Host Cell Prediction of Exosomes Using Morphological Features on

9 hours ago - Exosomes are extracellular nanovesicles released from any cells and found in any body-fluid. Because exosomes exhibit information of the...
0 downloads 3 Views 3MB Size
Subscriber access provided by Kaohsiung Medical University

B: Biomaterials and Membranes

Host Cell Prediction of Exosomes Using Morphological Features on Solid Surfaces Analyzed by Machine Learning Kazuki Ito, Yuta Ogawa, Keiji Yokota, Sachiko Matsumura, Tamiko Minamisawa, Kanako Suga, Kiyotaka Shiba, Yasuo Kimura, Ayumi Hirano-Iwata, Yuzuru Takamura, and Toshio Ogino J. Phys. Chem. B, Just Accepted Manuscript • DOI: 10.1021/acs.jpcb.8b01646 • Publication Date (Web): 17 May 2018 Downloaded from http://pubs.acs.org on May 17, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Host Cell Prediction of Exosomes Using Morphological Features on Solid Surfaces Analyzed by Machine Learning

Kazuki Ito,1 Yuta Ogawa,1 Keiji Yokota,1 Sachiko Matsumura,2 Tamiko Minamisawa,2 Kanako Suga,2 Kiyotaka Shiba,2 Yasuo Kimura,3 Ayumi Hirano-Iwata,4 Yuzuru Takamura,5 and Toshio Ogino*1,5

1

Yokohama National University, 79-5, Tokiwadai, Hodogaya-ku, Yokohama 240-8501, Japan

2

Japanese Foundation for Cancer Research, 3-8-31 Ariake, Koto-ku, Tokyo, 135-8550 Japan

3

Tokyo University of Technology, 1404-1, Katakura-Cho Hachioji, 192-0914 Japan

4

Tohoku University, 2-1-1, Katahira, Aoba-ku, Sendai, Miyagi, 980-8577, Japan

5

Japan Advanced Institute of Science and Technology, 1-1, Asahi-Dai, Nomi, Ishikawa, 923-1292,

Japan

E-Mail address: [email protected] Telephone Number: +81-045-339-4493

1

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ABSTRACT: Exosomes are extracellular nanovesicles released from any cells and found in any body-fluid. Because exosomes exhibit information of their host cells (secreting cells), their analysis is expected to be a powerful tool for early diagnosis of cancers. To predict the host cells, we extracted multi-dimensional feature data about size, shape, and deformation of exosomes immobilized on solid surfaces by atomic force microscopy (AFM). The key idea is combination of support vector machine (SVM) learning for individual exosome particles and their interpretation by principal component analysis (PCA). We observed exosomes derived from three different cancer cells on SiO2/Si, 3-aminopropyltriethoxysilane-modified-SiO2/Si, and TiO2 substrates by AFM. Then, 14-dimensional feature vectors were extracted from AFM particle data, and classifiers were trained in 14-dimensional space. The prediction accuracy for host cells of test AFM particles was examined by the cross-validation test. As a result, we obtained prediction of exosome host cells with the best accuracy of 85.2% for two-class SVM learning and 82.6% for three-class one. By PCA of the particle classifiers, we concluded that the main factors for prediction accuracy and its strong dependence on substrates are incremental decrease in the PCA-defined aspect-ratio of the particles with their volume.

2

ACS Paragon Plus Environment

Page 2 of 51

Page 3 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

INTRODUCTION

Exosomes, which are extracellular nanovesicles with a diameter of 30 – 150 nm, attract much attention as intercellular communication vehicles, and their characterization is expected to be a new tool for early diagnosis of disease.1,2 Exosomes are released from any cell, including dendritic cells,3 macrophages,4 lymphocytes,5 and tumor cells,6 through exocytosis to the extracellular space,7-9 and always found in any human body fluid, such as saliva,10,11 urine,12,13 plasma,14 breast milk,15,16 and ascite,17 as well as blood. These cell-derived particles contain mRNAs, miRNAs, glycan, and proteins specific to the host cell (secreting cells) and have lipid membranes whose compositions retain characteristics of the host cell membranes.18-20 Because of these properties, exosomes play an important role in long-range and short-range cell-to-cell communication.21-23 Recently, it has been widely accepted that exosomes are deeply involved in disease, cancer metastasis, and diagnosis.24-32 For example, amount of the circulating exosomal proteins released from ovarian cancer patients was larger than that from age-matched women without evidence of ovarian disease, and increased as progression stage of the cancer.33 S. Sharma et al. reported that exosome size in saliva samples from oral cancer patients is generally larger than that from healthy people.34 From these previous works, we can expect that exosome characterization is a promising diagnosis alternative to replace or complement the traditional cancer diagnosis, typically biopsies, which are accompanied with patient pain, long treatment time, and high cost. To establish early

3

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

diagnosis of cancer using exosomes, a highly reliable, high-throughput characterization technique is required. Up to now, several techniques for observation and characterization of exosomes have been developed.35 Transmission electron microscopy (TEM) is the most powerful technique for observation of real shape and size of exosomes.8,32,36,37 However, TEM equipment is expensive and the sample preparation for cryogenic TEM is complicated. Because atomic force microscopy (AFM)34,38-45 and scanning electron microscope (SEM)36 are easy-to-use methods, they have been often used to roughly evaluate number and size of exosomes. Optical measurement systems, such as nanoparticle tracking analysis (NTA),36,41 dynamic light scattering (DLS),36 and single particle interferometric reflectance imaging sensor (SP-IRIS),46 are also used owing to an advantage of no requirement for sample preparation. In these optical methods, however, accurate diameters of small particles are difficult to obtain because they are not based on direct observation of particles but on measurement of particle mobility. Recently, F. Liu et al. developed high-yield, high-purity total chips that isolate exosomes from biofluid according to their size.47 In these techniques, we can obtain statistical data of exosomes, such as size and density, which exhibit differences between healthy people and cancer patients.48,34,41 However, these particle parameters are very scattered and cannot be used for prediction of the host cells of the individual exosomes. Although analyses of the exosome contents, such as proteins and miRNA, would be the most informative way to predict the host cells, it is difficult to analyze individual particles owing to their small size.

4

ACS Paragon Plus Environment

Page 4 of 51

Page 5 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Here, we focus on the biophysical properties of exosomes. I. Parolini et al. reported that biophysical properties of exosome membranes, such as rigidity and fluidity, are controlled by the environmental pH.49 This means that the biophysical properties are important variables that characterize host cells. By using AFM, 3-dimensional, morphological information of individual exosomes can be obtained in buffer solution environment, which is similar to human body fluid, without special sample treatment, such as staining in TEM. Generally, samples should be immobilized on substrates in AFM imaging either by chemical bonding through appropriate linkers or simple physisorption. In the physisorption used in our experiments, the original sample morphology is deformed by immobilization, especially in soft materials observation, and it has been believed to be a disadvantage. However, plenty of information about biophysical properties, such as rigidity of exosome membrane and amount of protein contents, are included in the deformation fashion. Because interaction between the exosome membranes and solid surfaces are also involved in the observed deformation, we can examine various substrates to more clearly discriminate the host cells. Although feature extraction of biophysical properties from the individual exosomes is important for classifying the exosome types, reliable prediction of exosome host cells has been difficult only by use of characterization hardware. In this study, we investigated morphological deformation features of exosomes obtained from AFM images when they are immobilized on solid substrates in liquid environment. In conventional analysis, only one- or two-dimensional parameters, such as exosome volume, have been extracted

5

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

from AFM images. In the present technique, we extract morphological features of the individual immobilized exosomes as multi-dimensional data from the AFM particle images. The multi-dimensional AFM data are analyzed using a support vector machine (SVM), which is one of the data machine learning methods, and principal component analysis (PCA), which is one of the standard methods in multivariate analysis.50-52 In cancer stage prediction, SVM was used to analyze morphological features of cancer cells.53 SVM was also utilized to identify host cells of exosomes by comparison of mRNA expression levels.54 To the best of our knowledge, the present work is the first report that the SVM method was applied to morphological feature analysis of exosomes immobilized on substrate surfaces and that the prediction accuracy of the host cells was examined. Furthermore, we investigated the substrate material effects to optimize the interaction between exosomes and the substrate surfaces from the point of view of highly-reliable host cell prediction assisted by PCA.

