Predicting Protein Corona Formation on Polylactic Acid Microplastics Pre- and Post-Photoaging: The Importance of Optimal Imputation Methods

Abstract

Micro-nanoplastics (MNPs) enter biological systems, forming a protein corona (PC) by adsorbing proteins from bodily fluids, influencing their biological effects. Mass spectrometry-based proteomics characterizes PC composition, and recent advances have leveraged protein amino acid sequence-derived features to predict PC formation using a supervised random forest (RF) classifier. However, mass spectrometry often generates substantial missing values (MVs), which may hinder the model’s predictive performance and the understanding of protein–particle interactions. This study assessed the impact of 20 imputation methods on RF classifier performance in predicting human plasma PC formation on polylactic acid (PLA) and photoaged PLA microplastics (MPs), considering their rising ecological and health concerns. The results showed that five left-censored imputation methods (Zero, Half-min, Min, QRILC, GSimp) achieved the best performance, with high accuracy (0.80–0.82), AUC (0.78–0.84), precision (0.78–0.80), and recall (0.97–0.98). Protein spatial features, including secondary sheet structure (negative) and absolute solvent-accessible area (positive), were identified as key factors influencing protein adsorption onto MPs. Additionally, UV aging increased the importance ranking of features frac_aa_S and fraction_exposed_exposed_S, highlighting altered protein–MPs interactions, likely through hydrogen bonding and electrostatic forces. This study demonstrates the potential of left-censored imputation methods in enhancing RF classifier performance for predicting PC formation.