Melting point prediction employing k-nearest neighbor algorithms and genetic parameter optimization

Area of Science:

• Computational Chemistry
• Cheminformatics
• Physical Chemistry

Background:

• Accurate prediction of melting points is crucial for drug discovery and materials science.
• Existing methods often struggle with diverse chemical spaces and capturing complex molecular interactions.

Purpose of the Study:

• To apply and optimize the k-nearest neighbor (kNN) modeling technique for predicting melting points of organic molecules and drugs.
• To investigate the impact of molecular descriptors and weighting schemes on prediction accuracy.
• To compare model performance across different chemical spaces (general organic molecules vs. drugs).

Main Methods:

• Utilized a k-nearest neighbor (kNN) modeling approach.
• Employed two datasets: 4119 diverse organic molecules and 277 drugs.
• Investigated various molecular descriptors and four weighting schemes (arithmetic/geometric average, inverse distance, exponential weighting).
• Optimized the model using a genetic algorithm and validated via 25-fold Monte Carlo cross-validation.

Main Results:

• Exponential weighting scheme provided the best prediction results.
• kNN model achieved an average RMSE of 46.2 °C (r²=0.49) for organic molecules and 42.2 °C (r²=0.42) for drugs.
• Drug-based predictions were more accurate (RMSE=46.3 °C, r²=0.30) than general molecule-based predictions (RMSE=50.3 °C, r²=0.20) for the drug dataset.
• Identified inherent systematic error in kNN and limitations due to uncaptured liquid state interactions.

Conclusions:

• The kNN method, particularly with exponential weighting and genetic algorithm optimization, offers a viable approach for melting point prediction.
• Model performance is influenced by the chemical space of the training data.
• Future improvements require incorporating descriptors that better represent liquid state interactions.