What question did this study set out to answer?

This study aims to identify determinants affecting protein adsorption and abundance on nanoparticle surfaces using interpretable machine learning.

June 1, 2026Scientific Reports0 citationsOpen Access

Determinants of protein corona adsorption and abundance revealed by interpretable machine learning across nanoparticle systems

KLKeyuan Li ACAlexa Canchola FZFan Zhang

Key Points

This study aims to identify determinants affecting protein adsorption and abundance on nanoparticle surfaces using interpretable machine learning.
Developed large-scale deep neural network models for binary classification of protein adsorption and regression analysis of protein abundance.
Utilized a dataset from 83 studies containing 817 NP–PC samples and 2,497 proteins for model training.
Employed prevalence-based filtering and SHapley Additive exPlanations to interpret model predictions.
Adsorption models achieved high discriminative performance with AUC = 0.96.
Abundance models had a pooled R² of 0.67, and average per-protein R² of 0.40 on the test set.
Key predictors of adsorption include NP material class and surface chemistry, while abundance is influenced by handling and kinetic parameters.

Abstract

Abstract Nanoparticles (NPs) hold significant potential in biotechnology, including molecular sensing, controlled release systems, and therapeutic applications. However, their behavior in biological environments remains difficult to predict because proteins rapidly absorb onto NP surfaces, forming a protein corona (PC) that reshapes their surface properties and determines their biological identity, transport, and cellular interactions. In this study, we developed large-scale deep neural network (DNN) models to predict both protein adsorption (binary classification) and relative protein abundance (regression) on NP surfaces. We utilized a well-curated and comprehensive PC dataset comprising data from 83 peer-reviewed studies, 817 NP–PC samples, and 2,497 proteins, substantially expanding the scale and diversity compared with prior studies. Then, we employed a prevalence-based filtering strategy to mitigate sparsity and batch noise and trained over 200 machine learning models across proteins. The adsorption classification models achieved high discriminative performance (AUC = 0.96), while the abundance models achieved a pooled R² of 0.67 and an average per-protein R 2 of 0.40 on the test set. SHapley Additive exPlanations (SHAP) revealed that adsorption was predominantly governed by NP material class and surface chemistry, whereas abundance was more strongly influenced by experimental handling and kinetic parameters, particularly isolation and incubation time. Incorporation of applicability domain (AD) analysis enabled identification of reliable prediction regions, with in-AD predictions demonstrating higher confidence and reduced error for both tasks. Together, these results demonstrate that our DNN models can identify predictive drivers of PC composition and reveal feature associations consistent with patterns reported in prior mechanistic literature, offering a data-driven reference to inform nanomaterial design for biomedical and environmental applications.

问 AI

Bookmark

View Full Paper