Towards image-based cancer cell lines authentication using deep neural networks

Deogratias Mzurikwao,Mark Wass,Gianluca Marcelli,Martin Michaelis,Muhammad Usman Khan,Jindrich Cinatl,Oluwarotimi Williams Samuel,Chee Siang Ang

doi:10.1038/s41598-020-76670-6

Abstract

Although short tandem repeat (STR) analysis is available as a reliable method for the determination of the genetic origin of cell lines, the occurrence of misauthenticated cell lines remains an important issue. Reasons include the cost, effort and time associated with STR analysis. Moreover, there are currently no methods for the discrimination between isogenic cell lines (cell lines of the same genetic origin, e.g. different cell lines derived from the same organism, clonal sublines, sublines adapted to grow under certain conditions). Hence, additional complementary, ideally low-cost and low-effort methods are required that enable (1) the monitoring of cell line identity as part of the daily laboratory routine and 2) the authentication of isogenic cell lines. In this research, we automate the process of cell line identification by image-based analysis using deep convolutional neural networks. Two different convolutional neural networks models (MobileNet and InceptionResNet V2) were trained to automatically identify four parental cancer cell line (COLO 704, EFO-21, EFO-27 and UKF-NB-3) and their sublines adapted to the anti-cancer drugs cisplatin (COLO-704rCDDP1000, EFO-21rCDDP2000, EFO-27rCDDP2000) or oxaliplatin (UKF-NB-3rOXALI2000), hence resulting in an eight-class problem. Our best performing model, InceptionResNet V2, achieved an average of 0.91 F1-score on tenfold cross validation with an average area under the curve (AUC) of 0.95, for the 8-class problem. Our best model also achieved an average F1-score of 0.94 and 0.96 on the authentication through a classification process of the four parental cell lines and the respective drug-adapted cells, respectively, on a four-class problem separately. These findings provide the basis for further development of the application of deep learning for the automation of cell line authentication into a readily available easy-to-use methodology that enables routine monitoring of the identity of cell lines including isogenic cell lines. It should be noted that, this is just a proof of principal that, images can also be used as a method for authentication of cancer cell lines and not a replacement for the STR method.

Highlights

Short tandem repeat (STR) analysis is available as a reliable method for the determination of the genetic origin of cell lines, the occurrence of misauthenticated cell lines remains an important issue
As most of the available pre-trained models for transfer learning are trained on nonmedical images, we proposed a multi-stage transfer learning through an inter-mediate training step using breast cancer cell images, to make sure our selected model had the knowledge of medical images before fine-tuning it for our cancer cell lines cell line authentication task
InceptionResNet V2 was found to be the optimal model for our problem over MobileNet through the pilot classification task, using RGB images, nearest width shift data augmentation with transfer learning techniques (Method combination V)

Summary

Introduction

Short tandem repeat (STR) analysis is available as a reliable method for the determination of the genetic origin of cell lines, the occurrence of misauthenticated cell lines remains an important issue. Short Tandem Repeat (STR) analysis has been available as a reliable method to identify the genetic origin of a cancer cell line for a w hile[3]. In a study based on the analysis of 482 different human tumor cell lines, up to 96 cell lines were m isidentified[6] Another study[6] found that STR profiling alone is insufficient to exclude inter-species cross-contamination of human cell lines, and the study argued for the need of additional testing and authentication methods. Methods are needed that reliably discriminate between isogenic cell lines, such as clonal sublines and drug-adapted cancer cell lines, since this is not achievable using STR analysis. Automated image recognition approaches may be an additional way to authenticate cells that may complement STR analysis and enable the differentiation between isogenic cell lines[13]

Methods

Results

Discussion

Conclusion