EXPERIMENTAL AND ANALYTICAL SECTION

Sample Preparation and Characterization. In this study, 3 types of exosomes derived from cultured cancer cells, HT-29 (colorectal adenocarcinoma), HT-1080 (Fibrosarcoma), and MIA PaCa-2 (pancreatic cancer), were used. From the harvested supernatants, cell debris and the other contaminants were removed using differential centrifugation followed by density gradient centrifugation. The purified exosome-containing fraction was suspended in a phosphate buffered saline (PBS) solution. The exosome solutions were stored at −80°C and thawed before use. 6

ACS Paragon Plus Environment

Page 6 of 51

Page 7 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

To investigate effects of interface properties between exosomes and solid surfaces, we used 3 types of substrate: as-cleaned SiO2/Si, 3-aminopropyltriethoxysilane (APTES)-modified SiO2/Si (APTES/SiO2/Si), and as-cleaned TiO2. APTES, one of the molecules that form self-assembled monolayers (SAMs), modifies a negatively-charged, hydrophilic SiO2/Si surface to positively-charged, hydrophilic one. SiO2/Si substrates of 5 mm square were dipped in a piranha solution, which is a mixture of 98% H2SO4 and 33% H2O2 in a volume ratio of 3:1, for 10 min to remove organic contaminant. Cleaned substrates were then dipped into deionized water and dried in N2 blow. An APTES solution was prepared by diluting as-purchased APTES with dehydrated toluene to 0.02 vol %. To prepare APTES-modified SiO2/Si substrates, the as-cleaned SiO2/Si substrates were immersed in the APTES solution for 2 h, dipped into deionized water, and dried in N2 blow. TiO2 substrates of 5 mm square were cleaned in acetone and ethanol solutions for 5 min each. Then, the cleaned TiO2 substrates were irradiated with ultraviolet light for 2 h to hydrophilize the surface. In the exosome immobilization on various solid surfaces, the substrates were fixed on a bottom of a petri dish using double-sided adhesive tape, and a 10 µl of exosome-containing solution was dropped to the substrates. After exosome immobilization on the substrates for 1 hour, we rinsed the substrates with fresh PBS solutions for 3 times to remove suspending exosomes. Finally, we fulfilled the petri dished with fresh PBS. The exosome-fixed substrate surfaces were observed by the tapping mode of AFM (E-sweep, Hitachi High-Tech Science) in PBS solutions. We used Si-DF3 cantilevers with spring constant of

7

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

1.6 N/m and a tip radius of 10 nm provided from Hitachi High-Tech Science. Data Analysis. In this study we analyzed morphological features of the exosomes that are immobilized on substrates and consequently deformed. Figure 1 shows the outline of our analysis. First, exosome samples derived from cancer cells were immobilized on solid surfaces and observed by AFM (left of Figure 1a). Here, we refer to adsorbates observed in AFM as AFM particles. AFM particles were prescreened by masking (right of Figure 1a). 42-dimensional particle data were extracted from the masked image, but 14 components were used in SVM learning (Figure 1b) as data vectors. The particles were classified into three groups (Figure 1c) according to the shape, and only one type was used for SVM learning. The 14-dimensional data vectors were used t o train classifiers and cross-validation was performed (Figure 1d). Finally, mechanism of discrimination of host cells was analyzed by PCA (Figure 1e).

8

ACS Paragon Plus Environment

Page 8 of 51

Page 9 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Figure 1. Block diagram of data analysis: (a) prescreening of AFM particles, (b) data vector extraction, (c) selection of particles suitable for host cell prediction, (d) SVM learning, and (e) PCA of prediction accuracy. To obtain multi-dimensional morphological data from the deformed exosomes, we used “Gwyddion”, which is a free software for image analysis of scanning probe microscopy including AFM.55 To extract only exosome-originated particle data, we masked individual particles using a masking function of “Gwyddion”, and then excluded the particles located at the peripheries of the AFM image flame owing to their missing areas, as well as apparently complex or contamination particles as the prescreening (Figure 1a). We also excluded particles with heights smaller than 1.5 nm at this stage because they apparently are contaminants. Then, 42 feature data, such as a maximum height of a specific particle, were extracted from the individual masked particles in the AFM images. In the SVM learning for host cell prediction, we selected 14 components for feature vectors from the automatically extracted 42 components. We first removed some parameters because they are related only to positions, which are meaningless for exosome type identification. Then, we calculated correlation coefficients among the parameters. When two or more parameters exhibited strong correlation, we selected one parameter from them as the representative parameter. Table 1 shows the selected feature vector components that were used in the exosome characterization.

9

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 51

Table 1. Selected Components of the 14-Dimensional Feature Vectors component

definition of vector components

Zmax

maximum height value

Zm

mean height value

Zmed

median height value

Bmax

maximum height value on boundary

A0

projected area (2-dimensional area)

As

total surface area of a particle

Ah

area above half-height

Ac

area of convex hull

V

total volume above substrate level

Lb0

projected boundary length

Dmin

minimum boundary size (minimum width)

Dmax

maximum boundary size (maximum width)

Ri

maximum inscribed disc radius

Rm

mean radius (mean distance from the center)

The selected particles in AFM images were automatically classified into three deformation-type groups according to the threshold values of the median height component (Zmed) and the maximum height one (Zmax), as shown in eq. (1). Type I:

7 nm < Zmed

10 nm < Zmax

Type II:

3 nm < Zmed < 7 nm,

10 nm < Zmax

Type III:

3 nm < Zmed < 7 nm,

(1)

Zmax < 10 nm

Because thickness of the lipid bilayers that rap the exosomes is 4-5 nm, particles with Zmed smaller than 3 nm were excluded at this stage because they are contamination or exosome debris. We compared the type classification results performed by “Gwyddion” and the cross-sectional views drawn by manual operation. As a result, it was confirmed that the former results correctly correspond to three different deformation fashions, as described in the results and discussion section. The 10

ACS Paragon Plus Environment

Page 11 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

deformation-type ratio for each host cell was automatically calculated from the type classification results. The exosome diameters (dsph), which correspond to sphere exosomes suspending in solution, were estimated from the immobilized particle volume (V) according to eq. 2. య

ଷ௏

݀௦௣௛ = 2ටସగ

(2)

In our strategy, we first extract 14-dimensional feature vectors from the individual particles in AFM images obtained from an a priori known exosome solution, which are referred to as training vectors, and then plot the training vectors in a 14-dimensional feature space. Host cell for unknown exosome sample, which is referred to as test sample, will be predicted using SVM learning. Figure 2 shows the basic process of SVM learning to assign a test sample to the correct cell group, where we use a 2-dimensional feature space simply consisting of particle height and area instead of the 14-dimensional one to visualize the outline of our strategy. In this simplified model, height and area components are extracted from raw AFM image (Figure 2a,b) and plotted into a two-dimensional feature space (Figure 2c). Then, training data vectors from another training sample derived from different host cells are plotted using the same process. The basic idea behind the SVM approach consists of (A) plotting the training data vectors into a higher dimensional feature space through a linear or non-linear transformation function (which is called kernel) and (B) training hyperplanes that effectively split the training data vectors into two or more classes. In the model process of Figure 2, we are dealing with three-class problem. When the training data vectors are plotted in the feature space, we choose a suitable kernel for obtaining the optimized classifier used in the next step. In

