Multi-scale classification of single-cell gel electrophoresis assay using deep learning algorithm

https://doi.org/10.1016/j.bspc.2019.101672Get rights and content

Highlights

  • A novel automatic comet assay scoring algorithm is proposed.

  • The five different comet pattern can be characterized by the CNN algorithm.

  • There is no need for any image conditioning or pre-processing step.

  • The presented method does not need any meta-rules or some kind of threshold values.

Abstract

Structural and functional integrity of deoxyribonucleic acid (DNA) is crucial for the maintenance of hereditary information. However, by-products of cellular metabolism and physical or chemical factors may cause spontaneous DNA damage. The alkaline single-cell gel electrophoresis or comet assay analysis is an easy and reliable method for the determination genotoxic effects of chemical and physical factors. Simply, it is the electrophoretic analysis of intact/damaged DNA of a single cell on in a thin layer of agarose gel. The quantitative analysis of the comet assay images is performed manually by an expert researcher. In visual scoring, DNA nuclei are scored as 0, 1, 2, 3, and 4; and the correct scoring is crucial for the determination of the DNA damage. However, visual scoring depends on the professional experience of the researcher and it is a time consuming and exhausting task. Therefore, this evaluation is inevitable to have subjective results. To avoid this subjectivity and to show the effectiveness of deep learning algorithm on cell images, a Convolution Neural Network (CNN) based deep learning method is proposed to classify comet assay images. According to the results, CNN is trained and tested with high accuracy. The results show that CNN algorithm can successfully classify five different scores of comet assay images, and these results can also reduce the subjectivity.

Introduction

Deoxyribonucleic acid (DNA), the hereditary material of the cell, is the polymer of deoxyribonucleotide subunits [1]. Structural and functional integrity of DNA is crucial for the maintenance of hereditary information. However, by-products of cellular metabolism and physical or chemical factors may cause spontaneous DNA damage [2]. Such DNA lesions can inhibit replication and transcription. In consequence, unrepaired lesions may lead to lethal mutations or large-scale genomic errors that threaten cellular metabolism and organism.

DNA damage analysis, in addition to information about the pathophysiology of the disease, provides important information on the early biological effects of exposure to occupationally harmful chemicals [3]. Besides, DNA damage detection is used in the fields of ecotoxicology, molecular epidemiology, and genotoxicology. There are different methods to evaluate DNA damage. Amongst, comet assay or single-cell gel electrophoresis is the widely used and accepted analysis. The method is based on the microscopic evaluation of the electrophoretic mobility of DNA obtained from a single cell and its nucleus. In general, DNA and/or intact nucleus is isolated, and embedded in thin layers of agarose gel on a slide. The DNA in the agarose gel layers is moved in an alkaline electrophoretic environment. In alkaline media, the chromatin structure opens, and the DNA is partially denatured. Because of the effect of genotoxic factors, single and/or double-strand DNA breaks and DNA fragments of different molecular weights occur. Following electrophoresis, the DNA in the gel is labeled with a fluorescent DNA specific dye like ethidium bromide or silver-stained and observed by a microscope. Intact supercoiled chromatin DNA moves very slowly and observed spherical under the microscope. Depending on the extent of damage, DNA fragments of varying sizes move at differentially with respect to their molecular weight. This results in a form similar to a “comet” in fluorescence microscope images [2,4,5].

Comet assay analysis is the most used method in genotoxic studies due to its high sensitivity. This method also enables the detection of DNA damage in small sample sizes. Besides, it is cheap, reproducible method with short protocol duration [3]. Comet assay images are visually scored by expert researchers from 0 to 4 according to DNA damage levels. However, this scoring needs personal expertise and it may give subjective results.

In recent years, few studies have been made in the literature on the classification of these images [3,[6], [7], [8], [9], [10], [11], [12], [13], [14], [15]]. In the proposed study, unlike from the literature, comet scores were classified using the convolution neural network (CNN) method as a deep learning algorithm. Thus, objective and robust results were tried to be obtained by classifying the comet scores successfully. Also, the success of the deep learning algorithm in the classification of cell images was analyzed.

