Data-Driven Machine Learning Framework for the Regulation of Protein Adsorption on Surfaces

Protein
adsorption on surfaces is a highly complex process, governed
by intricate interactions between protein, surface and surrounding
environment. However, accurately predicting protein adsorption amounts
and precisely controlling adsorption behavior at surfaces remain significant
challenges. In this work, we present a data-driven machine learning
framework that systematically evaluates the determinants of protein
adsorption, thereby enabling robust prediction of adsorption amount
and adsorption behavior. Seven critical descriptors were selected
and used in the AutoGluon framework to construct three optimal models,
with the WeightedEnsemble_L2 (WE_L2) model exhibiting superior performance.
Furthermore, SHAP analysis quantified and evaluated the contributions
of the seven descriptors, revealing their respective promoting or
inhibiting effects on protein adsorption along with their percentage
impacts. Moreover, the WE_L2 model accurately predicts BSA adsorption
on various SAM surfaces, demonstrating excellent agreement with experimental
values. It also effectively captures the complex adsorption behavior
of HSA, including bilayer formation, electrostatic shielding effects,
and the influence of surface hydrophobicity. These findings demonstrate
that interpretable machine learning models not only facilitate precise
predictions of protein adsorption amounts, but also provide a powerful
tool for regulating interfacial protein adsorption.