Deep learning neural networks highly predict early onset of mouse and human pluripotent stem cell differentiation from light microscopy images

ARIEL WAISMAN; ALEJANDRO DAMIÁN LA GRECA; ALAN MIQUEAS MÖBBS; MARIA AGUSTINA SCARAFIA; NATALIA LUCÍA SANTÍN VELAZQUE; GABRIEL NEIMAN; LUCIA NATALIA MORO; CARLOS DANIEL LUZZANI; GUSTAVO EMILIO SEVLEVER; ALEJANDRA GUBERMAN; SANTIAGO GABRIEL MIRIUKA

Los Angeles

Congreso; International Society for STem Cell Research; 2019

ISSCR

Pluripotent stem cell (PSC) differentiation is a highly dynamic process in which both epigenetic, transcriptional and metabolic changes eventually lead to new cell identities. There modifications occur whithin hours to days and are generally identified by measuring gene expression changes and protein markers. PSC differentiation is also followed by important morphological transformations, but these are inherently subjective and thus are not used as a standard and quantitative measurement of cell differentiation. Our goal in this work was to use artificial intelligence techniques to automatically classify PSCs from early differentiating cells based on their morphology. For that, we made use of convolutional neural networks (CNNs), powerful algorithms that are particularly useful in computer vision. We induced differentiation of mouse embryonic stem cells (mESCs) to apiblast-like cells (EpiLCs) and took transmitted light microscopy images at several time-points from the initial stimulus. We found that several network architectures can be trained to recognize differentiationg from undifferentiation cells and correctly classify images with an accuracy higher than 99%. Successful prediction started only 20 minutes after the onset of EpiLC differentiation. WE also show that CNNs can be succesfully trained to predict whether mESCs were cultured in standard serum + LIF medium or in the recently developed defined culture conditions using MEK and GSK3 inhibitors plus LIF, that sustain the ground state of pluripotency. Moreover, these algorithms also display great performance in the classification of undifferentiated human induced PSCs (hiPSCs) compared to hiPSCs-derived early mesodermal cells. Although high training accuracy required strong computational power and the use of hundreds of images for each condition, once the CNN was trained it allowed to rapidly and accurately classify the query images. We believe that efficient cellular morphology recognition in a simple microscopic set up may have a significant impact on how cell assays are performed in the near future, ranging from experimental biology to quality control of cell cultures for the eventual application of PSCs to the clinic.