Detection and Classification of Chromosomes with Sister Chromatid Cohesion Defects Using Object Detection Models

Detection and Classification of Chromosomes with Sister Chromatid Cohesion Defects Using Object Detection Models | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Detection and Classification of Chromosomes with Sister Chromatid Cohesion Defects Using Object Detection Models Shinya Matsumoto, Miku Sojo, Kiyoshi Nishikawa, Kan Okubo, Takuya Abe This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7405049/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 16 Mar, 2026 Read the published version in Scientific Reports → Version 1 posted 11 You are reading this latest preprint version Abstract Sister chromatid cohesion (SCC) is mediated by a protein complex called cohesin and by regulatory proteins that control SCC function. A commonly used approach to evaluate the involvement of cohesin regulatory proteins is to classify the shape of the chromosomes after depletion of the target protein and analyze their distribution. Currently, shape classification is often performed manually by researchers, which is not only time-consuming but also subject to individual interpretation. Therefore, our research group developed image classification models for automating chromosome shape classification. However, in this method, unclassifiable chromosomes that arise when cropping single chromosomes must be removed manually, creating a significant barrier to the fully automated detection of SCC-defective chromosomes. In this study, we propose a method that utilizes an object detection model to detect chromosomes with SCC defects without the need to crop single chromosomes. Several pretrained object detection models were selected and fine-tuned, and their performances were compared. Among the models, the one based on You Only Look Once v8 (YOLOv8) achieved a maximum concordance rate of 89.40% with manual analysis and successfully identified differences in the distribution of wild-type (WT) and DDX11 −/− cells. These results indicate that the YOLOv8-based model enables fully automated analysis of SCC-defective chromosomes. Biological sciences/Biological techniques Biological sciences/Cell biology Biological sciences/Computational biology and bioinformatics Biological sciences/Genetics Biological sciences/Molecular biology Sister chromatid cohesion machine learning chromosome analysis Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Sister chromatid cohesion (SCC) is mediated by the cohesin complex, which consists of four core proteins: SMC3, SMC1, RAD21, and SA1/SA2 1,2 . Cohesins hold sister chromatids until the onset of mitosis to prevent premature chromosome segregation and aneuploidy. Cohesin function is regulated by several proteins 3 . DDX11, an evolutionarily conserved DNA helicase, is a critical regulator of cohesin. A mutation in DDX11 cause a genetic developmental disorder called Warsaw Breakage syndrome (WABS) 4 . SCC defects are typically examined by microscopic observation of a metaphase chromosome spread 5 . SCC-defective cells often display more open sister chromatids. However, a key issue with this method is that the evaluation and classification of chromosome shapes are performed manually by researchers, leading to subjective judgments and blurred boundaries between categories. Additionally, for quantitative analysis, a minimum of 50–100 metaphase chromosomes must be classified for each strain or sample, which makes this approach time-consuming and labor-intensive. These challenges increase the risks of human error and subjective bias. To overcome these issues, a machine learning model capable of automatically classifying chromosome shapes is required. In previous studies, we developed image classification models based on deep learning to detect chromosomes with SCC defects. The initial model, based on SqueezeNet, achieved a concordance rate of 73.1% with example answers (EAs) provided by the researcher. The second model, based on DenseNet161, achieved a concordance rate of 89.67% 7 . Moreover, our latest model, based on a Vision Transformer, succeeded in increasing the concordance rate to 93.02% 8 . As these models are specifically designed for individual chromosome classification, preprocessing steps are necessary to isolate single chromosomes from composite metaphase images. Thus, we utilized the contour extraction function of OpenCV image processing to crop the single chromosomes. This function detects chromosomal outlines and single chromosomes of rectangular crops based on coordinate information. The limitation of this method is that many overlapping chromosomes and two nearby chromosomes are mistakenly cropped together, making manual removal unavoidable. Moreover, owing to variations in the background intensity and staining levels across chromosome preparations, the parameters for edge detection and morphological processing require individual optimization for accurate chromosome contour extraction. These steps act as barriers to fully automating analysis. Object detection, which is a fundamental task in computer vision , is used in various fields such as medicine and manufacturing 9 , 10 . While image classification assigns a single class label to an entire image, object detection simultaneously localizes and classifies multiple objects within an image through a unified computational process. In the field of chromosome analysis, a recent study proposed an automated chromosome analysis system that applies an object detection method based on deep learning to detect chromosomal aberrations such as dicentric and ring chromosomes 11 . However, it is unclear whether object detection models can classify SCC-defective chromosomes, whose morphological changes are subtler than those of dicentric and ring chromosomes. In this study, we developed an object-detection model that enables the fully automated detection of chromosomes with SCC defects. In the proposed method, we used an object detection model to simultaneously detect single chromosomes from images containing multiple chromosomes and classify them into three types based on the positional relationship between sister chromatids. This approach enables the detection of chromosomes with SCC defects without the need for single-chromosome cropping or manual removal of unanalyzable chromosomes. As a result, comparing several pre-trained object detection models, the model based on You Only Look Once v8 (YOLOv8) achieved a maximum concordance rate of 89.40% with EA. The model revealed a difference between chromosomes from wild-type (WT) and DDX11 −/− cells, which have more SCC-defective chromosomes. These results show that object detection–based models can be used for the fully automated analysis of SCC-defective chromosomes. Material and methods Evaluation indicators The accuracy of each model was evaluated using mean Average Precision (mAP). mAP, a standard metric used to quantitatively evaluate the object detection performance by calculating the area under the precision–recall curve. It is commonly used to assess and compare the accuracies of object detection models. [email protected] refers to the mean Average Precision when the Intersection over Union (IoU) threshold is set to 0.5. IoU is a metric that measures the overlap between two bounding boxes and is calculated as the area of the intersection divided by the area of the union of the two bounding boxes. In [email protected] , a prediction is considered a true positive if the IoU between the predicted bounding box and the ground truth bounding box is 0.5 or higher. Similarly, [email protected] refers to the mean Average Precision when the IoU threshold was set to 0.75. [email protected] :0.95 represents the mean Average Precision over a range of IoU thresholds varying from 0.5 to 0.95, providing the average accuracy across different IoU thresholds. In addition to the mAPs, another metric—the concordance rate—was used. The concordance rate indicates the proportion of chromosomes that were accurately detected and classified by the model, representing how closely the model results matched those of EA. The concordance rate was calculated using the following equation (Eq. 1), where \(\:{N}_{TP}\) represents