11

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

process (A), we used three kernel functions: linear, polynomial, and radial basis function (RBF) kernels, and the arbitrary parameters included in those kernel functions were automatically determined so that the maximum accuracy in the prediction stage is obtained. The SVM-generated hyperplane is optimal in a sense that it has the largest distance between the hyperplane and the training data vectors

Figure 2. Experimental and analytical procedure of the SVM learning: (a) AFM image of exosome-immobilized substrate, (b) extraction of the training feature vectors, (c) data plots, (d) training of classifiers, and (e) prediction of the host cell of the test exosomes by (f) plotting the test particle feature vectors. Accuracy was examined by cross-validation after whole particle data were extracted and assigned to 10 feature vector subsets.

closest to the hyperplane, which is called support vectors. In the feature space, hyperplanes divide 12

ACS Paragon Plus Environment

Page 12 of 51

Page 13 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

the training data vectors from three different host cells into three classes (Figure 2d) using the SVM learning. Now, we have a SVM classifier for three class problem, as shown in Figure 2e. Host cells of the test samples are discriminated by the space where the feature vectors of the test particles (Figure 2f) are plotted (Figure 2e). In the present SVM learning, we used the e1071 package (a free software)56 implemented in the R language. The SVM function in the e1071 provides interfaces with the open source LIBSVM software, which is a library for SVM learning developed by C-C. Chang and C-J. Lin.57 Among the used SVM kernels (linear, polynomial, and RBF), the RBF kernel, which is the most widely used kernel in SVM learning, gave the largest accuracy in most cases.

Figure 3. Schematics of k-fold cross-validation. Performance for host cell prediction in our system was evaluated by k-fold cross-validation, which is a widely used method for prediction accuracy evaluation based on statistical analysis. Figure 3 schematically shows outline of our cross-validation procedure. The particle data vectors extracted from AFM images of whole host cancer cells are divided into k subsets with a same vector number, which are referred to as Subi ( i = 1, …., k ). In the first training, data vectors in the Subi (i = 2, …., k) 13

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

were used as the training vectors to train a classifier, and the data vectors in Sub1 were used as a test sample set. Rates of correct assignment of the individual test data vectors to the correct host cells are calculated. In the second training, data vectors in the Subi (i = 1, 3,…., k) are used as the training vectors and Sub2 as a test vector subset. In the i-th training ( i = 1, …., k ), all subsets except Subi were used for the training classifier, and Subi for validation of prediction, followed by calculation of the correct assignment rate. Thus, training and prediction were repeated k-times and correct assignment rates for whole trainings are obtained. In the multivariate analysis, PCA is often used to visualize the multidimensional data vectors by their projection in a low-dimensional feature space. In this method, new variables introduced by linear combination of the original vector components are used as new coordinates instead of the original ones with minimum loss of information. This approach is powerful to analyze the primary factors that determine the prediction accuracy. Two-dimensional variables, PC1 and PC2, were generated by R language, an open-source software for data analysis. Throughout our SVM learning and PCA analysis, we used raw AFM data without any data processing, such as mean-centering or scaling.

RESULTS AND DISCUSSION

Statistical Approach

14

ACS Paragon Plus Environment

Page 14 of 51

Page 15 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Figure 4. AFM images of exosomes derived from (a) HT-29, (b) HT-1080, and (c) MIA PaCa-2 on SiO2/Si substrates, scale bar = 1 µm. (d) Example of type classification, scale bar = 0.5 µm.

Figure 4 shows AFM images of particles derived from the HT-29, HT-1080, and MIA PaCa-2 exosomes on SiO2/Si substrate surfaces. Before SVM learning using feature data vectors, we separated the immobilized particles into three types according to the criteria of eq. (1): an isolated hemispherical particle (Type I), a hemispherical particle on a flat membrane (Type II), and an isolated single bilayer (Type III). Figure 4d shows an example of type classification result where particles are drawn in three-dimensions using “Gwyddion”, and the numbers (1, 2, and 3) on the individual particles represent Types I, II, and III, respectively. Figure 5a,c,e show examples of AFM 3-dimensional images of Type I, Type II, and Type III particles, respectively, and Figure 5b,d,f their cross-sectional views, respectively. In Figure 5a,b, height of the particle is about 20 nm and the 15

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

lateral diameter about 150 nm. This means that the particles were laterally expanded by an attractive force between the particle surface and the substrate one, resulting in flattened semispherical shape. In Figure 5b,d, the height of the protrusion and the flat base are about 15 nm and 5 nm, respectively. Because exosome membranes are lipid bilayers with thickness of about 5 nm, flat base of the particle is a ruptured exosome. Apparently, Type II particle should be a complex consisting of a ruptured exosome (lipid bilayer) and another deformed one. In Figure 5c,f, height of the flat membrane is about 5 nm, which is similar to a single bilayer thickness. This means that Type III

Figure 5. AFM 3-dimensional images of (a) Type I, (c) Type II, and (e) Type III particles, and (b,d,f) their cross-sectional views, respectively.

particles formed through adsorption, expansion, and rupture of exosomes. Every deformation type is 16

ACS Paragon Plus Environment

Page 16 of 51

Page 17 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

observed in every type of host cells on every surface used in the experiments. Many studies on adsorption of artificial lipid vesicles showed that deformation behavior, including expansion and rupture, depends on the lipid molecules that wrap the vesicles.58,59

Figure 6. Deformation type ratios on SiO2/Si substrates for exosomes derived from HT-29, HT-1080, and MIA PaCa-2.

Figure 6 shows the deformation type ratios on SiO2/Si substrates for exosomes derived from three host cells. HT-1080 exhibit the highest Type I ratio and HT-29 the lowest Type I ratio. MIA PaCa-2 exhibits the highest Type III ratio and HT-1080 the lowest Type III ratio. Rigidity of exosome membranes is closely related to efficiency in exosome release from cells or their fusion to another cells.49 It means that exosome rigidity is one of the important properties that represents exosome features. The membrane rigidity of exosomes is generally determined by their lipid composition, amount of membrane proteins, and that of skeletal proteins,41 and exosomes released from different cells in different environment have also different lipid composition and membrane proteins.49,60 Deformation of exosomes immobilized on solid surfaces is basically determined by the membrane rigidity, as demonstrated in artificial lipid vesicles.58,59 Therefore, deformation type ratio reflects the

17

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

difference in the membrane rigidity. The ratio of Type I and Type III is especially important because this ratio represents frequency of expansion and rupture of the immobilized exosomes. However, in addition to membrane rigidity, deformation of artificial vesicles depends on their size and their concentration even on the same substrate.59 Owing to dispersion in size and concentration of exosomes, as described in the next paragraph, it is difficult to determine the rigidity of the membrane only by the deformation types.

Figure 7. Diameter distributions of AFM particles derived from (a) HT-29, (b) HT-1080, and (c) MIA PaCa-2 on SiO2/Si substrates. Figure 7 shows diameter (dsph) distribution calculated by eq. (2) for exosomes derived from three

18

ACS Paragon Plus Environment

Page 18 of 51

Page 19 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

different host cells, and Table 2 their mean and median values. The values of dsph of almost all particles are in a range of 30-150 nm, which is widely accepted as a definition of exosome size. In Table 2, the mean and median dsphs obtained from the AFM images exhibit definite differences. Up to now, we have many reports on the size measurement of exosomes and proposals of their application to cancer diagnosis.27 For example, size distribution of exosomes obtained from oral cancer patient saliva shifts to larger dsph values compared with that of healthy donors.34,41 Thus, the size measurement of exosomes has been expected to be promising for future low-invasive diagnosis. In this study, we can also confirm definite differences in the size distribution of exosomes derived from different cancer cells, suggesting that we may obtain information about the type of the host cells and apply it for early diagnosis only by exosome characterization.

