Deciphering Molecular Embeddings with Centered Kernel Alignment.

Area of Science:

• Machine Learning
• Cheminformatics
• Computational Chemistry

Background:

• Analyzing nonlinear machine learning models presents significant challenges.
• Centered Kernel Alignment (CKA) is a tool for assessing embedding similarity, effective for neural networks but underutilized in cheminformatics.
• Random Forest (RF) models are popular in cheminformatics but lack the rotational invariance required by standard CKA.

Purpose of the Study:

• To adapt Centered Kernel Alignment (CKA) for analyzing Random Forest (RF) models.
• To develop a method that accounts for the properties of RF algorithms.
• To enable better understanding and explanation of RF model behavior in cheminformatics.

Main Methods:

• Adapted CKA by developing a specific kernel tailored to Random Forest (RF) properties.
• Validated the adapted CKA method by comparing its results with prediction similarity of RF models.
• Applied the RF-kernel CKA to analyze and explain RF models built from molecular and rooted fingerprints.

Main Results:

• The adapted CKA method shows strong correlation with the prediction similarity of RF models.
• Demonstrated the utility of CKA with the RF kernel for analyzing and explaining RF model behavior.
• Successfully applied the method to analyze models derived from molecular and rooted fingerprints.

Conclusions:

• The adapted CKA provides a robust method for analyzing RF models in cheminformatics.
• This approach enhances the interpretability of machine learning models in molecular sciences.
• The RF-kernel CKA is a valuable tool for understanding complex relationships in chemical data.