| Literature DB >> 32332167 |
Zhan Ban1, Peng Yuan1, Fubo Yu1, Ting Peng1, Qixing Zhou1, Xiangang Hu2.
Abstract
Protein corona formation is critical for the design of ideal and safe nanoparticles (Entities:
Keywords: cellular recognition; machine learning; nano-bio interface; nanotoxicity; protein corona
Mesh:
Substances:
Year: 2020 PMID: 32332167 PMCID: PMC7229677 DOI: 10.1073/pnas.1919755117
Source DB: PubMed Journal: Proc Natl Acad Sci U S A ISSN: 0027-8424 Impact factor: 11.205
Fig. 1.Overview of the qualitative and quantitative factors (652 data pieces). The superscript “a” represents that the other types of qualitative factors are shown in the . The superscript “b” represents that the factors are measured in the same culture with the same NP concentration The overall dataset is listed in the . The first to third quartiles were used to describe the data distribution. DMEM, Dulbecco’s modified Eagle medium; HP, human plasma; PDI, polydispersity index; PEG, polyethylene glycol; PS, polystyrene.
Fig. 2.Model performance of RF on protein corona compositions classified by physicochemical and functional properties and selected individual proteins. (A) Tenfold cross-validation was utilized to estimate the model prediction accuracy. (B–H) Each point or (I) each box represents one model built from a nine-tenth dataset, estimated by the correlation coefficient (R2). The model performance evaluated by the rootRMSE of physicochemical and functional properties and individual proteins is shown in the . The specific 178 proteins are listed in the .
Fig. 3.The similarity network visualizes heterogeneity of functional corona composition in models. Each node represents a data piece in the prediction models of functional corona compositions. The nodes are colored according to the priority factors, NP without modification (A–D) and surface modification (E–H). For well-performed RF models, the values of connected nodes are more than four times higher than the average in each proximity matrix. Tighter connections in each cluster indicate the higher homogeneity of nodes for the factor-response dependence learned by RF models. In contrast, the sparse connections represent the heterogeneity of nodes in terms of the NP properties and experimental conditions in the cluster. The similarity network of functional composition models for protein clusters is shown in . The full names of abbreviations are listed in .
Fig. 4.Experiments evaluating the functional and physicochemical composition predictions of protein coronas. (A–D) TEM micrograph representing the morphology and size of Fe3O4, Fe3O4-CIT, Au-COOH, and Au-NH2 NPs. The observed functional components of the protein coronas on the four NPs were measured by liquid chromatography tandem mass spectrometry. The model prediction accuracy was measured by the correlation coefficient (R2) and between observations and predictions of the functional (E, apolipoproteins; F, coagulation proteins; G, complement proteins; and H, immune proteins) and physicochemical (I and J) compositions of selected NPs. The heatmap visualizes the model performance for compositions classified by physicochemical properties, as measured by R2 (I) and RMSE (J). The model evaluations of clusterin and other proteins are shown in .
Fig. 5.Individual protein composition predictions evaluated by experiments. Correlation coefficients (R2 in A) and RMSE (B) of the predictions (C) and observations (D) evaluating the model performance regarding 71 individuals adsorbed on selected NPs (Fe3O4, Fe3O4-CIT, Au-COOH, and Au-NH2). The RPA values were normalized to the maximum value in each prediction or observation as the RPA index. The model performance and details for proteins are shown in .