By performing biomechanical measurements on cells using an automated atomic force microscope and analyzing the results with machine learning tools, an interdisciplinary team from the LAAS-CNRS and Restore laboratories has successfully classified hundreds of cells with a high success rate. These results are published in
ACS Applied Materials and Interfaces.
Healthy cells have different mechanical properties than pathological cells. Biomechanical measurements performed with an atomic force microscope (AFM) have shown that it is possible to distinguish cancer cells from normal cells. The results of these measurements could therefore be used for diagnosis, provided that AFM technology is capable of operating at high throughput.
To meet this challenge, a team from
LAAS-CNRS in partnership with the mechanobiology division of the
Restore laboratory (CNRS/French Blood Establishment/Inserm/Université Toulouse Paul Sabatier) has designed an automated AFM biomechanical measurement device that performs a large number of measurements in a limited time.
The first step involves immobilizing the cells on a microstructured chip. Biomechanical measurements are then performed automatically, thanks to control software that moves the AFM from one cell to another. To standardize the measurements, parameters such as the distribution of cells on the chip, the geometry of the AFM probe, and the movement speed were optimized and fixed. With this device, the LAAS team was able to measure nearly a thousand cells in two hours, whereas with a standard AFM, an entire day is needed to measure only a few dozen cells.
For each cell, the AFM records 16 force curves (variations in force as a function of the distance between the AFM probe and the cell wall), which allows the measurement and calculation of seven mechanical characteristics relevant for classifying healthy or cancerous cells. The large amount of data thus collected (over 100,000 characteristics) then allows the use of machine learning techniques to discriminate between healthy and cancerous cells.
Left: cells organized on fibronectin micropatterns, the insert at the bottom right shows a zoom on a cell. The AFM lever is the black triangle, the probe is located at the lower end of the triangle.
Middle, fluorescence image of a co-culture of cancer cells (in green) and healthy cells (the others). This process allows verification of the algorithm's predictions.
Right, result of the classification by the LAMDA fuzzy logic algorithm, trained to differentiate healthy cells and cancer cells.
© Étienne Dague
Using an artificial intelligence tool based on a fuzzy logic algorithm developed at LAAS, machine learning and system testing were performed on non-malignant and cancerous prostate cell lines, and then on non-malignant and cancerous skin fibroblast cell lines.
The tests demonstrated the device's ability to correctly classify 73% of the cells. Depending on the classification thresholds set, the system will produce more or fewer false positives (healthy cell classified as pathological) or false negatives (cancerous cell classified as healthy). This is why the researchers emphasize that tuning will need to be done, with clinicians, depending on the intended application (diagnosis, chemotherapy monitoring...).
The LAAS team continues to develop the system, testing other machine learning algorithms to improve the rate of accurate classifications. Another project is dedicated to the discrimination of pancreatic cancer cells. In addition, a new application is under study: quality control of mesenchymal stem cells intended for tissue regeneration in partnership with Restore.