonstrated strong generalization to toxicity and biological activity predictions.

Conclusions:

• The proposed framework offers a powerful and generalizable solution for drug-likeness prediction.
• This approach has significant potential to accelerate and enhance early-stage drug discovery.
• The integration of advanced AI models provides more accurate and adaptable drug candidate assessment.