Table 2. Mean and Median Values of the Calculated Diameters for Exosomes Derived from Three Different Host Cells host cell

mean (nm)

median (nm)

HT-29

57.8

47.1

HT-1080

72.0

55.6

MIA PaCa-2

51.1

44.5

However, we can compare only mean or median values of the measured exosomes and cannot obtain meaningful information about the individual exosomes because their size dispersion is much larger than the difference in the mean or median values between exosomes from different host cells, as shown in Figure 7. We have to develop a novel method that enables us to predict the host cells of the individual exosomes by simple, quick processes without expensive equipment. 19

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 20 of 51

Host Cell Prediction Using SVM Learning. We performed 2-class SVM learning for HT-1080 and Mia PaCa-2 on SiO2/Si, APTES/SiO2/Si, and TiO2 substrates and 3-class SVM learning for HT-29, HT-1080, and Mia PaCa-2 on SiO2/Si and TiO2 substrates. We could not examine 3-class SVM on APTES/SiO2/Si owing to shortage of sample amount. Although the ratios of the deformation types shown in Figure 6 are useful data for host cell prediction, we used only particles grouped to Type I, which were unruptured, simply deformed particle group. This is because the ruptured exosomes in the Type III have lost information about their contents and induce errors in size when membrane fusion occurred. Unruptured particles can be found in the Type II, but they were also excluded because they are placed on lipid bilayers, not immobilized directly on the substrates. In the SVM learning, particle numbers of each host cell were adjusted to the same by excluding excess particles using a random number function. Accuracy values of 10-fold cross-validation were obtained after automatically optimizing the parameters in the kernel functions, and accuracy is defined by percentage of test feature vectors that were correctly discriminated by the classifier. Prediction results based on the 10-fold cross-validation are shown in Table. 3. We also found that the results of accuracy do not strongly depend on the k-values (number of division) in our test.

Table 3. Host Cell Prediction Accuracy Obtained by SVM Classifiers host cells

substrate

total particle number

accuracy

HT-1080/MIA PaCa-2

SiO2/Si

444

77.9%

HT-1080/MIA PaCa-2

APTES/SiO2/Si

478

83.3%

HT-1080/MIA PaCa-2

TiO2

250

85.2%

HT-29/HT-1080/MIA PaCa-2

SiO2/Si

666

60.1%

HT-29/HT-1080/MIA PaCa-2

TiO2

375

82.6%

20

ACS Paragon Plus Environment

Page 21 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

The prediction accuracy depends on the substrates. Table S1 shows the surface chemistry of the substrate used in this study. In the two-class prediction for HT-1080 and MIA PaCa-2 in Table 3, the accuracy on SiO2 substrates, 77.9%, was the lowest and that on TiO2 surfaces, 85.2%, was the highest. We examined validity of the total sample numbers used for the present prediction. From each data set of HT-1080 and MIA PaCa-2, feature vectors fewer than the total numbers used in Table 3 were randomly chosen, and the accuracy of SVM learning for the 2-class prediction was examined on SiO2/Si, APTES/SiO2/Si, and TiO2 substrates. Figure S1 shows dependence of prediction accuracy on the number of feature vectors. Although the accuracy monotonically increases with increase in the number of the feature vectors, their increment increase in accuracy was not large when the number of data vectors is larger than 100. Moreover, in Figure S1, the accuracy order is always TiO2, APTES/SiO2/Si, and SiO2/Si in descending order. Therefore, validity of the accuracy order with respect to substrates in Table 3 can be verified in spite of the different total particle numbers. Our cross-validation results were affected by errors owing to tip deformation and other tip changes such as its apex contamination because we did not consider them. Better prediction results may be obtained by modeling the tip history effects. Because exosomes have soft, rather smooth surfaces, tip history effect may be small. However, when our method is applied to harder, rougher materials than exosomes, tip effect should be included in the SVM learning. In the three-class SVM learning, exosomes derived from three kinds of cancer cells (HT-29,

21

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

HT-1080, and MIA PaCa-2) were immobilized on the substrates, and the host cells were predicted in the same process as the two-class SVM. The accuracy with respect to substrates are 82.6% on the TiO2 and 60.1% on the SiO2 surfaces, as shown in Table 3. When using the total particle number of 375 for a SiO2 substrate, the prediction accuracy was 55.2%. Because this particle number produced 82.6% accuracy for TiO2, superiority of TiO2 substrates to SiO2 ones in prediction accuracy is verified also in the three-class SVM learning. To discuss the determinant factors involved in the prediction accuracy, we introduced PCA. In this technique, we reduced the dimension of the feature vectors to 2-dimensional space and visualized the data plots and the classifiers as scatter plots. Figure 8a,b,c shows scatter plots of the feature vectors obtained from particles on SiO2/Si, APTES/SiO2/Si, and TiO2 substrates, respectively, in the 2-class SVM learning for HT-1080 and MIA PaCa-2, where the vertical axis is the first principal component (PC1) and the horizontal axis the second principal component (PC2). The PC1 and PC2 are composed of linear combination of 14-dimensional components so that the largest dispersion in the individual feature vectors is obtained. In Figure 8, circular-blue points and triangular-red points represent feature vectors of the particles derived from HT-1080 and MIA PaCa-2, respectively. Blue-colored region and red-colored one are separated by the hyperplanes trained by the training feature vectors, and discriminated as HT-1080 and MIA PaCa-2, respectively. Factor loadings of the PC1 and the PC2 attached to the scatter plots in Figure 8 were automatically obtained by calculation using D language. Because factor loadings are individually calculated for each set of scatter plots and

22

ACS Paragon Plus Environment

Page 22 of 51

Page 23 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

classifiers, PC1 and PC2 as the coordinates in Figure 8a,b,c are not exactly the same, and quantitative comparison among these plots are meaningless. However, by using PCA of the data plots, we can discuss determinant factors to obtain higher accuracy in the exosome host cell prediction. From the factor loadings in Figure 8a,b,c, it is obvious in any PCA plot that the PC1 contains positive components associated with the particle height group (Zmax, Zm, Zmed) and the contact area group between individual particles and the substrate (A0, As, Ah, Ac). These factor loadings indicate that PC1 is strongly related to particle volumes. The PC2 Figure 8. Interpretation of experimentally-obtained prediction accuracy based on PCA: (a) discrimination between HT-1080 and MIA PaCa2 on (a) SiO2/Si, (b) APTES/SiO2/Si, and (c) TiO2/Si substrates. contains negative components of the particle height group and the positive ones of the contact area group, or vice versa. It means that PC2 is more related to the aspect ratio. Values of the factor loadings related to the height group in this system are much larger than those to the contact area group. To maximize the dispersion in PC2, weight of height factors should be enhanced compared with the contact area factors because increment increase in particle height with volume increase is smaller than that of contact area owing to deformation fashion, as shown in Figure 5a,b. The signs of the factor loadings of the height-group values and contact-area-group ones in PC2 often switch, as shown in Figure 8a,c. It can occur because PC2 is derived under a condition that PC2 is perpendicular to PC1. In our discussion, we only focus on the absolute values by ignoring which sign is assigned to the height-group components. Small particle plots always form a cluster with the highest plot density in any scatter plots because particle dimeters at the particle-number distribution 23

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure 8. Interpretation of experimentally-obtained prediction accuracy based on PCA: (a) discrimination between HT-1080 and MIA PaCa2 on (a) SiO2/Si, (b) APTES/SiO2/Si, and (c) TiO2/Si substrates.maximum are around