the number of true positives and \(\:{N}_{FN}\) represents the number of false negatives. To calculate the concordance rate, the IoU threshold was set to 0.45. The bounding boxes with a confidence score of 0.6 or higher were used as the model's predictions. $$\:\begin{array}{c}Concordance\:rate=\frac{Correctlydetectedandclassifiedchromosomes}{Labeledchromosomes}=\frac{{N}_{TP}}{{N}_{TP}+{N}_{FN}}\#\left(1\right)\end{array}$$ Cell culture WT TK6 cells were provided by Dr. Kouji Hirota (Department of Chemistry, Graduate School of Science, Tokyo Metropolitan University), and DDX11 −/− TK6 cells were established from WT TK6 cells in this study. These cells were cultured at 37°C in RPMI medium (Wako) supplemented with 5% house serum (Gibco), penicillin-streptomycin mix (Nacalai Tesque), 2 mM l-glutamine (Nacalai Tesque), and 100 µM sodium pyruvate. Chromosome preparation Chromosome preparations were performed as previously described 6 . Plasmid construction and transfection DDX11 knockout (KO)-Neo and DDX11 KO-Hyg vectors were generated from genomic PCR products combined with a neomycin or a hygromycin marker cassette. The left and right homology arms of DDX11 KO vectors were amplified using the primers 5’-aaaGGTACCacagtgttccgatgagaccacagtaggc-3’ and 5’- aaaCTCGAGggcctggcagctttcctcagtttctctg-3’ (for the left arm of the KO construct); and 5’- aaaGCGGCCGCatggttcctccagacacctgggccaag-3’ and 5’- aaaGTCGACgaaccaaagtgctgcctgcctctcagag-3’ (for the right arm of the KO construct). The amplified PCR products were cloned into DT-ApA / NEO R (provided by the Laboratory for Animal Resources and Genetic Engineering, Center for Developmental Biology, RIKEN Kobe), DT-ApA / HYG R or pLoxP vectors using the attached restriction sites 12 . A gRNA to introduce a DSB into the DDX11 locus was designed using CRISPR Direct ( https://crispr.dbcls.jp/ ). The CRISPR expression vector for DDX11 locus was designed to recognize 5’- aagtccctttacgtcacagc-3’. The pX330 vector (Addgene plasmid #42230) was used for the CRISPR-Cas9 system. Western blotting analysis Western blotting analysis Western blotting was performed using primary antibodies against DDX11 (Santa Cruz Biotechnology, sc-271711) and Topoisomerase I (Abcam, ab109374), followed by incubation with a horseradish peroxidase (HRP)-conjugated anti-mouse IgG secondary antibody (Cell Signaling Technology). Protein bands were visualized using ImmunoStar LD according to the manufacturer’s protocol. Growth curve WT TK6 and DDX11 −/− TK6 cells were cultured at 37°C in RPMI medium (Wako) supplemented with 5% house serum (Gibco), penicillin-streptomycin mix (Nacalai Tesque), 2 mM l-glutamine (Nacalai Tesque), and 100 µM sodium pyruvate. To plot growth curves, each cell line was cultured in three different wells of 24 well-plates and passaged every 24 h. Viable cell numbers were determined by flow cytometry. 25 µl of cell suspension were analyzed, and viable cells determined by forward scatter and side scatter were counted. Results Dataset preparation To establish an SCC-defective model cell line, we generated DDX11 KO cells from the human TK6 cell line. The details of the KO design are shown in Fig. 1 A and in the Materials and Methods section. The absence of DDX11 protein expression in the established DDX11 −/− cells was confirmed by western blot analysis (Fig. 1 B, Supplementary Fig. 1). DDX11 −/− cells were viable, but their proliferative capacity slightly decreased (Fig. 1 C). Chromosome spreads were prepared from WT cells and DDX11 −/− cells, and whole-chromosome images were captured. Each chromosome was manually labeled by displaying the location coordinates and types of chromosomes with bounding boxes using Coco-annotator 13 , a web-based image annotation tool. Chromosomes are classified into three types. Well-coherent tight chromosomes were classified as type A; chromosomes in which the arms were separated were classified as type B; and chromosomes in which sister chromatids were separated at the centromere were classified as type C 6 . These labels were used as EA, and representative images of labeled chromosomes and illustrations of each type are shown in Fig. 1 D-E. A total of 459 chromosomes from WT cells and 2046 chromosomes from DDX11 −/− cells were prepared. Since the rate of type C chromosomes was low in WT cells (~ 10%), we used chromosome images from DDX11 −/− cells, which have almost equal numbers of type A, type B, and type C chromosomes, for well-balanced training. To this end, we divided the chromosome images from DDX11 −/− cells into training, validation, and testing sets and used all chromosome images from WT cells for testing, as shown in Fig. 2 . Fine-tuning of models We compared four object detection models from the YOLO series released by the Ultralytics company: YOLOv5 14 , YOLOv5u 14 , YOLOv8 15 , and YOLO11 16 , and three other object detection models: Faster R-CNN 17 , SSD 18 , and DETR 19 . YOLO is a one-stage object detection method based on a convolutional neural network (CNN) and is known for its high detection accuracy and fast processing speed. YOLO11 is the latest version of the YOLO series, whereas the other YOLOs (YOLOv5, YOLOv5u, and YOLOv8) are earlier versions. SSD represents another one-stage object detection model using a CNN, whereas Faster R-CNN is a two-stage object detection model based on a CNN. In contrast, DETR is a transformer-based object-detection model. According to the respective model guidelines, the SSD model was pretrained on the Pascal VOC dataset 20 , whereas the other models were pretrained on the COCO dataset 21 . To achieve high accuracy with a limited number of chromosomes, fine-tuning was performed on the pre-trained object detection models. We fine-tuned all layers of each model sufficiently using 1270 chromosomes for training and 311 chromosomes for validation. Chromosome images were resized from 1440 × 1024 pixels to 1440 × 1440 pixels for YOLOv5, YOLOv5u, YOLOv8, YOLO11, and Faster R-CNN; 300 × 300 pixels for SSD; and 1125 × 800 pixels for DETR before being input. We repeated the training, validation, and testing cycles 3 times and used the average score from the three trials. The accuracy of chromosome detection with trained models First, we compared the accuracy of chromosome detection for each model. Specifically, the comparison focused solely on detection, disregarding the accuracy of the chromosome types predicted by the models. This comparison involved models based on YOLOv5, YOLOv5u, YOLOv8, YOLO11, Faster R-CNN, SSD, and DETR. For comparison, 459 chromosomes from WT cells were used. Table. Table 1 presents the concordance rates for [email protected] , [email protected] , and [email protected] :0.95. The concordance rates were 88.93% (YOLOv5u), 95.13% (YOLOv8 and YOLO11), and 95.57% (Faster R-CNN). [email protected] :0.95 were 69.87% (YOLOv5u), 74.20% (YOLOv8), 75.27% (YOLO11), and 70.27% (Faster R-CNN). The models based on YOLOv5, SSD, and DETR exhibited significantly lower accuracy. In the case of the YOLOv5-based model, this may be primarily due to its use of the same head design as the older YOLOv3 22 because YOLOv5u, which replaces the YOLOv5 head design with the one used in YOLOv8, showed a significant improvement in accuracy compared to YOLOv5. An outdated head design may limit the ability to detect small chromosomes. Similarly, the SSD-based model likely struggled because of the required reduction in the input image size according to the model specifications. The reduction in the input image size likely caused the chromosomes to appear even smaller, making detection more challenging. Unlike for DETR, the performance of DETR may be limited by the absence of a multiscale feature extraction mechanism. Consequently, small object features may have been lost during processing in the CNN backbone of DETR, leading to reduced detection accuracy. Detection and classification of chromosomes with trained models Next, we compared the classification accuracies of the trained models. Four models with high chromosome detection performances were examined, and 459 chromosomes from WT cells and 465 chromosomes from DDX11 −/− cells were used as the test data. Table 2 presents the results for the concordance rates, [email protected] , [email