AIComet: a reliable automated scoring program for DNA damage assessment in the comet assay Authors: Adrien Germot, Nicolas Elie, Léonie Ibazizene, Sterenn Guillemot, Poppy Evenden, Peter Møller, Raphaël Delépée, Matthieu Meryet-Figuiere, Pierre Lebailly Published: 2026 ABSTRACT The alkaline comet assay is widely used to assess genotoxicity in biomonitoring studies but lacks standardized scoring system and shows high inter-laboratory variability. Visual scoring is time-consuming, operator-dependent, and limited in sensitivity to low DNA damage levels. We developed a fully automated, open-access scoring model for comet assay analysis to address these limitations. The AIComet model performances were assessed and the model was compared to manual scoring performed in-house on the lymphocytes of 327 healthy volunteers, and compared to visual scoring by 11 trained investigators on open-access images from (Møller et al. 2023). The influence of a larger number of nuclei scored on the sensitivity of the results was evaluated. AIComet showed excellent results for comet classification and detection. Correlation with manual scoring by our laboratory investigator was excellent with a y = 1.02x − 0.155 linear regression equation and 0.92 R2. AIComet scores (51.0, 40.2, and 55.0 a.u) on the three datasets consistently fell close to the median values from the 13 investigators (50.4, 41.0, and 61.7 a.u). Regarding its sensitivity, N = 310 nuclei scored are needed to obtain a median relative difference of 10% between the score for N nuclei and the score based on all nuclei available, with a [0.4%–42.1%] 95% confidence interval.AIComet is a standardized, reproducible, time-efficient technique for comet assay scoring, reducing operator-dependent variability while improving sensitivity for detecting subtle differences in DNA damage. These findings could be important, primarily in biomonitoring studies with low-exposure settings, but also for reproducing comet assay results overall. Introduction One of the ten key characteristics of carcinogenicity is genotoxicity, which is the capacity to induce DNA damage, mutation or both (Smith et al. 2016). DNA damage includes DNA adducts, DNA strand breaks, DNA crosslinks, and DNA alkylation. Several techniques exist to assess this variety of damage such as the chromosome aberration test (CA), the cytokinesis-block micronucleus (CBMN) assay, and the comet assay (Gantchev et al. 2022; Rigutto et al. 2025). These methods are used for the regulatory safety assessment and marketing authorisation application of chemicals or pharmaceuticals as well as in human biomonitoring to assess the effects of environmental and occupational exposures. The alkaline comet assay, also referred as single-cell gel electrophoresis (SCGE) and first reported in 1984 (Ostling 1984), is a well-established test to assess DNA damage in vitro, in vivo, and in real life settings for humans, animals and fungi (Langie et al. 2015; OECD 2016). It is based on the electrophoresis of nuclei embedded in agarose gel, where DNA breaks relax the helix supercoiling and allow migration of the DNA loops toward the anode of the electrophoresis tank. It detects single- and double-strand breaks, as well as alkali-labile sites (Collins 2004; Azqueta et al. 2019), is relatively easy to use, fast and inexpensive to set up, and necessitates a relatively small number of individualized cells. Yet forty years after its introduction, the comet assay still faces limitations, mainly the absence of a consensus for interpreting its results to measure DNA damage, and a high inter-laboratory variability (Milić et al. 2021; Bivehed et al. 2023). It leads to conflicting outcomes and difficulties in comparing different studies. This variability has been extensively studied by the EU hCOMET COST Action, which has produced a compendium of recommended comet assay protocols (Collins et al. 2023). In the comet assay, DNA damage can be determined either by visual scoring or image analysis. The latter provides a variety of descriptors such as comet tail length (TL), percentage of DNA in the tail (%T DNA), and tail moment (TM). These descriptors are measured by semi-automated or fully automated systems. Open source pipelines, including the Comet Assay Software Project (CASP), the FIJI-based OpenComet plugin and the CellProfiler Comet assay pipeline (Gyori et al. 2014; González et al. 2012; Końca et al. 2003), require manual selection of the comet boundary, and are therefore highly time-consuming. Commercial analysis algorithms such as Komet, CometAssay IV and Comet Score may also need manual input and be time-consuming, but are also limited in the number of images analysed per run, restrictive in the microscopes used, and may cost thousands of dollars, so they cannot be used by all laboratories (Kumaravel and Jha 2006; G. et al. 2014; Tutty et al. 2022). A meta-analysis identified large coefficients of variation (CV) between studies using image analysis parameters, with 42.9%, 92.6% and 305% CVs for TL, %T DNA and TM respectively (Milić et al. 2021). Deep learning algorithms, which automatically segment comets including HiComet, DeepComet or AutoComet (Namuduri et al. (2019) Hong et al. (2020)) (Barbé et al. 2023), have shown limited prediction