Fouling of material surfaces due to biomolecule adsorption is a serious issue since it can decrease the performance and lifetime of devices such as biosensors, medical devices and water treatment membranes. Thus, functional surfaces with excellent antifouling properties are required to maintain device performance over the long term. The typical method for applying antifouling properties to materials is to modify the surface with a super hydrophilic zwitterionic polymer such as poly (2-methacryloyloxyethyl phosphorylcholine) (MPC) in the brush state and form a robust hydration layer on the surface. So far, many research papers and practical applications regarding such antifouling surface have been reported (Fig. 1a).

　It is well known that optimum surface condition is different depending on the water quality and water composition in the operating environment. Therefore, to maximize antifouling properties, it is essential to explore the optimum polymer structure for the target water environment and to design the surface brush state, especially the polymer density and molecular weight. Most studies have investigated brush states that exhibit high antifouling properties for specific water environments via experimental approaches. However, biomolecular adsorption is a complex phenomenon that results from the interaction of external and brush conditions (Fig. 1b), making a comprehensive understanding of the phenomenon difficult. Indeed, there has been little quantitative understanding in this research area, particularly regarding the contribution of brush density and molecular weight.

Fig. 1.	(a) Schematic illustration of the antifouling behavior induced by the zwitterionic polymer brush and the formation of a hydration layer. (b) Factors affecting the properties of molecular adsorption.

　Here, we developed a new machine learning-based approach to quantitatively understand the adsorption behavior of biomolecules. First, we surveyed previous literatures that focused on adsorption phenomena in which brush density and molecular weight were controlled. Adsorption data, which included brush structure (molecular weight, density, thickness, polymer type and substrate properties), solution conditions (pH, ionic strength, temperature, biomolecular charge and molecular weight) and control conditions (flow rate), was then collected and dataset with a total of 125 experimental results were prepared. In machine learning, the dataset can be trained by a computer to predict the amount of adsorption. Here, machine learning was conducted using a total of six linear and non-linear regression methods, and the random forest regression was adopted as the model with the highest prediction accuracy (Fig. 2a). The trained model was further analyzed using the SHapley Additive exPlanations (SHAP) method to quantify the impact of each feature on the adsorption of biomolecules. The results showed that the effect of polymer brush density was the most important, being three times more important than the effect of molecular weight (Fig. 2b).

Fig. 2.	Fig. 2. (a) Prediction performance of machine learning using random forest regression. (b) Estimated importance (mean SHAP value) of considered descriptors.

　Machine learning-aided prediction can also help to understand adsorption properties beyond current experimental facts or extract desirable surface conditions in practical environments. Thus, we further predicted biomolecular adsorption on zwitterionic polymer brushes concerning the molecular weight and density under various water environments. Here, the influence of the ionic strength was investigated. The consideration of the ionic strength is particularly crucial in the application of antifouling porous membranes for water treatment. Generally, groundwater and surface water have millimolar ranges of ionic strength, whereas seawater has approximately 700 mM. Therefore, we examined ionic strengths in the range of 1-1000 mM (Fig. 3). Mapping at 1 mM shows that the density is more effective than the molecular weight to enhance the antifouling properties Especially in the density region above 0.2 chains/cm², the biomolecular adsorption can effectively be inhibited.

Furthermore, in the low ionic strength region, it was suggested that there are two major adsorptive regions (1) and (2). Region (1) is a low-density, low-molecular-weight condition, where the surface is not fully covered with zwitterionic polymer, thus the surface is partially exposed and remains adsorptive even with increment of ionic strength. On the other hand, in region (2), adsorption occurs under low ionic strength conditions, but the adsorption property decreases with increasing ionic strength. This result agreed with previous report that the insertion of biomolecules into the brush is induced by the interaction between the charges of the grafted polymers and the biomolecules.

Fig. 3.	Ionic strength dependence of the mapping protein adsorption. The descriptors are fixed as follows: Sub_Ad = 450 ng cm-2, Pro_Conc = 1 g L-1, and Flow Rate = 0.01 mL min-1.

　To the best of our knowledge, this is the first report to accurately estimate the contribution of density and molecular weight to protein adsorption using a machine learning-based approach. This work quantitatively evaluated the importance of the polymer brush to the antifouling properties, which was empirically not understood until now. Although this study focuses only on the zwitterionic polymers based on a limited number of data sets, the approach may be applicable investigating various brush interfaces and could help for designing future antifouling surfaces.

Journal	：	ACS Applied Materials & Interfaces
Title	：	Machine-Learning-Aided Understanding of Protein Adsorption on Zwitterionic Polymer Brushes
Authors	：	Hiroto Okuyama, Yuuki Sugawara, Takeo Yamaguchi.
URL	：	https://pubs.acs.org/doi/full/10.1021/acsami.4c01401