QM/NN QSPR models with error estimation: vapor pressure and logP

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in chemistry

Background:

• Quantitative structure-property relationship (QSPR) models are crucial for predicting chemical properties.
• Accurate prediction of logP and vapor pressure is essential in various chemical applications.
• Quantum mechanical calculations provide valuable descriptors for QSPR modeling.

Purpose of the Study:

• To develop and validate QSPR models for predicting logP and vapor pressure of organic compounds.
• To utilize neural network interpretation of quantum mechanical descriptors.
• To assess the reliability and performance of the developed predictive models.

Main Methods:

• Quantitative structure-property relationship (QSPR) modeling using multilayer feedforward neural networks.
• Derivation of molecular descriptors from semiempirical Austin Model 1 (AM1) calculations.
• Cross-validation using the back-propagation of errors algorithm with multiple test sets.
• Ensemble approach combining predictions from multiple neural networks.

Main Results:

• Successfully developed QSPR models for logP and room temperature vapor pressure.
• Achieved reliable predictions validated through rigorous cross-validation.
• Demonstrated the effectiveness of neural network interpretation of quantum mechanical descriptors.
• Quantified prediction reliability using mean predicted values and standard deviations.

Conclusions:

• Neural network interpretation of AM1-derived descriptors provides a robust method for QSPR modeling.
• The developed models offer reliable predictions for logP and vapor pressure of organic compounds.
• Cross-validation confirms the generalizability and performance of the QSPR models.