LigEGFR

LigEGFR predicts pIC50 values and classifies hit compounds (pIC50 ≥ 6) for small molecules targeting human epidermal growth factor receptor (EGFR) using a deep learning model that integrates spatial graph embeddings and molecular descriptors.

Key Features:

• Deep learning integration: Integrates spatial graph embeddings with non-hashed and hashed molecular descriptors to model small-molecule bioactivity against EGFR.
• Prediction and classification performance: Demonstrates superior predictive accuracy and hit classification compared with baseline machine learning models and molecular docking approaches for identifying compounds with pIC50 ≥ 6.
• Large-scale cell line-based dataset: Trained on a comprehensive, non-redundant cell line-based dataset encompassing diverse chemical features to improve generalizability.
• Architectural and validation innovations: Processes non-hashed descriptors with convolutional layers and hashed descriptors with fully connected layers, and applies y-randomization and applicability domain analysis (ADAN) for robustness assessment.

Scientific Applications:

• pIC50 prediction: Estimate potency (pIC50) of small molecules against human EGFR to inform lead selection.
• Hit compound classification: Classify and prioritize compounds as hits using a pIC50 ≥ 6 threshold for downstream experimental validation.
• EGFR-targeted drug discovery: Support prioritization of candidate therapeutics in research on EGFR-associated lung cancer.

Methodology:

Uses a deep learning model combining spatial graph embeddings with non-hashed and hashed molecular descriptors, processes non-hashed descriptors via convolutional layers and hashed descriptors via fully connected layers, and performs y-randomization and applicability domain analysis (ADAN) while training on a non-redundant cell line-based dataset.