protected] , and [email protected] :0.95 for each model. The YOLOv8-based model achieved the highest concordance rate and mAP for both WT and DDX11 −/− cells, with a maximum concordance rate of 89.40%. Detection of SCC defects in DDX11 −/− cells with trained models Since the analyses output the distribution of each type of chromosome in WT and DDX11 −/− cells, we compared the distribution in each model and confirmed whether the models produced similar results to manual analysis. The rates of each type of chromosome obtained from manual analysis or the trained model are shown in Fig. 3 . In manual analysis, the proportion of each chromosome type in DDX11 −/− cells, compared to WT cells, changed as follows: type A decreased from 25.1–24.9%, type B decreased from 65.8–41.3%, and type C increased from 9.1–33.8%. Although all the trained models detected the difference between WT cells and DDX11 −/− cells, especially the YOLOv8-based model, which achieved the highest concordance rates with EA, outputted the most similar distribution with manual analysis. These results demonstrate that the YOLO based model can automatically detect SCC defects without human intervention. Discussion In this study, we demonstrated the fully automated detection of chromosomes with SCC defects using an object-detection model. We prepared chromosomal images from WT and DDX11 −/− cells and fine-tuned several pre-trained object detection models using these images. By comparing the performance of each model, the YOLOv8-based model achieved a maximum concordance rate of 89.40% with EA and demonstrated results comparable to those of manual analysis. Based on these results, the technical improvements in YOLOv8 were considered effective for the tasks in this study. Specifically, all C3 modules in YOLOv5 and v5u, which split the feature map into two paths, one direct and the other processed through bottleneck structures, were replaced with C2f in YOLOv8. The C2f module incorporated a cross-stage partial bottleneck with two convolutional layers that combined high-level features with contextual information. Additionally, YOLOv5 uses an anchor-based detection method and has a coupled head structure, whereas YOLOv5u and v8 adopt an anchor-free detection method and a decoupled head structure with two separate branches for object classification and predicted bounding box regression. This design allows each branch to focus on its tasks and improves the overall accuracy of the model. A comparison between YOLOv5s and YOLOv5u revealed a significant increase in chromosome detection accuracy, demonstrating the essential nature of these architectural improvements for chromosome analysis. Compared with YOLOv8, YOLO11 introduces two main architectural improvements: replacing the C2f module with the C3k2 module and incorporating the new C2PSA module. While the C3k2 module is designed to be faster and more efficient for feature aggregation, and the spatial attention mechanism in the C2PSA module allows the model to focus more effectively on important regions within the image, the performance of YOLO11 showed no improvement compared with YOLOv8. Thus, the modifications introduced to YOLO11 may not have been effective for the chromosome classification tasks in this study. Although the accuracy of the YOLOv8-based model seems sufficient for practical use, and future YOLO models are expected to achieve better performance, we raise two other challenges to increase accuracy. The first is to improve the labeling accuracy for both the training and test data. Some chromosomes were difficult to classify into a specific category, and incorrect labeling may have occurred. Refining the training data by having multiple researchers independently review and verify the labels could effectively improve the measurement accuracy. The second is accurate detection of overlapping chromosomes. The YOLOv8-based model occasionally failed to detect overlapping chromosomes (Fig. 4 ). To address this issue, the following two approaches were considered: (i) Application of soft Soft-NMS method 23 The current Non-Maximum Suppression (NMS) method integrates multiple bounding boxes for a single object by removing similar boxes. However, this method retains only one bounding box for overlapping objects, which can lead to detection errors. Soft-NMS measures the degree of overlap between the bounding boxes and reduces the confidence score if the overlap exceeds a certain threshold. This method allows the retention of information regarding overlapping bounding boxes while selecting the most confident bounding box. It has been reported that applying Soft-NMS improves the detection accuracy of overlapping chromosomes 11 , and accuracy improvements can be expected. (ii) Adoption of Oriented Bounding Box (OBB) 24 An OBB is a rotatable bounding box aligned with the orientation of objects, which enables close enclosure of objects. By using the OBB for elongated chromosomes oriented in multiple directions, the overlap of bounding boxes can be minimized, which is expected to improve the detection accuracy of overlapping chromosomes. Declarations Author Contribution K. N., K. O., and T. A. conceived and designed the experiments; S. M. and M. S. performed the experiments; S. M. and T. A. wrote the manuscript. Acknowledgments This work was supported by Grants from JSPS KAKENHI (22H05072 and 25K09513) to TA and JSPS KAKENHI (22K12170) to KN. Data Availability The datasets used and/or analyzed in the current study are available from the corresponding author upon reasonable request. References Peters, J.-M. & Nishiyama, T. Sister Chromatid Cohesion. Cold Spring Harb Perspect Biol 4 , a011130 (2012). van Schie, J. J. M. & de Lange, J. The Interplay of Cohesin and the Replisome at Processive and Stressed DNA Replication Forks. Cells 10 , 3455 (2021). Murayama, Y., Samora, C. P., Kurokawa, Y., Iwasaki, H. & Uhlmann, F. Establishment of DNA-DNA Interactions by the Cohesin Ring. Cell 172 , (465-477.e15) (2018). van Schie, J. J. M. et al. Warsaw Breakage Syndrome associated DDX11 helicase resolves G-quadruplex structures to support sister chromatid cohesion. Nat Commun 11 , 4287 (2020). Deng, W., Tsao, S. W., Lucas, J. N., Leung, C. S. & Cheung, A. L. M. A new method for improving metaphase chromosome spreading. Cytometry Part A 51 , 46–51 (2003). Ikemoto, D. et al. Application of neural network-based image analysis to detect sister chromatid cohesion defects. Sci Rep 13 , 2133 (2023). Matsumoto, S., Ikemoto, D., Abe, T., Okubo, K. & Nishikawa, K. Automatic classification of human chromosome shapes using convolutional neural network models. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences 2024EAL2053 (2025) doi:10.1587/transfun. EAL2053 (2024) Matsumoto, S., Okubo, K., Abe, T. & Nishikawa, K. Detection model of sister chromatid cohesion defects based on Vision Transformer. in 2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) 27–31 (IEEE, 2023). doi:10.1109/APSIPAASC58517.2023.10317257. Sergios Theodoridis & Konstantinos Koutroumbas. Pattern Recognition . (Elsevier, 2009). doi:10.1016/B978-1-59749-272-0.X0001-2. Kevin P. Murphy. Machine Learning: A Probabilistic Perspective . (MIT Press, 2012). Kang, S. et al. Chromosome analysis method based on deep learning: Counting chromosomes and detecting abnormal chromosomes. Biomed Signal Process Control 91 , 105891 (2024). Arakawa, H., Lodygin, D. & Buerstedde, J.-M. Mutant loxP vectors for selectable marker recycle and conditional knock-outs. BMC Biotechnol 1 , 7 (2001). Stefanics, D. & Fox, M. COCO Annotator. ACM SIGMultimedia Records 13 , 1–1 (2021). Glenn Jocher. Ultralytics YOLOv5. https://github.com/ultralytics/yolov5 (2020). Glenn Jocher, Ayush Chaurasia & Jing Qiu. Ultralytics YOLOv8. https://github.com/ultralytics/ultralytics (2023). Glenn Jocher & Jing Qiu. Ultralytics YOLO11. https://github.com/ultralytics/ultralytics (2024). Ren, S., He, K., Girshick, R. & Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. (2015). Liu, W. et al. SSD: Single Shot MultiBox Detector. in 21–37 (2016). doi:10.1007/978-3-319-46448-0_2. Carion, N. et al. End-to-End Object Detection with Transformers. (2020). Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J. & Zisserman, A. The Pascal Visual Object Classes (VOC) Challenge. Int J Comput Vis 88 , 303–338 (2010). Lin, T.