accuracy and often require additional training of the models to be applied on new datasets, so they are limited utility. DNA damage can also be determined by visual scoring, whereby nuclei are classified into categories depending on their morphologic characteristics and hence their level of damage (Møller et al. 2014). Visual scoring was used in a quarter of the studies published in 2020 (P Møller et al. 2023), but the distribution varies between research fields. In human biomonitoring studies, it was used in up to fifty percent of published studies before 2005 (Møller 2006) and was still used in more than forty percent of the published studies in 2020. Indeed, image analysis is rather unsuitable for the biomonitoring of healthy populations, since most nuclei observed are undamaged, which leads to a median of T% falling to as low as 1.6% (Milić et al. 2021). Populational studies include a large number of samples, often hundreds, so manual visual scoring is also labour-intensive and highly time-consuming, and can require weeks of analysis. As a result, the number of cells scored per individual is usually limited to a hundred cells or fewer (Collins et al. 2014). This reduces repeatability and may be insufficient to capture intra-sample heterogeneity (Forchhammer et al. 2008; Sharma et al. 2012). This limitation is especially critical in biomonitoring studies evaluating chronic low-dose exposures, where score differences between exposed and unexposed individuals are often low (Ladeira et al. 2024). The second main limitation of visual scoring is inter-investigator variability, with variation coefficients between investigators of 25 to 105% on premade slides (Garcia et al. 2004; Møller et al. 2023). There is therefore a critical need to increase the number of cells scored, to reduce inter-laboratory variations and to have a normalised tool for biomonitoring studies. In this study we developed an open-access fully automated scoring model for the comet assay. The AIComet model was designed to standardize DNA damage detection and quantification, independently of comet assay parameters. It drastically reduces manual analysis time, improves inter-study comparability, and enhances sensitivity for detecting subtle differences in DNA damage between exposed and non-exposed individuals. Materials and methods PBMC sample collection Human peripheral blood mononuclear cells (PBMCs) from ESMERALDA-14 sub-cohort were used (Roulland et al. 2004). Ethical approval was obtained from the ethical committee (CPP n°99-07, Caen, France). From heparinized blood samples, PBMCs were separated by Ficoll Hypaque gradient centrifugation following the manufacturer’s instructions. The PBMCs obtained were then suspended in a solution of RPMI1640 with 20% heat-inactivated foetal bovine serum (FBS), and 10% dimethyl sulfoxide) to ensure long-term viability of cells in liquid nitrogen. Alkaline comet assay protocol on PBMCs Cryopreserved cells were quickly thawed at 37 °C in a water bath. The aliquot was re-suspended in 6 mL of RPMI1640 with 10% heat-inactivated FBS and centrifuged at 200 g for 10 min. The pellet was resuspended in 1 mL of RPMI1640 with 10% heat-inactivated FBS. The viability of the cells was tested with Trypan blue. Cells were brought to a concentration of 80,000 cells/ mL in RPMI1640 with 10% FBS and centrifuged at 200 RCF for 10 min. The pelleted cells were resuspended in 0.8% low melting point agarose in PBS. Two gel spots of around 1200 cells were deposited per sample in a 15 μL droplet on a Gelbond® sheet. The Gelbond® sheet was immersed in cold lysis for 90 min, then directly transferred to the electrophoresis solution at 4 °C for 40 min. Electrophoresis was then performed with cold electrophoresis solution with a 0.8 V/cm power supply for 25 min. The sheet was then placed in cold PBS 1X for neutralisation twice for 5 min. The gels were then fixed by placing the sheet in cold absolute ethanol for 1.5 h. The sheet was left to dry completely, then cut to fit glass slides for microscope analysis. For staining, the slides were kept in 20 mL of TE buffer and 10 μL of SYBRTM Gold Nucleic Acid at room temperature for 20 min (Evenden et al. 2024). Visual scoring system in four classes There is no standard visual scoring system in the literature. Scoring systems are usually based on 3 to 6 categories (Møller 2006) (Collins 2004). We used a four-category scoring system where nuclei were graded into classes as follows (Fig. 1):Fig. 1Illustration of the four distinctive comet classes in a 4-class system. Class 0: No DNA migration. Completely round comet. No detectable comet tail. Class 1: Clear asymmetric DNA migration. Head is round and as big as class 0. Class 2: Tail is longer or larger than the head. Reduced but clearly visible head. Class 3: Vast majority of DNA in the tail. Head diameter very limited to absent Illustration of the four distinctive comet classes in a 4-class system. Class 0: No DNA migration. Completely round comet. No detectable comet tail. Class 1: Clear asymmetric DNA migration. Head is round and as big as class 0. Class 2: Tail is longer or larger than the head. Reduced but clearly visible head. Class 3: Vast majority of DNA in the tail. Head diameter very limited to absent Class 0: No DNA