Single-cell dispensing and \u2018real-time\u2019 cell classification using convolutional neural networks for higher efficiency in single-cell cloning

Julian Riba,Jonas Schoendube,Peter Koltay,Roland Zengerle,Stefan Zimmermann

doi:10.1038/s41598-020-57900-3

Abstract

Single-cell dispensing for automated cell isolation of individual cells has gained increased attention in the biopharmaceutical industry, mainly for production of clonal cell lines. Here, machine learning for classification of cell images is applied for ‘real-time’ cell viability sorting on a single-cell printer. We show that an extremely shallow convolutional neural network (CNN) for classification of low-complexity cell images outperforms more complex architectures. Datasets with hundreds of cell images from four different samples were used for training and validation of the CNNs. The clone recovery, i.e. the fraction of single-cells that grow to clonal colonies, is predicted to increase for all the samples investigated. Finally, a trained CNN was deployed on a c.sight single-cell printer for ‘real-time’ sorting of a CHO-K1 cells. On a sample with artificially damaged cells the clone recovery could be increased from 27% to 73%, thereby resulting in a significantly faster and more efficient cloning. Depending on the classification threshold, the frequency at which viable cells are dispensed could be increased by up to 65%. This technology for image-based cell sorting is highly versatile and can be expected to enable cell sorting by computer vision with respect to different criteria in the future.

Highlights

Single-cell dispensing for automated cell isolation of individual cells has gained increased attention in the biopharmaceutical industry, mainly for production of clonal cell lines
fluorescent-activated cell sorting (FACS) can be used to gate for viable cells based on forward (FSC) and side scatter (SSC) or on fluorescent viability dyes such as propidium iodide or 7AAD, which are DNA intercalating substances that do not pass through intact cell membranes and only label dead cells, or CMFDA or Calcein AM which are membrane-permeable substances that are turned over to fluorescent products by viable cells
The convolutional neural network (CNN)-based classification for viability prediction was implemented into the existing software for cell detection and controlling of the instrument (c.sight single-cell printer, cytena GmbH, Germany) without any hardware changes

Summary

Introduction

Single-cell dispensing for automated cell isolation of individual cells has gained increased attention in the biopharmaceutical industry, mainly for production of clonal cell lines. Based on early regulatory guidelines released by the US federal drug administration (FDA) and other regulatory bodies, the production cell line for a recombinant product is to be cloned from a single progenitor cell[2] This is required in order to minimize population heterogeneity and facilitate isolation and subsequent selection of high producing clones, which could be otherwise overgrown by fast growing but low producing clones. In practice, cell samples often contain significant fractions of cells that are dead or are difficult to grow, which can result in low clone recovery It is not uncommon in industrial cell line development workflows, that only 20% of the isolated single cells grow to usable colonies[7]. Instead of synchronizing the algorithm with a fixed dispensing frequency, the droplet generation is continuously adapted to the timing of the image processing step

Methods

Results

Conclusion