-Y. et al. Microsoft COCO: Common Objects in Context. (2014). Redmon, J. & Farhadi, A. YOLOv3: An Incremental Improvement. (2018). Bodla, N., Singh, B., Chellappa, R. & Davis, L. S. Soft-NMS — Improving Object Detection with One Line of Code. in 2017 IEEE International Conference on Computer Vision (ICCV) 5562–5570 (IEEE, 2017). doi:10.1109/ICCV.2017.593. Zand, M., Etemad, A. & Greenspan, M. Oriented Bounding Boxes for Small and Freely Rotated Objects. IEEE Transactions on Geoscience and Remote Sensing 60 , 1–15 (2022). Tables Table 1 and 2 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.jpeg Table 1. Object detection metrics of the chromosome detection tasks. Table2.jpeg Table 2. Object detection metrics of the chromosome classification tasks in WT cells (A) and DDX11 -/- cells (B). matsumotoetalSupplementaryfigure250707.pptx Cite Share Download PDF Status: Published Journal Publication published 16 Mar, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 29 Dec, 2025 Editor invited by journal 25 Nov, 2025 Reviews received at journal 25 Oct, 2025 Reviews received at journal 23 Oct, 2025 Reviewers agreed at journal 15 Oct, 2025 Reviewers agreed at journal 15 Oct, 2025 Reviewers agreed at journal 10 Sep, 2025 Reviewers invited by journal 10 Sep, 2025 Editor assigned by journal 05 Sep, 2025 Submission checks completed at journal 23 Aug, 2025 First submitted to journal 23 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7405049","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":513763434,"identity":"1ad51b38-9f29-4a53-8243-16c30eed8a5e","order_by":0,"name":"Shinya Matsumoto","email":"","orcid":"","institution":"Tokyo Metropolitan University","correspondingAuthor":false,"prefix":"","firstName":"Shinya","middleName":"","lastName":"Matsumoto","suffix":""},{"id":513763435,"identity":"189c65c9-ed70-43c8-8282-8fa852130b69","order_by":1,"name":"Miku Sojo","email":"","orcid":"","institution":"Tokyo Metropolitan University","correspondingAuthor":false,"prefix":"","firstName":"Miku","middleName":"","lastName":"Sojo","suffix":""},{"id":513763436,"identity":"7d690087-e601-4647-b076-cdc5b7e8d920","order_by":2,"name":"Kiyoshi Nishikawa","email":"","orcid":"","institution":"Tokyo Metropolitan University","correspondingAuthor":false,"prefix":"","firstName":"Kiyoshi","middleName":"","lastName":"Nishikawa","suffix":""},{"id":513763437,"identity":"8c092b66-7c41-45ae-8ed7-a5553bdd0744","order_by":3,"name":"Kan Okubo","email":"","orcid":"","institution":"Tokyo Metropolitan University","correspondingAuthor":false,"prefix":"","firstName":"Kan","middleName":"","lastName":"Okubo","suffix":""},{"id":513763438,"identity":"71a9b0cb-9ad2-4363-9776-e656f874ceec","order_by":4,"name":"Takuya Abe","email":"data:image/png;base64,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","orcid":"","institution":"Tohoku Medical and Pharmaceutical University","correspondingAuthor":true,"prefix":"","firstName":"Takuya","middleName":"","lastName":"Abe","suffix":""}],"badges":[],"createdAt":"2025-08-19 06:23:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7405049/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7405049/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-43009-6","type":"published","date":"2026-03-16T15:59:01+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":91560993,"identity":"8f82c702-f55f-46b3-be10-25ba41430d85","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":102619,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Schematic representation of the \u003cem\u003eDDX11\u003c/em\u003e gene locus and gene targeting knockout construct. (Closed boxes) Exons; (Marker) drug resistance genes. The vector was designed to delete the exon encoding the walker B motif of DDX11. (B) Whole-cell lysates were prepared from cells of the indicated genotypes. DDX11 and TOP1 (loading control) were detected by Western blotting. (C) Growth curves of the indicated cell lines. Cells (1 × 10\u003csup\u003e5\u003c/sup\u003e) of the indicated genotypes were inoculated in 1 mL of medium, counted, and passaged every 24 h. Distribution of each type of chromosome in WT TK6 and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e-/-\u003c/em\u003e\u003c/sup\u003e TK6 cells. (D) Representative images of each type of chromosome. A typical chromosome spread picture. Type A, type B, and type C chromosomes were boxed by green, yellow, and red squares. (E) Image and cartoon showing each type of chromosome.\u003c/p\u003e","description":"","filename":"Figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/ca962505b1738ee1a09a40d5.jpeg"},{"id":91560992,"identity":"07fd9f00-a321-4801-8ee1-1823eb9bcf82","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":105529,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of dataset division.\u003c/p\u003e","description":"","filename":"Figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/ec594bf17ad2fb686c957975.jpeg"},{"id":91560996,"identity":"7c5d3580-9c8f-4d54-86d1-9e3455133431","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":95220,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of each type of chromosome in WT and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e-/-\u003c/em\u003e\u003c/sup\u003e cells. (A) EA (B) YOLOv5u (C) YOLOv8 (D) YOLO11 (E) Faster R-CNN.\u003c/p\u003e","description":"","filename":"Figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/cfc16d03ef671ae9be97ca11.jpeg"},{"id":91560994,"identity":"a463c527-fe8d-497a-93da-07b276c5ce1b","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":44073,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative images of overlapping chromosomes. The two overlapping chromosomes indicated by the red arrow were grouped by the YOLOv8-based object detection model. Type A, type B, and type C chromosomes were boxed by green, yellow, and red squares.\u003c/p\u003e","description":"","filename":"Figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/13626c41cbeb2881cd0175cc.jpeg"},{"id":105223591,"identity":"8b2962f8-189b-462b-8fcd-4e53dae86b99","added_by":"auto","created_at":"2026-03-23 16:08:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":942372,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/de5bb1da-692a-404d-8be9-3ad6686c8e9a.pdf"},{"id":91560990,"identity":"539eb7c4-d26e-45ed-92b9-87a006f96c04","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"jpeg","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":69289,"visible":true,"origin":"","legend":"\u003cp\u003eTable 1. Object detection metrics of the chromosome detection tasks.\u003c/p\u003e","description":"","filename":"Table1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/f7a8d0c5544a3a788c10d665.jpeg"},{"id":91562364,"identity":"c8131f53-5bc2-434e-9fb4-3fc13da4dff7","added_by":"auto","created_at":"2025-09-17 18:53:35","extension":"jpeg","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":90835,"visible":true,"origin":"","legend":"\u003cp\u003eTable 2. Object detection metrics of the chromosome classification tasks in WT cells (A) and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e-/-\u003c/em\u003e\u003c/sup\u003e cells (B).\u003c/p\u003e","description":"","filename":"Table2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/8993b560c952134bff702f7b.jpeg"},{"id":91561010,"identity":"100f786e-edc7-4115-b016-cfa641b28c88","added_by":"auto","created_at":"2025-09-17 18:45:35","extension":"pptx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":2481947,"visible":true,"origin":"","legend":"","description":"","filename":"matsumotoetalSupplementaryfigure250707.