Applied machine learning for predicting the lanthanide-ligand binding affinities.

Area of Science:

• Computational Chemistry and Materials Science
• Machine Learning Applications in Chemistry

Background:

• Metal-ligand complex binding affinities are crucial for applications like drug design and chelation therapy.
• Traditional molecular modeling is computationally expensive and lacks accuracy for large-scale screening of metal-ligand interactions.
• Experimental data-driven approaches offer a promising alternative for predicting metal-binding affinities.

Purpose of the Study:

• To develop a machine learning framework for predicting the binding affinities (logK1) of lanthanide cations with diverse molecular ligands.
• To identify key molecular, metallic, and solvent features influencing metal-ligand binding affinities.
• To accelerate the screening and design of novel ligands for targeted applications.

Main Methods:

• Trained six supervised machine learning algorithms (RF, KNN, SVM, KRR, MLP, AdaBoost) on a large dataset of experimental logK1 values.
• Validated the models using a 10-fold cross-validation procedure.
• Performed feature engineering and importance analysis to identify relevant predictive features.

Main Results:

• The machine learning framework demonstrated excellent predictive ability for lanthanide-ligand binding affinities.
• Feature importance analysis identified critical factors influencing binding affinity.
• The best-performing AdaBoost model predicted logK1 values for approximately 71 million compounds in the PubChem database.

Conclusions:

• The developed machine learning framework significantly enhances the prediction accuracy and efficiency of metal-ligand binding affinities.
• This approach enables rapid screening of vast chemical spaces, facilitating the discovery of new ligands.
• The methodology holds potential for accelerating ligand design in diverse fields, including drug discovery and materials science.