30 nm and the number distribution is extended to the larger diameter side, as clearly shown in Figure 7. Before analyzing individual data plots by PCA, we summarize what scatter plots express in the PC coordinate. PC1 simply represents a particle volume derived under the condition of maximum dispersion, and relatively larger particles are plotted in the upper (larger) side with respect to PC1. PC2 is related to aspect ratio, but not exactly aspect ratio of the real particle. We refer to the aspect ratio defined by PC2 as PC2-aspect-ratio. The real aspect ratio of exosomes immobilized on solid surfaces decreases as the volume become larger. PC2-aspect-ratio represents incremental decrease in aspect ratio as particle volume increases, as shown in the top illustration in Figure 9a,b,c. When incremental decrease in PC2-aspect-ratio with increase in the volume is relatively small, particles are plotted in the positive side with respect to PC2 when factor loadings of the height-related-group have positive sign. When incremental decrease in PC2-aspect-ratio is relatively large, which corresponds to larger lateral expansion of the particles, the particles are plotted in the negative side. In the factor loading system of negative height-group loadings, plots with respect to PC2 are horizontally flipped. Now, we discuss how the host cells are predicted with high accuracy from AFM morphology of the particles immobilized on the substrates. In the shape simulation of vesicles adsorbed on solid surfaces without rupture, vesicle deformation depends on lipid composition,48 vesicle size61 and

24

ACS Paragon Plus Environment

Page 24 of 51

Page 25 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

membrane proteins.62 Lipid composition dependence of deformation of artificial lipid vesicles was also experimentally demonstrated using AFM.63 Exosomes have their own lipid composition, membrane proteins, and skeletal proteins that are originated in their host cells and, therefore, exhibit their own rigidity. Moreover, exosomes contain proteins and miRNAs specific to the host cells, and their amount also varies with cancer progress.30 These contents also influence deformation fashion, resulting in difference in the 14-dimensional feature vectors. We have known that particle size distribution is similar among the host cells used in this work. From SVM learning of exosomes, we can obtain plenty of information such as volume dependence of rigidity to predict their host cell, and we can visualize the primary reason why they are discriminated by SVM learning. In Figure 8a, both particle plots are distributed in wide range of PC1, which means both volume dispersion are similar. Compared with MIA PaCa-2, the paticles from HT-1080 exhibits relatively negative shift with respect to PC2 when PC1 increases. Because a vector plot with a larger PC2 value corresponds to a particle with smaller increment decrease in aspect ratio when its volume is enlarged. Some particles from MIA PaCa-2 tends to have larger PC2 values with increase in PC1. This means that some exosomes from MIA PaCa-2 are more rigid than those from HT-1080. However, rest of the particles from MIA PaCa-2 exhibit a behavior similar to HT-1080 exosomes, which causes prediction errors. Moreover, small particles indicated by an arrow have small dispersion with respect to PC2. This is the reason why SiO2/Si substrate yields the lowest prediction accuracy among three two-class predictions in Figure 8. In Figure 8b, particles from HT-1080 on APTES/SiO2/Si /Si has a behavior

25

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

similar to that in Figure 8a, whereas particles from MIA PaCa-2 exhibit simultaneous decrease with respect to PC2 with increase in particle volume represented as increase in PC1. Generally, when volume of vesicles immobilized on solid surfaces increases, their shape is more flattened by lateral expansion compared with small vesicles, resulting in lower aspect ratio, as indicated by the previous work.57 The behavior of MIA PaCa-2 in Figure 8b can be interpreted by a rapid decrease in the aspect ratio, as represented by decrease in PC2, with increase in their volume. In this system, incremental decrease in the aspect ratio for MIA PaCa-2 is clearly larger than that for HT-1080, and prediction accuracy higher than that in SiO2/Si substrates was obtained. In Figure 8c, PC2 values for both particles from HT-1080 and MIA PaCa-2 tend to increase with PC1 value. Note that Figure 8c is drawn by the factor loading system of negative height-related-group and positive area-related-group. When particle volume increases, both contact area and height increase. If the height increase is relatively small, increase in PC2 will be larger, as drawn in the illustration, because the corresponding incremental increase in PC2 consists of an increase in the area-related components and relatively small incremental increase in height-related components. It is obvious that, when particle volume (PC1) increases, the incremental increase in PC2 for MIA PaCa-2 is almost always larger than that for HT-1080. Therefore, it is concluded that the highest prediction accuracy was obtained in this system because MIA PaCa-2 is rapidly flattened as the particle volume increases. Similar results were obtained in the three-class SVM learning for HT-1080, MIA PaCa-2, and HT-29, as shown in Figure S2. When using SiO2/Si substrates, as shown in Figure S2a, dispersion of

26

ACS Paragon Plus Environment

Page 26 of 51

Page 27 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

PC1-dependent PC2 values is small in the small particle region (small PC1 values), which is similar to Figure 8a, whereas dispersion is large for plots of large particle volume (large PC1 values). This means that three types of exosomes exhibit similar shape change for their volume increase, resulting in poor prediction accuracy. When using TiO2 substrates, as shown in Figure S2b, flattening behavior of HT-29 particles with respect to volume increase is comparatively large, that of MIA PaCa-2 intermediate, and that of HT-1080 small. Here, Figure S2b is drawn by the factor loading system of negative height-related-group and positive area-related-group. Because flattening effects clearly discriminate these three regions, the prediction accuracy is much higher than that on SiO2/Si substrates. Note that the prediction is achieved in 14-dimensional space, and PCA is only used to interpret the discrimination basis, not used in the prediction. Thus, in the host cell prediction, one of the determinant factors has been found to be the incremental decrease in the PC2-aspect-ratio. For the PC2-aspect-ratio or the deformation caused by immobilization on solid surfaces, we have two main factors. One is the rigidity of exosomes, and the other exosome/solid interaction, as clearly demonstrated in Figure 8 and Figure S2. Next, we discuss origin of substrate dependence of the prediction accuracy. Figure 9 shows the scatter plots when exosomes from HT-1080 cells were immobilized on three types of substrates. In Figure 9, circular-blue points and blue-colored region, triangular-red points and red-colored region, and crossed-green points and green-colored region represent feature vectors of the particles on SiO2/Si, TiO2/Si, and APTES/SiO2/Si substrates, respectively. These plots well demonstrate that exosomes

27

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

immobilized on TiO2 exhibit large incremental increase with respect to PC2, which can be interpreted by that the smallest incremental decrease in aspect ratio with increase in the volume is realized on TiO2. In other words, lateral expansion of exosomes is comparatively suppressed on TiO2. Interaction between lipid membranes and solid substrates have been intensively studied.58,64-67 Lipid bilayer formation on TiO2 surfaces was hard by the vesicle fusion method

Figure 9. Scatter plots for the HT-1080 exosomes on SiO2/Si, APTES/SiO2/Si, and TiO2/Si substrates. whereas it was easy on SiO2 and mica surfaces.66 Because exosome surfaces are negatively charged,68 surface charge of the substrate surface also affects the deformation fashion. Because interaction force between HT-1080 exosomes and positively charged APTES surfaces should be high, HT-1080 exosomes are most easily flattened on APTES/SiO2/Si substrates, as shown in Figure 9. In addition to lipid membrane type and exosome size, interaction between the substrate surfaces and membrane proteins of exosomes would be involved in the deformation fashion. In this sense, our 28

ACS Paragon Plus Environment

Page 28 of 51

Page 29 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

method is an empirical one, but even if we do not know the concrete membrane composition and protein species, we can predict the host cell by replacing the unknown materials with multi-dimensional deformation data.

CONCLUSIONS

In this study, we have proposed a new method to predict the host cell of exosomes using multi-dimensional data vectors about size, shape, and deformation caused by immobilization on solid substrates. The key idea is combination of SVM learning for extracted data vectors and their interpretation by PCA. We observed exosomes derived from HT-29, HT-1080, and MIA PaCa-2 immobilized on SiO2/Si, APTES/SiO2/Si, and TiO2 substrates by AFM in buffer solution environment. Deformation types were automatically classified into three groups using the criteria about particle height, and the deformation type ratio in each exosome are obtained. We also found that mean values of exosomes released from different cells differ from each other. However, these statistical data are not powerful to discriminate host cells of individual particles. We then introduced SVM learning for 14-dimensional feature vectors extracted from AFM particle images, and classifier was trained in 14-dimensional space using the data plots. The prediction accuracy for new particles was examined by the cross-validation test, and the host cells of two kinds or three kinds of exosomes were predicted. As a result, we succeeded in prediction of the exosome host cells with the best accuracy of 85.2% for two-class SVM learning and 82.6% for three-class one. We found that the prediction accuracy strongly depends on the substrates. To investigate primary factors that determine the prediction accuracy, we introduced PCA and found that particle-volume dependence of incremental decrease in the aspect ratio is the most determinant factor, where the aspect ratio and the volume are defined by linear combination of the 14 factor loadings. Validity of this method is biophysically verified because exosome features, such as their membrane rigidity and membrane 29

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

protein species, can be displayed as size-dependent deformation fashion with large dispersion. Because AFM is now an easy-to-use, high-throughput tool, the present technique is promising for early diagnosis for disease.