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7405049/v1/325630a1411c3ad29b58f04a.pptx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Detection and Classification of Chromosomes with Sister Chromatid Cohesion Defects Using Object Detection Models","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSister chromatid cohesion (SCC) is mediated by the cohesin complex, which consists of four core proteins: SMC3, SMC1, RAD21, and SA1/SA2\u003csup\u003e1,2\u003c/sup\u003e. Cohesins hold sister chromatids until the onset of mitosis to prevent premature chromosome segregation and aneuploidy. Cohesin function is regulated by several proteins\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. DDX11, an evolutionarily conserved DNA helicase, is a critical regulator of cohesin. A mutation in DDX11 cause a genetic developmental disorder called Warsaw Breakage syndrome (WABS)\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eSCC defects are typically examined by microscopic observation of a metaphase chromosome spread\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. SCC-defective cells often display more open sister chromatids. However, a key issue with this method is that the evaluation and classification of chromosome shapes are performed manually by researchers, leading to subjective judgments and blurred boundaries between categories. Additionally, for quantitative analysis, a minimum of 50\u0026ndash;100 metaphase chromosomes must be classified for each strain or sample, which makes this approach time-consuming and labor-intensive. These challenges increase the risks of human error and subjective bias. To overcome these issues, a machine learning model capable of automatically classifying chromosome shapes is required.\u003c/p\u003e\u003cp\u003eIn previous studies, we developed image classification models based on deep learning to detect chromosomes with SCC defects. The initial model, based on SqueezeNet, achieved a concordance rate of 73.1% with example answers (EAs) provided by the researcher. The second model, based on DenseNet161, achieved a concordance rate of 89.67%\u003csup\u003e7\u003c/sup\u003e. Moreover, our latest model, based on a Vision Transformer, succeeded in increasing the concordance rate to 93.02%\u003csup\u003e8\u003c/sup\u003e. As these models are specifically designed for individual chromosome classification, preprocessing steps are necessary to isolate single chromosomes from composite metaphase images. Thus, we utilized the contour extraction function of OpenCV image processing to crop the single chromosomes. This function detects chromosomal outlines and single chromosomes of rectangular crops based on coordinate information. The limitation of this method is that many overlapping chromosomes and two nearby chromosomes are mistakenly cropped together, making manual removal unavoidable. Moreover, owing to variations in the background intensity and staining levels across chromosome preparations, the parameters for edge detection and morphological processing require individual optimization for accurate chromosome contour extraction. These steps act as barriers to fully automating analysis.\u003c/p\u003e\u003cp\u003eObject detection, which is a fundamental task in computer vision\u003csup\u003e,\u003c/sup\u003e is used in various fields such as medicine and manufacturing\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. While image classification assigns a single class label to an entire image, object detection simultaneously localizes and classifies multiple objects within an image through a unified computational process. In the field of chromosome analysis, a recent study proposed an automated chromosome analysis system that applies an object detection method based on deep learning to detect chromosomal aberrations such as dicentric and ring chromosomes\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. However, it is unclear whether object detection models can classify SCC-defective chromosomes, whose morphological changes are subtler than those of dicentric and ring chromosomes.\u003c/p\u003e\u003cp\u003eIn this study, we developed an object-detection model that enables the fully automated detection of chromosomes with SCC defects. In the proposed method, we used an object detection model to simultaneously detect single chromosomes from images containing multiple chromosomes and classify them into three types based on the positional relationship between sister chromatids. This approach enables the detection of chromosomes with SCC defects without the need for single-chromosome cropping or manual removal of unanalyzable chromosomes. As a result, comparing several pre-trained object detection models, the model based on You Only Look Once v8 (YOLOv8) achieved a maximum concordance rate of 89.40% with EA. The model revealed a difference between chromosomes from wild-type (WT) and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003ecells, which have more SCC-defective chromosomes. These results show that object detection\u0026ndash;based models can be used for the fully automated analysis of SCC-defective chromosomes.\u003c/p\u003e"},{"header":"Material and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eEvaluation indicators\u003c/h2\u003e\u003cp\u003eThe accuracy of each model was evaluated using mean Average Precision (mAP). mAP, a standard metric used to quantitatively evaluate the object detection performance by calculating the area under the precision\u0026ndash;recall curve. It is commonly used to assess and compare the accuracies of object detection models. [email protected] refers to the mean Average Precision when the Intersection over Union (IoU) threshold is set to 0.5. IoU is a metric that measures the overlap between two bounding boxes and is calculated as the area of the intersection divided by the area of the union of the two bounding boxes. In [email protected], a prediction is considered a true positive if the IoU between the predicted bounding box and the ground truth bounding box is 0.5 or higher. Similarly, [email protected] refers to the mean Average Precision when the IoU threshold was set to 0.75. [email protected]:0.95 represents the mean Average Precision over a range of IoU thresholds varying from 0.5 to 0.95, providing the average accuracy across different IoU thresholds. In addition to the mAPs, another metric\u0026mdash;the concordance rate\u0026mdash;was used. The concordance rate indicates the proportion of chromosomes that were accurately detected and classified by the model, representing how closely the model results matched those of EA. The concordance rate was calculated using the following equation (Eq.\u0026nbsp;1), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{TP}\\)\u003c/span\u003e\u003c/span\u003e represents the number of true positives and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{FN}\\)\u003c/span\u003e\u003c/span\u003e represents the number of false negatives. To calculate the concordance rate, the IoU threshold was set to 0.45. The bounding boxes with a confidence score of 0.6 or higher were used as the model's predictions.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}Concordance\\:rate=\\frac{Correctlydetectedandclassifiedchromosomes}{Labeledchromosomes}=\\frac{{N}_{TP}}{{N}_{TP}+{N}_{FN}}\\#\\left(1\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eCell culture\u003c/h3\u003e\n\u003cp\u003eWT TK6 cells were provided by Dr. Kouji Hirota (Department of Chemistry, Graduate School of Science, Tokyo Metropolitan University), and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e TK6 cells were established from WT TK6 cells in this study. These cells were cultured at 37\u0026deg;C in RPMI medium (Wako) supplemented with 5% house serum (Gibco), penicillin-streptomycin mix (Nacalai Tesque), 2 mM l-glutamine (Nacalai Tesque), and 100 \u0026micro;M sodium pyruvate.