BitterTranslate: A Natural Language Processing and Machine Learning-Based Framework for Mapping Bitter Taste Receptor...

Area of Science:

• Pharmacology
• Computational Biology
• Artificial Intelligence

Background:

• Bitter taste receptors (TAS2Rs) are G protein-coupled receptors (GPCRs) found on the tongue and in extraoral tissues.
• TAS2R ligands have applications in improving drug compliance, treating diseases, and understanding receptor functions.
• Current machine learning approaches for predicting TAS2R activators require extensive feature engineering.

Purpose of the Study:

• To explore the potential of transformer models for predicting TAS2R ligands.
• To develop an efficient algorithm for TAS2R agonist prediction using AI.

Main Methods:

• Developed BitterTranslate, an algorithm utilizing two Bidirectional Encoder Representations from Transformers (BERT) models.
• One BERT model extracts features from ligand SMILES strings; the other extracts features from GPCR amino acid sequences.
• An XGBoost classifier was pre-trained on GPCR-ligand data and fine-tuned on TAS2R-ligand data.

Main Results:

• BitterTranslate achieves 80% precision and 65% recall across all TAS2Rs.
• For TAS2R14, the algorithm demonstrates 83% precision and 88% recall.
• The method simplifies feature extraction for TAS2R ligand prediction.

Conclusions:

• Transformer models, specifically BERT, are effective for predicting TAS2R ligands.
• BitterTranslate offers a streamlined approach to drug discovery for TAS2Rs.
• This AI-driven method reduces the need for specialized software in feature extraction.