Label-free and rapid mechanics of single cells under high-density co-culture conditions by deep learning image recognition-assisted atomic force microscopy

Abstract

Mechanical cues play an important role in regulating cellular activities. Cells are able to sense and respond to the mechanical cues present in the extracellular physical microenvironment via mechanotransduction, which can ultimately shape the functions and behaviors of the cells themselves as well as their microenvironments during numerous developmental, physiological and pathological processes. The development of human diseases such as cancer is generally accompanied by unique changes in the mechanical properties of cells and their physical microenvironments, and discoveries in the field of physical oncology are beginning to be translated into new therapeutic strategies for cancer. Delineating the mechanical properties of biological tissues in various dimensions from individual cells to organs is therefore fundamental for dissecting the mysteries of life and advancing human healthcare. In particular, atomic force microscopy (AFM)-based force spectroscopy has become a powerful, standard and multifunctional toolbox for characterizing the various mechanical properties of single cells at the micro/nanoscale. However, current studies of AFM-based single-cell mechanical measurements rely mainly on the experience of the experimental operator to move the AFM probe to the target cells for subsequent force measurements, which often results in a time-consuming and laborious experimental process. In addition, cell coculture has been widely used in the field of life sciences to examine intercellular interactions. Nevertheless, in current cell coculture studies, cells are commonly labelled with fluorescent molecules so that one can visually identify the specific cell types in the coculture, which can affect the behaviors of the fluorescently labelled cells. Consequently, developing a method that allows AFM to measure the mechanical properties of cells under coculture conditions in an efficient and fluorescence-independent way will significantly benefit the applications of AFM in the field of mechanobiology. Previously, we presented a method based on the combination of AFM and deep learning optical image recognition, which can precisely move the AFM probe to individual targeted cells to perform mechanical measurements under low-density co-culture conditions (nearly no contact between different cell types in the co-culture). Here, we present a study of deep learning image recognition-assisted AFM to rapidly probe the mechanical properties of single living cells grown in high-density co-culture conditions (with different cell types in contact with each other in the co-culture) without the need for fluorescent labelling. In this work, AFM experiments were performed with a commercial JPK NanoWizard AFM (Bruker, Santa Barbara, USA), which was mounted on an inverted optical microscope (Nikon, Tokyo, Japan). Three types of cells, MGC-803 (a human gastric cancer cell line), HGC-27 (a human undifferentiated gastric cancer cell line), and HMrSV5 (a human peritoneal mesothelial cell line), were used. All three types of cells were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin at 37°C (5% CO2 and 95% air) in Petri dishes. During the experiments, the RPMI-1640 medium was replaced by CO2-independent Leibovitz’s L-15 medium, and the AFM experiments were performed at 37°C (the commercial AFM used here has a heater system). MGC-803 cells (stained with the DiI dye) were co-cultured with HGC-27 cells (stained with the DiO dye). Both optical bright-field images and corresponding fluorescent images of co-cultured cells were recorded. The YOLOX deep learning neural network was used for directly recognizing cell types from optical bright-field images. The fluorescent images were used to assist in the preparation of the training datasets (Supplementary Figure S1) and to verify the detection results of the deep learning image recognition model. In a previous study under low-density co-culture conditions, we reduced the complexity of the YOLOX neural network to improve the detection speed without reducing the detection accuracy. Under high-density co-culture conditions, where cell recognition becomes much more difficult, we found that reducing the complexity of the YOLOX model resulted in decreased detection accuracy. We examined the recognition performances of four YOLO series neural networks (YOLOX, YOLOv5, YOLOv7, and YOLOv8), and the experimental results revealed that the YOLOX model had the highest detection precision (89.25%) (Supplementary Table S1) and the best detection result (Supplementary Figure S2) and could meet the experimental requirements. Hence, the YOLOX neural network was used here. The AFM spherical probe (a microsphere attached to the tipless cantilever) was used in the indentation assay to measure the Young’s modulus of the cells, and the AFM single-cell probe (a living HMrSV5 cell attached to the tipless cantilever) was used in the single-cell force spectroscopy (SCFS) assay to measure the adhesion force of the cells. More experimental details (e.g., cell sample

1. Introduction

Mechanical cues play an important role in regulating cellular activities. Cells are able to sense and respond to the mechanical cues present in the extracellular physical microenvironment via mechanotransduction, which can ultimately shape the functions and behaviors of the cells themselves as well as their microenvironments during numerous developmental, physiological and pathological processes. The development of human diseases such as cancer is generally accompanied by unique changes in the mechanical properties of cells and their physical microenvironments, and discoveries in the field of physical oncology are beginning to be translated into new therapeutic strategies for cancer. Delineating the mechanical properties of biological tissues in various dimensions from individual cells to organs is therefore fundamental for dissecting the mysteries of life and advancing human healthcare.

In particular, atomic force microscopy (AFM)-based force spectroscopy has become a powerful, standard and multifunctional toolbox for characterizing the various mechanical properties of single cells at the micro/nanoscale. However, current studies of AFM-based single-cell mechanical measurements rely mainly on the experience of the experimental operator to move the AFM probe to the target cells for subsequent force measurements, which often results in a time-consuming and laborious experimental process. In addition, cell coculture has been widely used in the field of life sciences to examine intercellular interactions. Nevertheless, in current cell coculture studies, cells are commonly labelled with fluorescent molecules so that one can visually identify the specific cell types in the coculture, which can affect the behaviors of the fluorescently labelled cells. Consequently, developing a method that allows AFM to measure the mechanical properties of cells under coculture conditions in an efficient and fluorescence-independent way will significantly benefit the applications of AFM in the field of mechanobiology.