\u003c/p\u003e\n\u003ch3\u003eChromosome preparation\u003c/h3\u003e\n\u003cp\u003eChromosome preparations were performed as previously described\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003ch3\u003ePlasmid construction and transfection\u003c/h3\u003e\n\u003cp\u003e\u003cem\u003eDDX11\u003c/em\u003e knockout (KO)-Neo and \u003cem\u003eDDX11\u003c/em\u003e KO-Hyg vectors were generated from genomic PCR products combined with a neomycin or a hygromycin marker cassette. The left and right homology arms of \u003cem\u003eDDX11\u003c/em\u003e KO vectors were amplified using the primers 5\u0026rsquo;-aaaGGTACCacagtgttccgatgagaccacagtaggc-3\u0026rsquo; and 5\u0026rsquo;- aaaCTCGAGggcctggcagctttcctcagtttctctg-3\u0026rsquo; (for the left arm of the KO construct); and 5\u0026rsquo;- aaaGCGGCCGCatggttcctccagacacctgggccaag-3\u0026rsquo; and 5\u0026rsquo;- aaaGTCGACgaaccaaagtgctgcctgcctctcagag-3\u0026rsquo; (for the right arm of the KO construct). The amplified PCR products were cloned into \u003cem\u003eDT-ApA\u003c/em\u003e/\u003cem\u003eNEO\u003c/em\u003e\u003csup\u003e\u003cem\u003eR\u003c/em\u003e\u003c/sup\u003e (provided by the Laboratory for Animal Resources and Genetic Engineering, Center for Developmental Biology, RIKEN Kobe), \u003cem\u003eDT-ApA\u003c/em\u003e/\u003cem\u003eHYG\u003c/em\u003e\u003csup\u003e\u003cem\u003eR\u003c/em\u003e\u003c/sup\u003e or pLoxP vectors using the attached restriction sites\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. A gRNA to introduce a DSB into the \u003cem\u003eDDX11\u003c/em\u003e locus was designed using CRISPR Direct (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://crispr.dbcls.jp/\u003c/span\u003e\u003cspan address=\"https://crispr.dbcls.jp/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The CRISPR expression vector for \u003cem\u003eDDX11\u003c/em\u003e locus was designed to recognize 5\u0026rsquo;- aagtccctttacgtcacagc-3\u0026rsquo;. The pX330 vector (Addgene plasmid #42230) was used for the CRISPR-Cas9 system.\u003c/p\u003e\n\u003ch3\u003eWestern blotting analysis\u003c/h3\u003e\n\u003cdiv class=\"Heading\"\u003eWestern blotting analysis\u003c/div\u003e\u003cp\u003eWestern blotting was performed using primary antibodies against DDX11 (Santa Cruz Biotechnology, sc-271711) and Topoisomerase I (Abcam, ab109374), followed by incubation with a horseradish peroxidase (HRP)-conjugated anti-mouse IgG secondary antibody (Cell Signaling Technology). Protein bands were visualized using ImmunoStar LD according to the manufacturer\u0026rsquo;s protocol.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eGrowth curve\u003c/h2\u003e\u003cp\u003eWT TK6 and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e TK6 cells were cultured at 37\u0026deg;C in RPMI medium (Wako) supplemented with 5% house serum (Gibco), penicillin-streptomycin mix (Nacalai Tesque), 2 mM l-glutamine (Nacalai Tesque), and 100 \u0026micro;M sodium pyruvate. To plot growth curves, each cell line was cultured in three different wells of 24 well-plates and passaged every 24 h. Viable cell numbers were determined by flow cytometry. 25 \u0026micro;l of cell suspension were analyzed, and viable cells determined by forward scatter and side scatter were counted.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003eDataset preparation\u003c/h2\u003e\u003cp\u003eTo establish an SCC-defective model cell line, we generated \u003cem\u003eDDX11\u003c/em\u003e KO cells from the human TK6 cell line. The details of the KO design are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA and in the Materials and Methods section. The absence of DDX11 protein expression in the established \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells was confirmed by western blot analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB, Supplementary Fig.\u0026nbsp;1). \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells were viable, but their proliferative capacity slightly decreased (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC). Chromosome spreads were prepared from WT cells and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, and whole-chromosome images were captured. Each chromosome was manually labeled by displaying the location coordinates and types of chromosomes with bounding boxes using Coco-annotator\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, a web-based image annotation tool. Chromosomes are classified into three types. Well-coherent tight chromosomes were classified as type A; chromosomes in which the arms were separated were classified as type B; and chromosomes in which sister chromatids were separated at the centromere were classified as type C\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. These labels were used as EA, and representative images of labeled chromosomes and illustrations of each type are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eD-E. A total of 459 chromosomes from WT cells and 2046 chromosomes from \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells were prepared. Since the rate of type C chromosomes was low in WT cells (~\u0026thinsp;10%), we used chromosome images from \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, which have almost equal numbers of type A, type B, and type C chromosomes, for well-balanced training. To this end, we divided the chromosome images from \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells into training, validation, and testing sets and used all chromosome images from WT cells for testing, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eFine-tuning of models\u003c/h2\u003e\u003cp\u003eWe compared four object detection models from the YOLO series released by the Ultralytics company: YOLOv5\u003csup\u003e14\u003c/sup\u003e, YOLOv5u\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, YOLOv8\u003csup\u003e15\u003c/sup\u003e, and YOLO11\u003csup\u003e16\u003c/sup\u003e, and three other object detection models: Faster R-CNN\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, SSD\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, and DETR\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. YOLO is a one-stage object detection method based on a convolutional neural network (CNN) and is known for its high detection accuracy and fast processing speed. YOLO11 is the latest version of the YOLO series, whereas the other YOLOs (YOLOv5, YOLOv5u, and YOLOv8) are earlier versions. SSD represents another one-stage object detection model using a CNN, whereas Faster R-CNN is a two-stage object detection model based on a CNN. In contrast, DETR is a transformer-based object-detection model. According to the respective model guidelines, the SSD model was pretrained on the Pascal VOC dataset\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e, whereas the other models were pretrained on the COCO dataset\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. To achieve high accuracy with a limited number of chromosomes, fine-tuning was performed on the pre-trained object detection models. We fine-tuned all layers of each model sufficiently using 1270 chromosomes for training and 311 chromosomes for validation. Chromosome images were resized from 1440 \u0026times; 1024 pixels to 1440 \u0026times; 1440 pixels for YOLOv5, YOLOv5u, YOLOv8, YOLO11, and Faster R-CNN; 300 \u0026times; 300 pixels for SSD; and 1125 \u0026times; 800 pixels for DETR before being input. We repeated the training, validation, and testing cycles 3 times and used the average score from the three trials.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eThe accuracy of chromosome detection with trained models\u003c/h2\u003e\u003cp\u003eFirst, we compared the accuracy of chromosome detection for each model. Specifically, the comparison focused solely on detection, disregarding the accuracy of the chromosome types predicted by the models. This comparison involved models based on YOLOv5, YOLOv5u, YOLOv8, YOLO11, Faster R-CNN, SSD, and DETR. For comparison, 459 chromosomes from WT cells were used.\u003c/p\u003e\u003cp\u003eTable. Table\u0026nbsp;1 presents the concordance rates for [email protected], [email protected], and [email protected]:0.95. The concordance rates were 88.93% (YOLOv5u), 95.13% (YOLOv8 and YOLO11), and 95.57% (Faster R-CNN). [email protected]:0.95 were 69.87% (YOLOv5u), 74.20% (YOLOv8), 75.27% (YOLO11), and 70.27% (Faster R-CNN). The models based on YOLOv5, SSD, and DETR exhibited significantly lower accuracy. In the case of the YOLOv5-based model, this may be primarily due to its use of the same head design as the older YOLOv3\u003csup\u003e22\u003c/sup\u003e because YOLOv5u, which replaces the YOLOv5 head design with the one used in YOLOv8, showed a significant improvement in accuracy compared to YOLOv5. An outdated head design may limit the ability to detect small chromosomes. Similarly, the SSD-based model likely struggled because of the required reduction in the input image size according to the model specifications. The reduction in the input image size likely caused the chromosomes to appear even smaller, making detection more challenging. Unlike for DETR, the performance of DETR may be limited by the absence of a multiscale feature extraction mechanism. Consequently, small object features may have been lost during processing in the CNN backbone of DETR, leading to reduced detection accuracy.