Neural network modeling for estimation of partition coefficient based on atom-type electrotopological state indices

Area of Science:

• Computational chemistry
• Cheminformatics

Background:

• Accurate prediction of the n-octanol/water partition coefficient (log P) is crucial for drug discovery and environmental fate assessment.
• Traditional methods for log P prediction often struggle with complex molecular structures.

Purpose of the Study:

• To develop and validate a novel computational method for predicting log P values.
• To leverage atom-type electrotopological-state (E-state) indices and artificial neural networks (ANNs) for enhanced prediction accuracy.

Main Methods:

• Utilized an extended set of E-state indices, including detailed descriptors for amino, carbonyl, and hydroxy groups.
• Employed a 39-5-1 artificial neural network architecture with molecular weight and 38 atom-type E-state indices as inputs.
• Compared ANN performance against multilinear regression (MLR) using both training and independent test sets.

Main Results:

• ANNs achieved high accuracy on the training set (r² = 0.90, RMS(LOO) = 0.46) and superior performance on the test set (predictive r² = 0.94, RMS = 0.41) compared to MLR (r² = 0.86, RMS = 0.72).
• MLR showed good performance on the training set (r² = 0.87, RMS(LOO) = 0.55) but lower predictive power on the test set.
• The nonlinear capabilities of ANNs enabled the detection of complex relationships between E-state indices and log P.

Conclusions:

• The developed ANN-based method provides an accurate, fast, and reliable approach for estimating log P values, even for complex organic molecules.
• This computational strategy offers a significant improvement over traditional methods like MLR for log P prediction.