Supporting Information Available: Dependence of prediction accuracy on number of feature vectors in cross-validation test; surface properties of the substrate surfaces; PCA in the three-class SVM learning for HT-1080, MIA PaCa-2, and HT-29.

Acknowledgments

This work was partly supported by a Grant-in-Aid for Scientific Research (15K13361) from MEXT and the CREST program of the Japan Science and Technology Agency (JPMJCR14F3). The authors thank Prof. Sugimoto for helpful discussion about multivariate analysis and SVM learning.

Author Information Corresponding Author *[email protected] ORCID Toshio Ogino: 0000-0002-3241-5652 Present address The Instrumental Analysis Center, Yokohama National University, 79-5, Tokiwadai, Hodogaya-ku, Yokohama, 240-8501, Japan. Notes The authors declare no competing financial interest. 30

ACS Paragon Plus Environment

Page 30 of 51

Page 31 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

REFERENCES

(1) Vlassov, A. V.; Magdaleno, S.; Setterquist, R.; Conrad, R. Exosomes: Current Knowledge of their Composition, Biological Functions, and Diagnostic and Therapeutic Potentials. Biochimica et Biophysica Acta 2012,1820, 940–948. (2) Raposo, G.; Stoorvogel, W. Extracellular Vesicles: Exosomes, Microvesicles, and Friends. J. Cell Biol. 2013, 200, 373–383. (3) Lamparski, H. G.; Metha-Damani, A.; Yao, J.-Y.; Patel, S.; Hsu, D.-H.; Ruegg, C.; Le Pecq J.-B. Production and Characterization of Clinical Grade Exosomes Derived from Dendritic Cells. J. Immunol. Methods 2002, 270, 211–226. (4) Bhatnagar, S.; Shinagawa, K.; Castellino, F. J.; Schorey, J. S. Exosomes Released from Macrophages Infected with Intracellular Pathogens Stimulate a Proinflammatory Response in Vitro and in Vivo. Blood 2007, 110, 3234–3244. (5) Escola, J.-M.; Kleijmeer, M. J.; Stoorvogel, W.; Griffith, J. M.; Yoshie, O.; Geuze H. J. Selective Enrichment of Tetraspan Proteins on the Internal Vesicles of Multivesicular Endosomes and on Exosomes Secreted by Human B-Lymphocytes. J. Biol. Chem. 1998, 273, 20121–20127. (6) Iero, M.; Valenti, R.; Huber, V.; Filipazzi, P.; Parmiani, G.; Fais, S.; Rivoltini, L. Tumour-Released Exosomes and their Implications in Cancer Immunity. Cell Death Differ. 2008, 15, 80–88.

31

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(7) Heijnen, H. F. G.; Schiel, A. E.; Fijnheer, R.; Geuze, H. J.; Sixma, J. J. Activated Platelets Release Two Types of Membrane Vesicles: Microvesicles by Surface Shedding and Exosomes Derived from Exocytosis of Multivesicular Bodies and Alpha-Granules. Blood. 1999, 94, 3791– 3799. (8) Van Der Pol, E.; Hoekstra, A. G.; Sturk, A.; Otto, C.; Van Leeuwen, T. G.; Nieuwland, R. Optical and Non-Optical Methods for Detection and Characterization of Microparticles and Exosomes. J. Thromb. Haemost. 2010, 8, 2596–2607. (9) Gruenberg J.; van der Goot, F. G. Mechanisms of Pathogen Entry through the Endosomal Compartments. Nat. Rev. Mol. Cell Biol. 2006, 7, 495–504. (10) Michael, A.; Bajracharya, S. D.; Yuen, P. S. T.; Zhou, H.; Star, R. A.; Illei, G. G.; Alevizos, I. Exosomes from Human Saliva as a Source of MicroRNA Biomarkers. Oral Dis. 2010, 16, 34–38. (11) Ogawa, Y.; Kanai-Azuma, M.; Akimoto, Y.; Kawakami, H.; Yanoshita, R. Exosome-Like Vesicles with Dipeptidyl Peptidase IV in Human Saliva. Biol. Pharm. Bull. 2008, 31, 1059–1062. (12) Nilsson, J.; Skog, J.; Nordstrand, A.; Baranov, V.; Mincheva-Nilsson, L.; Breakefield, X. O.; Widmark, A. Prostate Cancer-Derived Urine Exosomes: a Novel Approach to Biomarkers for Prostate Cancer. Br. J. Cancer 2009, 100, 1603–1607. (13) Pisitkun, T.; Shen, R.-F.; Knepper, M. A. Identification and Proteomic Profiling of Exosomes in Human Urine. Proc. Natl. Acad. Sci. 2004, 101, 13368–13373.

32

ACS Paragon Plus Environment

Page 32 of 51

Page 33 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

(14) Orozco, A. F.; Jorgez, C. J.; Ramos-Perez, W. D.; Popek, E. J.; Yu, X.; Kozinetz, C. A.; Bischoff, F. Z.; Lewis, D. E. Placental Release of Distinct DNA-Associated Micro-Particles into Maternal Circulation: Reflective of Gestation Time and Preeclampsia. Placenta 2009, 30, 891–897. (15) Lässer, C.; Alikhani, V. S.; Ekström, K.; Eldh, M.; Paredes, P. T.; Bossios, A.; Sjöstrand, M.; Gabrielsson, S.; Lötvall, J.; Valadi, H. Human Saliva, Plasma and Breast Milk Exosomes Contain RNA: Uptake by Macrophages. J. Transl. Med. 2011, 9, 9. (16) Admyre, C.; Johansson, S. M.; Qazi, K. R.; Filén, J.-J.; Lahesmaa, R.; Norman, M.; Neve, E. P. A.; Scheynius, A.; Gabrielsson, S. Exosomes with Immune Modulatory Features are Present in Human Breast Milk. J. Immunol. 2007, 179, 1969–1978. (17) Andre, F.; Schartz, N. E. C.; Movassagh, M.; Flament, C.; Pautier, P.; Morice, P.; Pomel, C.; Lhomme, C.; Escudier, B.; Le Chevalier, T.; et al. Malignant Effusions and Immunogenic Tumour-Derived Exosomes. Lancet 2002, 360, 295–305. (18) Valadi, H.; Ekström, K.; Bossios, A.; Sjöstrand, M.; Lee, J. J.; Lötvall, J. O. Exosome-Mediated Transfer of mRNAs and MicroRNAs is a Novel Mechanism of Genetic Exchange between Cells. Nat. Cell Biol. 2007, 9, 654–659. (19) Batista, B. S.; Eng, W. S.; Pilobello, K. T.; Hendricks-Muñoz, K. D.; Mahal, L. K. Identification of a Conserved Glycan Signature for Microvesicles. J Proteome Res. 2011, 10, 4624–4633. (20) van den Boorn, J. G.; Daßler, J.; Coch, C.; Schlee, M.; Hartmann, G. Exosomes as Nucleic Acid Nanocarriers. Adv. Drug Deliv. Rev. 2013, 65, 331–335.