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eDetection and classification of chromosomes with trained models\u003c/h2\u003e\u003cp\u003eNext, we compared the classification accuracies of the trained models. Four models with high chromosome detection performances were examined, and 459 chromosomes from WT cells and 465 chromosomes from \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells were used as the test data. Table\u0026nbsp;2 presents the results for the concordance rates, [email protected], [email protected], and [email protected]:0.95 for each model. The YOLOv8-based model achieved the highest concordance rate and mAP for both WT and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, with a maximum concordance rate of 89.40%.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDetection of SCC defects in\u003c/b\u003e \u003cb\u003eDDX11\u003c/b\u003e\u003csup\u003e\u003cb\u003e\u0026minus;/\u0026minus;\u003c/b\u003e\u003c/sup\u003e \u003cb\u003ecells with trained models\u003c/b\u003e\u003c/p\u003e\u003cp\u003eSince the analyses output the distribution of each type of chromosome in WT and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, we compared the distribution in each model and confirmed whether the models produced similar results to manual analysis. The rates of each type of chromosome obtained from manual analysis or the trained model are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. In manual analysis, the proportion of each chromosome type in \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, compared to WT cells, changed as follows: type A decreased from 25.1\u0026ndash;24.9%, type B decreased from 65.8\u0026ndash;41.3%, and type C increased from 9.1\u0026ndash;33.8%. Although all the trained models detected the difference between WT cells and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells, especially the YOLOv8-based model, which achieved the highest concordance rates with EA, outputted the most similar distribution with manual analysis. These results demonstrate that the YOLO based model can automatically detect SCC defects without human intervention.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we demonstrated the fully automated detection of chromosomes with SCC defects using an object-detection model. We prepared chromosomal images from WT and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003e cells and fine-tuned several pre-trained object detection models using these images. By comparing the performance of each model, the YOLOv8-based model achieved a maximum concordance rate of 89.40% with EA and demonstrated results comparable to those of manual analysis. Based on these results, the technical improvements in YOLOv8 were considered effective for the tasks in this study. Specifically, all C3 modules in YOLOv5 and v5u, which split the feature map into two paths, one direct and the other processed through bottleneck structures, were replaced with C2f in YOLOv8. The C2f module incorporated a cross-stage partial bottleneck with two convolutional layers that combined high-level features with contextual information. Additionally, YOLOv5 uses an anchor-based detection method and has a coupled head structure, whereas YOLOv5u and v8 adopt an anchor-free detection method and a decoupled head structure with two separate branches for object classification and predicted bounding box regression. This design allows each branch to focus on its tasks and improves the overall accuracy of the model. A comparison between YOLOv5s and YOLOv5u revealed a significant increase in chromosome detection accuracy, demonstrating the essential nature of these architectural improvements for chromosome analysis.\u003c/p\u003e\u003cp\u003eCompared with YOLOv8, YOLO11 introduces two main architectural improvements: replacing the C2f module with the C3k2 module and incorporating the new C2PSA module. While the C3k2 module is designed to be faster and more efficient for feature aggregation, and the spatial attention mechanism in the C2PSA module allows the model to focus more effectively on important regions within the image, the performance of YOLO11 showed no improvement compared with YOLOv8. Thus, the modifications introduced to YOLO11 may not have been effective for the chromosome classification tasks in this study.\u003c/p\u003e\u003cp\u003eAlthough the accuracy of the YOLOv8-based model seems sufficient for practical use, and future YOLO models are expected to achieve better performance, we raise two other challenges to increase accuracy. The first is to improve the labeling accuracy for both the training and test data. Some chromosomes were difficult to classify into a specific category, and incorrect labeling may have occurred. Refining the training data by having multiple researchers independently review and verify the labels could effectively improve the measurement accuracy. The second is accurate detection of overlapping chromosomes. The YOLOv8-based model occasionally failed to detect overlapping chromosomes (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). To address this issue, the following two approaches were considered:\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e(i) Application of soft Soft-NMS method\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eThe current Non-Maximum Suppression (NMS) method integrates multiple bounding boxes for a single object by removing similar boxes. However, this method retains only one bounding box for overlapping objects, which can lead to detection errors. Soft-NMS measures the degree of overlap between the bounding boxes and reduces the confidence score if the overlap exceeds a certain threshold. This method allows the retention of information regarding overlapping bounding boxes while selecting the most confident bounding box. It has been reported that applying Soft-NMS improves the detection accuracy of overlapping chromosomes\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, and accuracy improvements can be expected.\u003c/p\u003e\u003cp\u003e(ii) Adoption of Oriented Bounding Box (OBB)\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eAn OBB is a rotatable bounding box aligned with the orientation of objects, which enables close enclosure of objects. By using the OBB for elongated chromosomes oriented in multiple directions, the overlap of bounding boxes can be minimized, which is expected to improve the detection accuracy of overlapping chromosomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eK. N., K. O., and T. A. conceived and designed the experiments; S. M. and M. S. performed the experiments; S. M. and T. A. wrote the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgments\u003c/h2\u003e\u003cp\u003eThis work was supported by Grants from JSPS KAKENHI (22H05072 and 25K09513) to TA and JSPS KAKENHI (22K12170) to KN.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analyzed in the current study are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePeters, J.-M. \u0026amp; Nishiyama, T. Sister Chromatid Cohesion. \u003cem\u003eCold Spring Harb Perspect Biol\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, a011130 (2012).\u003c/li\u003e\n\u003cli\u003evan Schie, J. J. M. \u0026amp; de Lange, J. The Interplay of Cohesin and the Replisome at Processive and Stressed DNA Replication Forks. \u003cem\u003eCells\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 3455 (2021).\u003c/li\u003e\n\u003cli\u003eMurayama, Y., Samora, C. P., Kurokawa, Y., Iwasaki, H. \u0026amp; Uhlmann, F. Establishment of DNA-DNA Interactions by the Cohesin Ring. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e172\u003c/strong\u003e, (465-477.e15) (2018).\u003c/li\u003e\n\u003cli\u003evan Schie, J. J. M. \u003cem\u003eet al.\u003c/em\u003e Warsaw Breakage Syndrome associated DDX11 helicase resolves G-quadruplex structures to support sister chromatid cohesion. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 4287 (2020).