Detection of DNA damage is a very important issue in comet assay. However, there are few studies about fully automated computer-aided detection of DNA damage. Furthermore, there is no study or developed algorithm for fully automatic segmentation and DNA damage scoring using CNN in the literature.

Turan et al. combined dynamic time warping method and decision tree in a software program to measure DNA damage and to score it. In their study, each single comet assay was manually extracted, then the center of comet head was marked with giving different scores (Score 0, Score 1, Score 2, Score 3) [16]. Mani et al. proposed a standalone tool named “CoMat” for the detection and quantification of the DNA damage by using a software program developed with Visual Studio. CoMat can process different scores of image formats such as JPEG, TIFF, BMP, and PNG [17]. Ganapathy et al. proposed an automated software to detect and quantify the DNA damage by analyzing comet assay images. Support vector machine method is used to classify images into two groups. These groups are Class 1 and Class 2 that are silver-stained images with lightly or moderately damaged cells and silver-stained images with heavily damaged cells respectively [18]. Quintana et al. used automated comet assay analysis by using an image processing algorithm. In preprocessing step unsharp mask is done. Segmentation was performed using a learning algorithm. After performing several tests, it was determined that the most efficient algorithm was the K-Means cluster based on the spatial relationship of neighboring pixels and the general techniques of object detection with thresholding [19]. Sreelatha et al. developed an automated algorithm to analyze DNA damage using silver-stained comet assay for clinical applications. In the preprocessing stage, a contrast enhancement method is used. Then Gaussian filtering is applied to the contrast-enhanced image. After that Otsu’s method is evaluated for the segmentation stage. The comet has separated into three regions. These regions are comet head, tail, and background. In this method, the edges of this region cannot be defined clearly. Therefore, the fuzzy-based algorithm is adopted for clustering [20]. Sreelatha et.al proposed a shading correction algorithm using morphological bottom-hat or top-hat transformation as a preprocessing step. A homomorphic filtering is applied to overcome the problem of silver-stained comet assay the tail region that merged with the background. Then the Otsu method is used to convert binaries image. During morphological filtering operations: morphological closing operation is used to find highly damaged cells and also opening algorithm is evaluated for eliminating noise in the background [14]. Kızıltan et al. proposed a semi-automatic comet assay analysis tool to obtain reliable and accurate measurement results. In this analysis, the head of the comet is segmented from the tail with using the tail moment [6]. Gyori et al. developed a method to detect comets based on geometric properties. They also tried to segment the comet heads using image intensity. Head segmentation of comet is done for lightly damage or no tail DNA since heavily damaged cell tail has a geometrically larger area than the head [10]. Vojnovic et al. focused on algorithms, based on delineating the head from the tail of a comet. This procedure allows detecting very low level of DNA damage. For this purpose, a threshold level is used on the normalized image to determine all bright objects. This threshold level is calculated from the intensity value of the histogram maximum frequency. This procedure can segment the whole area of comet including with both heads and tails together [21]. Gonzalez et al. applied an open-source software, CellProfiler, to analyze automatically comet on the digital image. The boundary of the comet is determined manually. Then, the CellProfiler calculate automatically the measurements of tail and head areas [22]. Sansone et al. developed a comet assay analysis algorithm for minimizing user interaction and for getting more reproducible measurements. The developed algorithm is divided into two steps. In the first step, the comet is detected with Gaussian filtering and morphological operator. In the second step, the comet is segmented as head, halo, and tail with fuzzy clustering to identify areas [7]. Helma et al. proposed an open-source program to analyze the comet assay [12]. Bocker et al. described an automatic image analysis system including software and hardware together to get minimum human interactions. The comet assay image analysis is evaluated in a two-step. These are the automatic comet classification and the calculation of related parameters [9].

Section snippets

Cell and image properties

The comet assay images used in this study were obtained from mammary carcinoma MCF-7 cells. Cell images were obtained by fluorescence microscopy (Nikon, Eclipse E600). 500 comet assay cell images at total were analyzed. Each image was in RGB format and 150 × 150 pixel size. These dimensions were chosen to cover the area from the head to the tail end of the Comet in each image.