33

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(21) Kosaka, N.; Iguchi, H.; Yoshioka, Y.; Takeshita, F.; Matsuki, Y.; Ochiya, T. Secretory Mechanisms and Intercellular Transfer of MicroRNAs in Living Cells. J. Biol. Chem. 2010, 285, 17442–17452. (22) Pegtel, D. M.; Cosmopoulos, K.; Thorley-Lawson, D. A.; van Eijndhoven, M. A. J.; Hopmans, E. S.; Lindenberg, J. L.; de Gruijl, T. D.; Würdinger, T.; Middeldorp, J. M. Functional Delivery of Viral MiRNAs via Exosomes. Proc. Natl. Acad. Sci. 2010, 107, 6328–6333. (23) Camussi, G.; Deregibus, M. C.; Bruno, S.; Cantaluppi V.; Biancone, L. Exosomes/Microvesicles as a Mechanism of Cell-to-Cell Communication. Kidney Int. 2010, 78, 838–848. (24) Wolfers, J.; Lozier, A.; Raposo, G.; Regnault, A.; Théry, C.; Masurier, C.; Flament, C.; Pouzieux, S.; Faure, F.; Tursz, T.; et al. Tumor-Derived Exosomes are a Source of Shared Tumor Rejection Antigens for CTL Cross-Priming. Nat. Med. 2001, 7, 297–303. (25) Al-Nedawi, K.; Meehan, B.; Micallef, J.; Lhotak, V.; May, L.; Guha, A.; Rak, J. Intercellular Transfer of the Oncogenic Receptor EGFRvIII by Microvesicles Derived from Tumour Cells. Nat. Cell Biol. 2008, 10, 619–624. (26) Svensson, K. J.; Kucharzewska, P.; Christianson, H. C.; Sköld, S.; Löfstedt, T.; Johansson, M. C.; Mörgelin, M.; Bengzon, J.; Ruf, W.; Belting, M. Hypoxia Triggers a Proangiogenic Pathway Involving Cancer Cell Microvesicles and PAR-2-Mediated Heparin-Binding EGF Signaling in Endothelial Cells. Proc. Natl. Acad. Sci. 2011, 108, 13147–13152.

34

ACS Paragon Plus Environment

Page 34 of 51

Page 35 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

(27) Grange, C.; Tapparo, M.; Collino, F.; Vitillo, L.; Damasco, C.; Deregibus, M. C.; Tetta, C.; Bussolati, B.; Camussi, G. Microvesicles Released from Human Renal Cancer Stem Cells Stimulate Angiogenesis and Formation of Lung Premetastatic Niche. Cancer Res. 2011, 71, 5346–5356. (28) Zitvogel, L.; Regnault, A.; Lozier, A.; Wolfers, J.; Flament, C.; Tenza, D.; Ricciardi-Castagnoli, P.; Raposo, G.; Amigorena, S. Eradication of Established Murine Tumors Using a Novel Cell-Free Vaccine: Dendritic Cell Derived Exosomes. Nat. Med. 1998, 4, 594–600. (29) Bobrie, A.; Krumeich, S.; Reyal, F.; Recchi, C.; Moita, L. F.; Seabra, M. C.; Ostrowski, M.; Théry, C. Rab27a Supports Exosome-Dependent and -Independent Mechanisms that Modify the Tumor Microenvironment and can Promote Tumor Progression. Cancer Res. 2012, 72, 4920–4930. (30) Peinado, H.; Alečković, M.; Lavotshkin, S.; Matei, I.; Costa-Silva, B.; Moreno-Bueno, G.; Hergueta-Redondo, M.; Williams, C.; García-Santos, G.; Ghajar, C.; et al. Melanoma Exosomes Educate Bone Marrow Progenitor Cells toward a Pro-Metastatic Phenotype through MET. Nat. Med. 2012, 18, 883–891. (31) Valenti, R.; Huber, V.; Filipazzi, P.; Pilla, L.; Sovena, G.; Villa, A.; Corbelli, A.; Fais, S.; Parmiani, G.; Rivoltini, L. Human Tumor-Released Microvesicles Promote the Differentiation of Myeloid Cells with Transforming Growth Factor-β-Mediated Suppressive Activity on T Lymphocytes. Cancer Res. 2006, 66, 9290–9298.

35

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(32) Lee, E.-Y.; Park, K.-S.; Yoon, Y. J.; Lee, J.; Moon, H.-G.; Jang, S. C.; Choi, K.-H.; Kim, Y.-K.; Gho, Y. S. Therapeutic Effects of Autologous Tumor-Derived Nanovesicles on Melanoma Growth and Metastasis. PLoS One 2012, 7, e33330. (33) Taylor, D. D.; Gercel-Taylor, C. MicroRNA Signatures of Tumor-Derived Exosomes as Diagnostic Biomarkers of Ovarian Cancer. Gynecol. Oncol. 2008, 110, 13–21. (34) Sharma, S.; Gillespie, B.; Palanisamy, V.; Gimzewski, J. K. Quantitative Nanostructural and Single-Molecule Force Spectroscopy Biomolecular Analysis of Human-Saliva-Derived Exosomes. Langmuir 2011, 27, 14394–14400. (35) Varenne, F.; Makky, A.; Gaucher-Delmas, M.; Violleau, F.; Vauthier, C. Multimodal Dispersion of Nanoparticles: A Comprehensive Evaluation of Size Distribution with 9 Size Measurement Methods. Pharm. Res. 2016, 33, 1220–1234. (36) Chernyshev, V. S.; Rachamadugu, R.; Tseng, Y. H.; Belnap, D. M.; Jia, Y.; Branch, K. J.; Butterfield, A. E.; Pease III, L. F.; Bernard, P. S.; Skliar, M. Size and Shape Characterization of Hydrated and Desiccated Exosomes. Anal. Bioanal. Chem. 2015, 407, 3285–3301. (37) Choi, H.; Mun, J. Y. Structural Analysis of Exosomes Using Different Types of Electron Microscopy. Appl. Microsc, 2017, 47, 171–175. (38) Yuana, Y.; Oosterkamp, T. H.; Bahatyrova, S.; Ashcroft, B.; Rodriguez, P. G.; Bertina, R. M.; Osanto, S. Atomic Force Microscopy: A Novel Approach to the Detection of Nanosized Blood Microparticles. J. Thromb. Haemost. 2010, 8, 315–323.

36

ACS Paragon Plus Environment

Page 36 of 51

Page 37 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

(39) Hardij, J.; Cecchet, F.; Berquand, A.; Gheldof, D.; Chatelain, C.; Mullier, F.; Chatelain, B.; Dogné, J.-M. Characterisation of Tissue Factor-Bearing Extracellular Vesicles with AFM: Comparison of Air-Tapping-Mode AFM and Liquid Peak Force AFM. J. Extracell. Vesicles 2013, 2, 21045. (40) Marton, A.; Vizler, C.; Kusz, E.; Temesfoi, V.; Szathmary, Z.; Nagy, K.; Szegletes, Z.; Varo, G.; Siklos, L.; Katona, R. L.; et al. Melanoma Cell-Derived Exosomes Alter Macrophage and Dendritic Cell Functions in Vitro. Immunol. Lett. 2012, 148, 34–38. (41) Zlotogorski-Hurvitz, A.; Dayan, D.; Chaushu, G.; Salo, T.; Vered, M. Morphological and Molecular Features of Oral Fluid-Derived Exosomes: Oral Cancer Patients versus Healthy Individuals. J. Cancer Res. Clin. Oncol. 2016, 142, 101–110. (42) Sharma, S.; Rasool, H. I.; Palanisamy, V.; Mathisen, C.; Schmidt, M.; Wong, D. T.; Gimzewski, J. K. Structural-Mechanical Characterization of Nanoparticle Exosomes in Human Saliva, Using Correlative AFM, FESEM, and Force Spectroscopy. ACS Nano 2010, 4, 1921–1926. (43) Parisse, P.; Rago, I.; Severino, L. U.; Perissinotto, F.; Ambrosetti, E.; Paoletti, P.; Ricci, M.; Beltrami, A. P.; Cesselli, D.; Casalis, L. Atomic Force Microscopy Analysis of Extracellular Vesicles. Eur. Biophys. J. 2017, 46, 813–820. (44) Et-Thakafy, O.; Delorme, N.; Gaillard, C.; Mériadec, C.; Artzner, F.; Lopez, C.; Guyomarc’h, F. Mechanical Properties of Membranes Composed of Gel-Phase or Fluid-Phase Phospholipids Probed on Liposomes by Atomic Force Spectroscopy. Langmuir 2017, 33, 5117−5126.

