Toward Assay-Aware Bioactivity Model(er)s: Getting a Grip on Biological Context

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and computational biology
• Drug discovery and development

Background:

• Proteochemometric (PCM) models are crucial for predicting protein-ligand interactions in early drug discovery.
• Existing PCM models often overlook the diversity of biological assays, leading to potential inaccuracies.
• Lack of standardized assay metadata hinders data interpretation and model performance.

Purpose of the Study:

• To develop and validate assay descriptors to capture biological context.
• To improve the accuracy and specificity of protein-ligand interaction predictions.
• To enhance data curation and understanding of bioactivity data.

Main Methods:

• Utilized dimensionality reduction on ChEMBL assay descriptions to create text embeddings.
• Applied clustering to group similar assays based on embedded descriptions.
• Integrated BioBERT-based assay embeddings into PCM models.
• Compared performance against simpler methods like bag-of-words.

Main Results:

• Assay embeddings effectively capture relevant features and group assays meaningfully.
• Clustering assays improved purity and matched manual categorizations.
• Models incorporating assay embeddings showed improved predictive accuracy (average R² increased from 0.67 to 0.69).
• Bag-of-words approach did not yield improvements.

Conclusions:

• Biological assay context is a critical factor for accurate bioactivity modeling.
• Assay embeddings enhance PCM model performance and allow prediction for specific endpoints.
• The developed method aids in assay categorization, data curation, and understanding biological context.