DrugTar improves druggability prediction by integrating large language models and gene ontologies

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Target discovery is critical for developing therapeutics for complex chronic diseases.
• Computational druggability prediction methods show promise but often rely on limited sequence-based features.
• Existing machine learning methods face challenges with feature engineering, reproducibility, and accessibility.

Purpose of the Study:

• To investigate the utility of raw protein sequence and structure for druggability prediction.
• To develop a high-performance computational method for predicting drug targets.
• To improve the efficiency of target discovery in drug development.

Main Methods:

• Leveraged deep learning techniques integrating protein sequence embeddings (ESM-2) and gene ontologies.
• Developed DrugTar, a novel algorithm for predicting target druggability.
• Utilized both protein sequence and structure data for model training and validation.

Main Results:

• Protein sequence embeddings proved more informative than protein structure for druggability prediction.
• DrugTar achieved high performance, with AUC and PRC values of 0.94.
• DrugTar outperformed existing state-of-the-art methods in predicting druggability.

Conclusions:

• DrugTar effectively predicts drug target druggability using advanced deep learning.
• The developed algorithm streamlines target discovery, addressing a key bottleneck in therapeutic development.
• DrugTar offers a powerful, accessible tool for accelerating drug discovery research.