E-GuARD

E-GuARD addresses data scarcity in early drug discovery by combining self-distillation, active learning, and expert-guided molecular generation to improve the detection of compounds interfering with biological assays.

Key Features:

• Iterative data enrichment: E-GuARD iteratively enriches training data with interference-relevant molecules.
• Self-distillation: Models learn from newly generated, informative compounds through iterations to facilitate understanding of underrepresented structural patterns.
• Active learning: Integrated approach to select informative samples for model improvement.
• Expert-guided molecular generation: Guided molecular augmentation to enhance predictive performance of quantitative structure-interference relationship (QSIR) models.
• Quantitative structure-interference relationship modeling: Generates QSIR models with superior performance for predicting assay interference.

Scientific Applications:

• Thiol reactivity prediction: Detection of compounds causing thiol reactivity interference in biological assays.
• Redox reactivity prediction: Identification of compounds with redox reactivity that may interfere with assays.
• Nanoluciferase inhibition prediction: Detection of compounds that inhibit nanoluciferase and cause assay interference.
• Firefly luciferase inhibition prediction: Identification of firefly luciferase inhibitors that interfere with biological assays.
• Early drug discovery screening: Scalable solution for mitigating assay interference across different interference mechanisms.

Methodology:

E-GuARD integrates self-distillation, active learning, and expert-guided molecular generation in an iterative framework that enriches training data with interference-relevant molecules.