37

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(45) Sebaihi, N.; de Boeck, B.; Yuana, Y.; Nieuwland, R.; Pétry, J. Dimensional Characterization of Extracellular Vesicles Using Atomic Force Microscopy. Meas. Sci. Technol. 2017, 28, 034006. (46) Daaboul, G. G.; Gagni, P.; Benussi, L.; Bettotti, P.; Ciani, M.; Cretich, M.; Freedman, D. S.; Ghidoni, R.; Ozkumur, A. Y.; Piotto, C.; et al. Digital Detection of Exosomes by Interferometric Imaging. Sci. Rep. 2016, 6, 37246. (47) Liu, F.; Vermesh, O.; Mani, V.; Ge, T. J.; Madsen, S. J.; Sabour, A.; Hsu, E.-C.; Gowrishankar, G.; Kanada, M.; Jokerst, J. V.; et al. The Exosome Total Isolation Chip. ACS Nano 2017, 11, 10712– 10723. (48) Dimitrievski, K.; Kasemo, B. Simulations of Lipid Vesicle Adsorption for Different Lipid Mixtures. Langmuir 2008, 24, 4077–4091. (49) Parolini, I.; Federici, C.; Raggi, C.; Lugini, L.; Palleschi, S.; Milito, A. D.; Coscia, C.; Iessi, E.; Logozzi, M.; Molinari, A.; et al. Microenvironmental pH is a Key Factor for Exosome Traffic in Tumor Cells. J. Biol. Chem. 2009, 284, 34211–34222. (50) Jesse, S.; Kalinin, S. V. Principal Component and Spatial Correlation Analysis of Spectroscopic-Imaging Data in Scanning Probe Microscopy. Nanotechnology 2009, 20, 085714. (51) Ziatdinov, M.; Maksov, A.; Li, L.; Sefat, A. S.; Maksymovych, P.; Kalinin, S. V. Deep Data Mining in a Real Space: Separation of Intertwined Electronic Responses in a Lightly Doped BaFe2As2. Nanotechnology 2016, 27, 475706.

38

ACS Paragon Plus Environment

Page 38 of 51

Page 39 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

(52) Yamanishi, J.; Iwase, S.; Ishida, N.; Fujita, D. Multivariate Analysis for Scanning Tunneling Spectroscopy Data. Appl. Surf. Sci. 2018, 428, 186-190. (53) Glotsos, D.; Kalatzis, I.; Spyridonos, P.; Kostopoulos, S.; Daskalakis, A.; Athanasiadis, E.; Ravazoula, P.; Nikiforidis, G.; Cavouras, D. Improving Accuracy in Astrocytomas Grading by Integrating a Robust Least Squares Mapping Driven Support Vector Machine Classifier into a Two Level Grade Classification Scheme. Comput. Methods Programs Biomed. 2008, 90, 251–261. (54) Ko, J.; Bhagwat, N.; Yee, S. S.; Ortiz, N.; Sahmoud, A.; Black, T.; Aiello, N. M.; McKenzie, L.; O’Hara, M.; Redlinger, C.; et al. Combining Machine Learning and Nanofluidic Technology to Diagnose Pancreatic Cancer Using Exosomes, ACS Nano 2017, 11, 11182–11193. (55) Nečas, D.; Klapetek, P. Gwyddion: An Open-Source Software for SPM Data Analysis. Cent. Eur. J. Phys. 2012, 10, 181–188. (56) Meyer, D.; Dimitriadou, E.; Hornik, K.; Weingessel, A.; Leisch, F.; Chang, C.-C.; Lin, C.-C. Package ‘e1071’ 2015, 1–62. (57) Chang, C.-C.; Lin, C.-J. LIBSVM : A Library for Support Vector Machines. ACM Trans. Intell. Syst. Technol. 2011, 2,3, 27. (58) Keller, C. A.; Kasemo, B. Surface Specific Kinetics of Lipid Vesicle Adsorption Measured with a Quartz Crystal Microbalance. Biophys. J. 1998, 75, 1397–1402. (59) Reviakine, I.; Brisson, A. Formation of Supported Phospholipid Bilayers from Unilamellar Vesicles Investigated by Atomic Force Microscopy. Langmuir 2000, 16, 1806–1815.

39

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(60) Meckes Jr., D. G.; Raab-Traub, N. Microvesicles and Viral Infection. J. Virol. 2011, 85, 12844– 12854. (61) Dimitrievski, K. Deformation of Adsorbed Lipid Vesicles as a Function of Vesicle Size. Langmuir 2010, 26, 3008–3011. (62) Dimitrievski, K. Influence of Lipid-Bilayer-Associated Molecules on Lipid-Vesicle Adsorption. Langmuir 2010, 26, 5706–5714. (63) Nakano, K.; Tozuka, Y.; Yamamoto, H.; Kawashima, Y.; Takeuchi, H. A Novel Method for Measuring Rigidity of Submicron-Size Liposomes with Atomic Force Microscopy. Int. J. Pharm. 2008, 355, 203–209. (64) Tero, R.; Watanabe, H.; Urisu, T. Supported Phospholipid Bilayer Formation on Hydrophilicity-Controlled Silicon Dioxide Surfaces. Phys. Chem. Chem. Phys. 2006, 8, 3885–3894. (65) Nabika, H.; Fukasawa, A.; Murakoshi, K. Control of the Structure of Self-Spreading Lipid Membrane by Changing Electrolyte Concentration. Langmuir 2006, 22, 10927–10931. (66) Tero, R.; Ujihara, T.; Urisu, T. Shape Transformation of Adsorbed Vesicles on Oxide Surfaces: Effect of Substrate Material and Photo-Irradiation. Trans. Mater. Res. Soc. Jpn. 2009, 34, 183–188. (67) Isono, T.; Tanaka, H.; Ogino, T. Effect of Chemical Modification of the Substrate Surface on Supported Lipid Bilayer Formation. e-J. Surf. Sci. Nanotechnol. 2007, 5, 99–102.

40

ACS Paragon Plus Environment

Page 40 of 51

Page 41 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

(68) Sokolova, V.; Ludwig, A.-K.; Hornung, S.; Rotan, O.; Horn, P. A.; Epple, M.; Giebel, B. Characterisation of Exosomes Derived from Human Cells by Nanoparticle Tracking Analysis and Scanning Electron Microscopy. Colloids Surf. B: Biointerfaces 2011, 87, 146–150.

41

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

TOC GRAPHIC

42

ACS Paragon Plus Environment

Page 42 of 51

Page 43 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

308x199mm (96 x 96 DPI)

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

270x189mm (150 x 150 DPI)

ACS Paragon Plus Environment

Page 44 of 51

Page 45 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

212x209mm (96 x 96 DPI)

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

328x262mm (96 x 96 DPI)

ACS Paragon Plus Environment

Page 46 of 51

Page 47 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

371x134mm (149 x 149 DPI)

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

71x114mm (150 x 150 DPI)

ACS Paragon Plus Environment

Page 48 of 51

Page 49 of 51 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

192x283mm (150 x 150 DPI)

ACS Paragon Plus Environment

The Journal of Physical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

350x167mm (150 x 150 DPI)

ACS Paragon Plus Environment

Page 50 of 51