A High-Performance and Interpretable pK_a Prediction Framework Integrating Count-Based Fingerprints and Ensemble...

Area of Science:

• Computational chemistry
• Environmental science
• Machine learning

Background:

• Accurate prediction of acid dissociation constant (pKa) is crucial for understanding organic compound environmental fate.
• Traditional binary Morgan fingerprints (B-MF) lack stoichiometric information, hindering accurate pKa modeling.
• Substituent effects significantly influence pKa, necessitating improved predictive methods.

Purpose of the Study:

• Develop an interpretable machine learning framework for high-performance pKa prediction.
• Integrate count-based Morgan fingerprints (C-MF) to capture functional group multiplicity.
• Enhance pKa prediction accuracy and generalizability for environmental applications.

Main Methods:

• Developed a machine learning framework using count-based Morgan fingerprints (C-MF) and ensemble algorithms.
• Employed SHAP-based recursive feature elimination (SHAP-RFE) for model optimization.
• Defined the applicability domain using the AD_SAL method for reliable predictions.

Main Results:

• Count-based Morgan fingerprints (C-MF) outperformed traditional binary Morgan fingerprints (B-MF).
• The optimized Catboost model achieved high accuracy (test-set R² = 0.890, RMSE = 1.026).
• SHAP analysis confirmed chemically intuitive feature importance and model interpretability.
• External validation demonstrated strong generalizability with R² = 0.890 and RMSE = 0.942.

Conclusions:

• The developed framework provides a robust and generalizable tool for accurate pKa prediction.
• The model's interpretability facilitates understanding of substituent effects on pKa.
• This approach has significant potential for environmental risk assessment and chemical safety evaluations.