Image analysis algorithm

The block diagram of the proposed method used in the study is shown in Fig. 1. The algorithm was evaluated by MATLAB(R),

Results

In this study, 500 single-cell gel electrophoresis assay images were used; equally, consist of five different scores. The representative images of five different comet assay scores are shown in Fig. 4. Fifty percent of all images were used for training, and the rest of the images were used for testing. The trained images were not used for testing. In the training image set, there were 50 images of comet assay score 0, 50 images of comet assay score 1, 50 images of comet assay score 2, 50 images

Discussion

The quantitative analysis of the comet assay images is performed manually expert researchers using various image macros (ImageJ etc.). Image analysis software is used to calculate the amount of DNA in the head and tail, and the length of the tail is also measured. Different software metrics are used to quantify DNA damage. In visual scoring, a tail moment of DNA is expressed as arbitrary units (AU). Nuclei are scored as 0, 1, 2, 3, and 4 by a blinded observer according to the apparent relative

Conclusion

In proposed work, a deep learning-based classification algorithm has been developed for comet (single-cell gel electrophoresis) assay images. The CNN algorithm has high sensitivity, specificity, and accuracy results. Scores of comet assay images are carried out by expert researchers. This may have subjective consequences, and also may differ according to the visual abilities of the experts. This study shows that with the high-performance scoring results, CNN can objectively determine the comet

Acknowledgments

The Department of Molecular Biology and Genetics, Faculty of Arts and Science, Baskent University is greatly acknowledged for comet assay and images. Prof. Dr. Özlem Darcansoy İşeri is greatly acknowledged for her contribution.

Declaration of Competing Interest

The authors report no declarations of interest.

References (29)

  • G. Litjens et al.

    A survey on deep learning in medical image analysis

    Med. Image Anal.

    (2017)
  • T. Fawcett

    An introduction to ROC analysis

    Pattern Recognit. Lett.

    (2006)
  • X. Zhao et al.

    Modification of lymphocyte DNA damage by carotenoid supplementation in postmenopausal women

    Am. J. Clin. Nutr.

    (2006)
  • B. Alberts

    Molecular Biology of the Cell

    (2015)
  • Cited by (7)

    • Integrating convolutional neural networks, kNN, and Bayesian optimization for efficient diagnosis of Alzheimer's disease in magnetic resonance images

      2023, Biomedical Signal Processing and Control
      Citation Excerpt :

      Its components are presented next. We implement a convolutional neural network to extract deep features from MRIs as it was successful in numerous biomedical engineering applications [23–35]. The designed CNN is composed of one input layer used to receive the MRI, three successive two-dimensional convolutional layers (CL) used to analyse the receive MRI and compute its deep features, and pooling layers (PL) which are introduced between successive convolutional layers to gradually decrease the spatial size of the MRI representation so as the CNN performs with less computation power.

    • Agarose waste derived toxicologically screened carbon dots as dual sensor: A mechanistic insight into luminescence and solvatochromic behaviour

      2020, Nano-Structures and Nano-Objects
      Citation Excerpt :

      Thus, it has been considered as a good feeding stock for the synthesis of C-dots. The application of agarose in the separation of DNA, purification of proteins, immune-diffusion and electrophoresis has made it one of the widely used materials in molecular biology [30,31]. However, the safe disposal and management of agarose waste is one of the prime concerns before the researchers for the protection of communal health and ecosystem [32].

    • Faster R-CNN approach for detection and quantification of DNA damage in comet assay images

      2020, Computers in Biology and Medicine
      Citation Excerpt :

      consists of a flexible software-based on computer vision techniques (such as equalization and filtering) and K-Means algorithm to automatically detect comets and perform features extraction, without assessing DNA damage. On the other hand, there are works where only classification task is automatically performed on already segmented comet assay images[19–22]. Turan etal.[19]

    • A Review on Brain Tumor Prediction using Deep Learning

      2023, 2023 International Conference on Advancement in Computation and Computer Technologies, InCACCT 2023
    View all citing articles on Scopus
    View full text