Quantitation of Suicidal Netosis Using Convolutional Neural Networks in Human Neutrophils

Abstract

Suicidal NETosis (Neutrophil Extracellular Traps) is a mode of cell death involving de-condensation of chromatin and fusion with cytoplasmic granule proteins to create a net-like structure for entrapping pathogenic elements. Despite its name, NETosis does not occur exclusively in neutrophils but rather in all myeloid cells of the immune system except dendritic cells. A growing list of diseases and conditions implicate NETs in their etiology and the progression of symptoms. These conditions include sepsis, cystic fibrosis, lupus nephritis, acute respiratory distress syndrome (ARDS) and deep vein thrombosis. There are several different quantitative methods to assess NETosis making comparative analysis across the field of NETosis research extremely challenging. To further define NETosis and facilitate new discoveries in this field, we have developed a novel assay that combines high-content microscopy with Convolutional Neural Networks (CNN). CNN is the revolutionary image recognition technology used in popular social media facial features recognition systems to 1) quantify the percentage of NETotic and non-NETotic nuclei in a cell population, 2) classify NETotic, necrotic and normal nuclei in a mixed population of neutrophils and 3) differentiate between two different NETotic nuclei generated by two of the most commonly used agonists in NETosis research, the PKC activator PMA and A23187, a Ca+2 ionophore. Furthermore we have optimized neutrophil isolation and culture conditions to yield reproducible results with low inter-and intra-assay variability. Image acquisition was performed on adherent neutrophils using the automated BD Pathway Bio-imaging Systems with autofocusing function that allows imaging of approximately 10% of a well in a 96-well plate. An algorithm was created for image segmentation and recognition of NETotic nuclei based on size, DNA staining intensity and pattern. Data generated by this semi-automated analysis system were used to train a CNN to count NETotic and non-NETotic nuclei. These methods allowed us to calculate the EC50 for PMA (2.2 nM) and A23187 (2.5 uM) and to differentiate between their distinct mechanisms of NETosis. CNNs are able to classify NETotic phenotypes, without the use of expensive antibodies or reagents. Our assay represents a crucial first step for standardizing and quantitating NETosis, which will allow better data interpretation across different labs and provide a potential screening tool for known and as yet undiscovered NETosis agonists. DisclosuresNo relevant conflicts of interest to declare.