Deep learning assisted mechanotyping of individual cells through repeated deformations and relaxations in undulating channels.

Mechanical properties of cells are important features that are tightly regulated and are dictated by various pathologies. Deformability cytometry allows for the characterization of the mechanical properties at a rate of hundreds of cells per second, opening the way to differentiating cells via mechanotyping. A remaining challenge for detecting and classifying rare sub-populations is the creation of a combined experimental and analysis protocol that approaches the maximum potential classification accuracy for single cells. In order to find this maximum accuracy, we designed a microfluidic channel that subjects each cell to repeated deformations and relaxations and provides a comprehensive set of mechanotyping parameters. We track the shape dynamics of individual cells with high time resolution and apply sequence-based deep learning models for feature extraction. In order to create a dataset based solely on differing mechanical properties, a model system was created with treated and untreated HL60 cells. Treated cells were exposed to chemical agents that perturb either the actin or microtubule networks. Multiple recurrent and convolutional neural network architectures were trained using time sequences of cell shapes and were found to achieve high classification accuracy based on cytoskeletal properties alone. The best model classified two of the sub-populations of HL60 cells with an accuracy over 90%, significantly higher than the 75% we achieved with traditional methods. This increase in accuracy corresponds to a fivefold increase in potential enrichment of a sample for a target population. This work establishes the application of sequence-based deep learning models to dynamic deformability cytometry.