Machine-learning-based predictions of imprinting quality using ensemble and non-linear regression algorithms

Area of Science:

• Polymer Chemistry
• Materials Science
• Computational Chemistry

Background:

• Molecularly imprinted polymers (MIPs) are versatile artificial materials with tailored recognition sites.
• MIPs offer stability, reusability, and high selectivity but optimizing synthesis conditions is challenging.
• Current optimization methods are time-consuming, costly, and resource-intensive.

Purpose of the Study:

• To investigate the application of machine learning (ML) for predicting the imprinting factor (IF) of MIPs.
• To overcome the limitations of traditional experimental optimization of MIP synthesis.
• To identify key factors influencing MIP selectivity and performance.

Main Methods:

• Utilized non-linear regression ML algorithms: classification and regression tree, support vector regression, k-nearest neighbors, and ensemble methods (gradient boosting, random forest, extra trees).
• Employed a mutual information feature selection method to identify critical parameters affecting IF.
• Trained ML models using experimentally derived data, including pH, template type, monomer type, solvent, K_MIP, and K_NIP.

Main Results:

• The gradient boosting (GB) algorithm demonstrated superior performance in predicting the IF.
• GB achieved the highest R-squared value (0.871), indicating a strong model fit.
• GB also yielded the lowest Mean Absolute Error (MAE = -0.982) and Mean Square Error (MSE = -2.303).

Conclusions:

• Machine learning, particularly gradient boosting, is a powerful tool for predicting MIP imprinting factors.
• ML-based prediction significantly accelerates the optimization of MIP synthesis, reducing experimental costs and time.
• This approach enhances the efficiency and accuracy of developing selective molecularly imprinted polymers.