Identifying Polymers that Bind or Reject Proteins with Machine Learning: Handling Categorical Features within a GPR...

Area of Science:

• Materials Science
• Biochemistry
• Computational Biology

Background:

• Understanding polymer-protein interactions is crucial across diverse fields like medicine, food science, and water treatment.
• Applications range from enhancing enzyme stability to minimizing protein adsorption in nanomedicine.

Purpose of the Study:

• To identify polymers with maximal and minimal binding affinities to a panel of proteins using machine learning.
• To evaluate the efficacy of various Gaussian Process Regression (GPR) models, particularly those incorporating categorical features.

Main Methods:

• Development and testing of polymer libraries with diverse monomer compositions.
• Application of Gaussian Process Regression (GPR) models, including Multiplicative kernel, Additive kernel, EzGP, LVGP, and LMGP.
• Utilizing Förster resonance energy transfer (FRET) data for binding strength quantification to generate machine learning datasets.

Main Results:

• The Latent Variable Gaussian Processes (LVGP) model demonstrated superior performance on the polymer-protein binding dataset.
• Polymers exhibiting high protein affinity possessed positive charges and hydrophobic benzyl groups.
• Polymers with strong protein repulsion were dominated by negatively charged monomers, with some cationic units.

Conclusions:

• LVGP is a powerful tool for predicting polymer-protein interactions, aiding in the design of materials with specific binding properties.
• The study elucidates the relationship between polymer chemical structure (charge, hydrophobicity) and protein binding affinity.
• Findings can guide the development of novel polymers for applications requiring controlled protein interactions.