[19p-222-1~12] 12.6 Nanobiotechnology

[19p-222-3] Host Cell Prediction of Exosomes by Machine Learning of AFM Images and its Principal Component Analysis

Kazuki Ito¹, Chika Sugimoto¹, Kiyotaka Shiba², Ayumi Hirano-Iwata³, Yuzuru Takamura⁴, 〇Toshio Ogino^1,4 (1.Yokohama Natl. Univ., 2.Found. Cancer Res., 3.Tohoku Univ., 4.JAIST)

Keywords:machine learning, Exosome, atomic force microscopy

Exosomes are extracellular nanovesicles with diameters of 30 - 150 nm, and attract much interest as new disease markers for early diagnosis. We extracted 14-dimensional data vectors from atomic force microscopy (AFM) images of individual exosomes. The key idea toward host cell prediction is a combination of support vector machine (SVM) learning for exosome particles and their interpretation by principal component analysis (PCA). Prediction accuracy for unknown particles was examined by the cross-validation test. We succeeded in host cell discrimination with the best accuracy of 85.2% for two-class SVM learning and 82.6% for three-class one. The discrimination accuracy strongly depends on the substrate types. The primary factors for the accuracy were analyzed by PCA, and we found that size-dependent deformation fashion is the important factor.