\u003c/li\u003e\n\u003cli\u003eDeng, W., Tsao, S. W., Lucas, J. N., Leung, C. S. \u0026amp; Cheung, A. L. M. A new method for improving metaphase chromosome spreading. \u003cem\u003eCytometry Part A\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, 46\u0026ndash;51 (2003).\u003c/li\u003e\n\u003cli\u003eIkemoto, D. \u003cem\u003eet al.\u003c/em\u003e Application of neural network-based image analysis to detect sister chromatid cohesion defects. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 2133 (2023).\u003c/li\u003e\n\u003cli\u003eMatsumoto, S., Ikemoto, D., Abe, T., Okubo, K. \u0026amp; Nishikawa, K. Automatic classification of human chromosome shapes using convolutional neural network models. \u003cem\u003eIEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences\u003c/em\u003e 2024EAL2053 (2025) doi:10.1587/transfun. EAL2053 (2024)\u003c/li\u003e\n\u003cli\u003eMatsumoto, S., Okubo, K., Abe, T. \u0026amp; Nishikawa, K. Detection model of sister chromatid cohesion defects based on Vision Transformer. in \u003cem\u003e2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)\u003c/em\u003e 27\u0026ndash;31 (IEEE, 2023). doi:10.1109/APSIPAASC58517.2023.10317257.\u003c/li\u003e\n\u003cli\u003eSergios Theodoridis \u0026amp; Konstantinos Koutroumbas. \u003cem\u003ePattern Recognition\u003c/em\u003e. (Elsevier, 2009). doi:10.1016/B978-1-59749-272-0.X0001-2.\u003c/li\u003e\n\u003cli\u003eKevin P. Murphy. \u003cem\u003eMachine Learning: A Probabilistic Perspective\u003c/em\u003e. (MIT Press, 2012).\u003c/li\u003e\n\u003cli\u003eKang, S. \u003cem\u003eet al.\u003c/em\u003e Chromosome analysis method based on deep learning: Counting chromosomes and detecting abnormal chromosomes. \u003cem\u003eBiomed Signal Process Control\u003c/em\u003e \u003cstrong\u003e91\u003c/strong\u003e, 105891 (2024).\u003c/li\u003e\n\u003cli\u003eArakawa, H., Lodygin, D. \u0026amp; Buerstedde, J.-M. Mutant loxP vectors for selectable marker recycle and conditional knock-outs. \u003cem\u003eBMC Biotechnol\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 7 (2001).\u003c/li\u003e\n\u003cli\u003eStefanics, D. \u0026amp; Fox, M. COCO Annotator. \u003cem\u003eACM SIGMultimedia Records\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 1\u0026ndash;1 (2021).\u003c/li\u003e\n\u003cli\u003eGlenn Jocher. Ultralytics YOLOv5. \u003cem\u003ehttps://github.com/ultralytics/yolov5\u003c/em\u003e (2020).\u003c/li\u003e\n\u003cli\u003eGlenn Jocher, Ayush Chaurasia \u0026amp; Jing Qiu. Ultralytics YOLOv8. \u003cem\u003ehttps://github.com/ultralytics/ultralytics\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eGlenn Jocher \u0026amp; Jing Qiu. Ultralytics YOLO11. \u003cem\u003ehttps://github.com/ultralytics/ultralytics\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eRen, S., He, K., Girshick, R. \u0026amp; Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. (2015).\u003c/li\u003e\n\u003cli\u003eLiu, W. \u003cem\u003eet al.\u003c/em\u003e SSD: Single Shot MultiBox Detector. in 21\u0026ndash;37 (2016). doi:10.1007/978-3-319-46448-0_2.\u003c/li\u003e\n\u003cli\u003eCarion, N. \u003cem\u003eet al.\u003c/em\u003e End-to-End Object Detection with Transformers. (2020).\u003c/li\u003e\n\u003cli\u003eEveringham, M., Van Gool, L., Williams, C. K. I., Winn, J. \u0026amp; Zisserman, A. The Pascal Visual Object Classes (VOC) Challenge. \u003cem\u003eInt J Comput Vis\u003c/em\u003e \u003cstrong\u003e88\u003c/strong\u003e, 303\u0026ndash;338 (2010).\u003c/li\u003e\n\u003cli\u003eLin, T.-Y. \u003cem\u003eet al.\u003c/em\u003e Microsoft COCO: Common Objects in Context. (2014).\u003c/li\u003e\n\u003cli\u003eRedmon, J. \u0026amp; Farhadi, A. YOLOv3: An Incremental Improvement. (2018).\u003c/li\u003e\n\u003cli\u003eBodla, N., Singh, B., Chellappa, R. \u0026amp; Davis, L. S. Soft-NMS \u0026mdash; Improving Object Detection with One Line of Code. in \u003cem\u003e2017 IEEE International Conference on Computer Vision (ICCV)\u003c/em\u003e 5562\u0026ndash;5570 (IEEE, 2017). doi:10.1109/ICCV.2017.593.\u003c/li\u003e\n\u003cli\u003eZand, M., Etemad, A. \u0026amp; Greenspan, M. Oriented Bounding Boxes for Small and Freely Rotated Objects. \u003cem\u003eIEEE Transactions on Geoscience and Remote Sensing\u003c/em\u003e \u003cstrong\u003e60\u003c/strong\u003e, 1\u0026ndash;15 (2022).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Sister chromatid cohesion, machine learning, chromosome analysis","lastPublishedDoi":"10.21203/rs.3.rs-7405049/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7405049/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSister chromatid cohesion (SCC) is mediated by a protein complex called cohesin and by regulatory proteins that control SCC function. A commonly used approach to evaluate the involvement of cohesin regulatory proteins is to classify the shape of the chromosomes after depletion of the target protein and analyze their distribution. Currently, shape classification is often performed manually by researchers, which is not only time-consuming but also subject to individual interpretation. Therefore, our research group developed image classification models for automating chromosome shape classification. However, in this method, unclassifiable chromosomes that arise when cropping single chromosomes must be removed manually, creating a significant barrier to the fully automated detection of SCC-defective chromosomes. In this study, we propose a method that utilizes an object detection model to detect chromosomes with SCC defects without the need to crop single chromosomes. Several pretrained object detection models were selected and fine-tuned, and their performances were compared. Among the models, the one based on You Only Look Once v8 (YOLOv8) achieved a maximum concordance rate of 89.40% with manual analysis and successfully identified differences in the distribution of wild-type (WT) and \u003cem\u003eDDX11\u003c/em\u003e\u003csup\u003e\u003cem\u003e\u0026minus;/\u0026minus;\u003c/em\u003e\u003c/sup\u003ecells. These results indicate that the YOLOv8-based model enables fully automated analysis of SCC-defective chromosomes.\u003c/p\u003e","manuscriptTitle":"Detection and Classification of Chromosomes with Sister Chromatid Cohesion Defects Using Object Detection Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-17 18:45:30","doi":"10.21203/rs.3.rs-7405049/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-12-30T04:38:53+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-11-25T15:11:19+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-25T04:31:02+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-24T02:08:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"268718009215349670246396181165126381265","date":"2025-10-15T10:55:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"133792370125466604964083452310855900465","date":"2025-10-15T07:55:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"300779656270258048465958597206404445782","date":"2025-09-10T04:51:30+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-10T04:50:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-05T04:37:39+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-23T05:10:04+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-08-23T05:07:44+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"56e2ea5c-b583-4889-ba77-ba7530b39246","owner":[],"postedDate":"September 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":54583641,"name":"Biological sciences/Biological techniques"},{"id":54583642,"name":"Biological sciences/Cell biology"},{"id":54583643,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":54583644,"name":"Biological sciences/Genetics"},{"id":54583645,"name":"Biological sciences/Molecular biology"}],"tags":[],"updatedAt":"2026-03-23T16:05:19+00:00","versionOfRecord":{"articleIdentity":"rs-7405049","link":"https://doi.org/10.1038/s41598-026-43009-6","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-03-16 15:59:01","publishedOnDateReadable":"March 16th, 2026"},"versionCreatedAt":"2025-09-17 18:45:30","video":"","vorDoi":"10.1038/s41598-026-43009-6","vorDoiUrl":"https://doi.org/10.1038/s41598-026-43009-6","workflowStages":[]},"version":"v1","identity":"rs-7405049","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7405049","identity":"rs-7405049","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}