Machine Learning for Quantitative Prediction of Protein Adsorption on Well-Defined Polymer Brush Surfaces with Diverse Chemical Properties

Polymer informatics has attracted increasing attention because machine learning can establish quantitative structure–property relationships in polymer materials. Understanding and controlling protein adsorption on polymer surfaces are crucial for various applications, such as protein immobilization supports, biosensors, and antibiofouling surfaces. However, protein adsorption is a complex phenomenon that is difficult to predict quantitatively owing to the involvement of multiple factors. Therefore, this study aims to establish a machine learning model for protein adsorption on densely packed polymer brushes with various chemical structures, as these surfaces are well-suited for analyzing structure–property correlations between the polymer’s chemical structure and adsorption amount during initial protein adsorption. Two proteins, bovine serum albumin (BSA) and lysozyme, are adopted as target proteins, with the expectation that differences in their charge profiles will be reflected in the resulting machine learning model. The descriptors of the polymer brush surfaces include their grafted structures (thickness) and chemical properties, which are described by the contact angle and ζ potential. This allows physicochemical knowledge to be incorporated into the machine learning model. Random forest exhibits the best performance in all situations, accurately predicting the amounts of adsorbed BSA and lysozyme. In addition, the prediction of the contact angle and ζ potential by machine learning also enables a quantitative and explainable prediction of protein adsorption based on theoretical molecular descriptors, ensuring that no characteristics are overlooked. Moreover, the model is used to analyze the contributions of electrostatic and hydrophobic interactions to protein adsorption. In conclusion, a machine learning model is developed to predict protein adsorption on polymer brush surfaces, incorporating descriptors such as the grafted structure, contact angle, and ζ potential. It provides quantitative predictions and analyzes the roles of electrostatic and hydrophobic interactions, advancing the design of functional polymer surfaces for applications in biosensors and antifouling technologies.高分子信息学日益受到学界关注，这是因为机器学习能够在高分子材料中构建定量的结构-性能关联关系。理解并调控高分子表面的蛋白质吸附，对于蛋白质固定化载体、生物传感器以及抗生物污染表面等诸多应用场景至关重要。然而，蛋白质吸附是一类复杂现象，由于涉及多种影响因素，其定量预测颇具挑战。为此，本研究旨在针对具有不同化学结构的致密刷型高分子刷表面的蛋白质吸附问题构建机器学习模型——这类表面非常适合用于分析初始蛋白质吸附过程中，高分子化学结构与吸附量之间的结构-性能关联。本研究选用牛血清白蛋白（bovine serum albumin，BSA）和溶菌酶作为目标蛋白，预期二者的电荷分布差异可在最终构建的机器学习模型中得到体现。该模型的高分子刷表面描述符涵盖接枝结构（如厚度）与化学性质，其中化学性质通过接触角和ζ电势表征，这使得物理化学相关知识能够被融入机器学习模型之中。随机森林（Random Forest）在所有测试场景中均表现最优，可精准预测BSA与溶菌酶的吸附量。此外，通过机器学习对接触角与ζ电势的预测，还可基于理论分子描述符实现蛋白质吸附的定量且可解释性预测，确保不会遗漏任何特征信息。进一步而言，本模型还可用于分析静电相互作用与疏水相互作用对蛋白质吸附的贡献程度。综上，本研究构建了可预测高分子刷表面蛋白质吸附的机器学习模型，其纳入了接枝结构、接触角与ζ电势等描述符。该模型不仅可实现定量预测，还可分析静电与疏水相互作用的调控机制，从而推动面向生物传感器与防污技术应用的功能性高分子表面设计发展。