Multidimensional mechano-tomography of biological cells: novel modes and machine learning data analysis

New multidimensional AFM modalities, RingingMode and FT-NanoDMA allow collecting images of physical properties of cell surfaces and dynamical mechanical properties of the cell body for multiple frequencies simultaneously. These images show the distributions of nonspecific adhesion, energy losses due to the probe disconnection, size of the molecules covering the cell surface, viscoelastic properties of the cell membrane, storage, loss, and static elastic moduli of the cell body, etc. Repeating imaging at different load force/depth, it is possible to build “mechano-tomography” of the cell body. FT-NanoDMA allows simultaneous recording of more than 20 dynamical mechanical parameters, whereas RingingMode gives eight additional channels of information presenting the physical properties of the cell surface.

This wealth of information can effectively be analyzed using a machine learning approach to define cell phenotype and its specific stage.  “Classical” deep learning methods, however, are not directly applicable to AFM imaging because they require an unrealistically large number of images for training. This is technically impossible because AFM is a relatively slow technique to obtain the required number of images. I will describe our solution which is based on a substantial reduction of the dimension of the data space by using so-called surface parameters. An example of this approach will be described. The steps to produce sufficient statistics using bootstrap methods will be overviewed. An important problem of overtraining will be discussed. Examples of applications of this method will be given.