Prediction and Interpretation Microglia Cytotoxicity by Machine Learning.

Area of Science:

• Computational chemistry
• Neuroscience
• Drug discovery

Background:

• Neuroinflammation driven by microglia is key in neurodegenerative diseases.
• Plant compounds are promising drug candidates, but require cytotoxicity screening.
• High-throughput screening is costly and inefficient for early-stage compound evaluation.

Purpose of the Study:

• To develop a machine learning model for predicting microglia cytotoxicity of plant compounds.
• To identify effective computational methods for early-stage drug screening.
• To provide a reliable tool for selecting non-cytotoxic compounds in drug discovery.

Main Methods:

• A microglia cytotoxicity classifier was built using machine learning (ML).
• Data splitting was based on molecule murcko generic scaffold.
• ML approaches combined with molecular representations were assessed; feature selection via recursive feature elimination integrated with support vector machine (RFE-SVC) was applied.
• Shapley additive explanations (SHAP) were used for model interpretation.

Main Results:

• The ECFP4-RFE-SVM model achieved high accuracy (ACC=0.99, BA=0.99, F1=0.99, MCC=0.97) in classifying microglia cytotoxicity.
• SHAP analysis identified key substructures associated with cytotoxicity.
• The model demonstrated reliable classification capability and interpretability.

Conclusions:

• The ECFP4-RFE-SVM model is a reliable tool for predicting microglia cytotoxicity.
• SHAP analysis provides insights into cytotoxicity predictions and guides molecular modifications.
• This computational approach enhances early-stage drug screening for neurodegenerative diseases.