Synergizing Machine Learning, Conceptual Density Functional Theory, and Biochemistry: No-Code Explainable Predictive...

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Aromatic amines (AAs) are a significant class of chemical mutagens.
• Predictive modeling for mutagenic activity is crucial for risk assessment.
• Existing QSAR models often lack transparency and accessibility.

Purpose of the Study:

• To develop OECD-compliant, No-Code predictive models for AA mutagenicity.
• To integrate machine learning (ML) with conceptual density functional theory (CDFT).
• To identify key molecular descriptors for predicting mutagenic activity.

Main Methods:

• Utilized a dataset of 251 AAs with rigorous cross-validation (LOOCV).
• Employed the GFN2-xTB method for descriptor computation in vacuum and aqueous phases.
• Evaluated CDFT electrophilicity schemes (PSL, GCV, CDP) and a novel Log QP descriptor.
• Applied ML models: SPAARC, RandomTree, JCHAID*, and a Multilayer Perceptron.

Main Results:

• Achieved robust internal validation (Avg. Correct Classifications = 76%) and external validation (Avg. Correct Classifications = 79%).
• Identified the second CDP electrophilicity definition (ω_+Vac^CDP2+, ω_+Aq^CDP2+) and Log QP descriptors (LogQP1_+Vac) as most predictive.
• Highlighted the importance of metabolic activation, aqueous properties, and CDP/Log QP descriptors.

Conclusions:

• A replicable, No-Code methodology for QSAR research was established.
• The study demonstrates the power of combining ML with CDFT for toxicology prediction.
• Findings facilitate broader access to advanced predictive modeling for mutagenicity.