migration. Completely round comet. No detectable comet tail. Class 1: Clear asymmetric DNA migration. Head is round and as big as class 0. Class 2: Tail is longer or larger than the head. Reduced but clearly visible head. Class 3: Vast majority of DNA in the tail. Head diameter very limited to absent. Image acquisition The slides were digitized using a VS120 slide scanner (Olympus, Japan). The system was equipped with a 20× objective lens (numerical aperture, NA = 0.8), an ORCA Flash 4 camera (Hamamatsu, Japan), and a LED illumination source (LUMINACOR). For each deposit, a 5-mm × 5-mm or smaller region was imaged using the FITC filter set (excitation/emission). The z-stack depth was manually defined by the operator to ensure optimal focus. Images were saved in the proprietary Olympus CellSens VSI format, with a spatial resolution of 0.325 µm/pixel, corresponding to image dimensions of 15,390 × 15,390 pixels. Image pre-processing The images were opened using QuPath software (version 0.6). A custom Groovy script was developed to automate image normalization, tiling, and export for subsequent analyses. For each image, the primary fluorescence channel was automatically adjusted based on histogram statistics. To standardize intensity scaling across the dataset, the script computed the optimal display range by excluding the upper and lower 1.5% of pixel intensity values (saturation parameter = 0.015). This normalization step reduced brightness variability due to differences in staining efficiency, acquisition parameters, or illumination conditions. After normalization, image tiles were automatically exported using the QuPath’s TileExporter function. Each image was subdivided into overlapping tiles of 2000 × 2000 pixels with a 400-pixel overlap. Tiles were saved in PNG format. Since the number of cells varies for each deposit, the visual scores are normalized to a score per 100 cells. The scores for each sample were calculated by the formula below: Visual score = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{n}_{0}\times 0+ {n}_{1}\times 1+{n}_{2}\times 2+ {n}_{3}\times 3 }{\left({n}_{0}+{n}_{1}+{n}_{2}+{n}_{3}\right)*3}$$\end{document}*100, n0,1,2,3 the number of cells in each category. This scoring system gives a final visual score varying from 0 to 100 arbitrary units (a.u). All results present in this paper are reported as 0–100 a.u. A specific “undefined” label was assigned to structures that did not correspond to actual nuclei, or odd-looking comets (e.g. very large and irregular comets) as described by the OECD test guideline no.489, and that should not be scored (OECD 2016). Manual image annotations for model development Ground-truth annotations were produced by a trained member of the laboratory (AG) who had prior experience in scoring comet assays. Annotation was performed using a custom interface based on the LabelMe software (https://github.com/wkentaro/labelme/releases). The final dataset consisted of 2203 annotated images in PNG format along with YOLO-compatible label files, in which comets were labelled by class annotated with bounding-box coordinates (centre x, centre y, width, height). The dataset was then divided into three subsets using stratified random sampling to preserve class distribution across partitions: 70% for training, 15% for testing and hyperparameter tuning, and 15% for final validation (Peng et al. 2025). The dataset is composed of 24,581 nuclei scored by the expert, including 16,057, 2,665, 2,105, 3,351 and 403 nuclei for class 0, 1, 2, 3 or undefined, respectively. Model training and fine-tuning We employed the YOLOv11l architecture, a state-of-the-art real-time object detector combining a convolutional backbone with a unified classification–localization head (Redmon et al. 2016). The model was initialized using weights pre-trained on the Common Objects in Context (COCO) dataset, enabling efficient transfer learning for our biological imaging task (Lin et al. 2015). Model training based on our manually annotated dataset was performed using the Ultralytics YOLOv11 framework (Sapkota et al. 2025). The standard YOLO loss function, which combines localization, and classification components, was optimized using Adam with default hyperparameters to develop a final automated scoring model that we called AIComet. Scoring with the model A Python script was developed to apply the trained AIComet model to newly acquired images. For each PNG file, the model produced an annotated output image displaying bounding boxes and class labels for all detected objects. A confidence threshold with a default value of 0.3 can be adapted depending on the quality of images used. A companion text file was generated for each image, containing the predicted class and normalized bounding-box coordinates for all detections (Fig. 2). The script ultimately produced a.csv summary file reporting the total number of detected objects assigned to classes 0, 1, 2, 3, or “undefined” for each processed image. The source code for the algorithm developed in this study is available at the following URL: https://git.unicaen.fr/nicolas.elie/aicomet.Fig. 2Overview of the comet assay analysis workflow. The assay produces a Gelbond® film mounted on a slide, containing eight deposits. Each deposit is digitalized as a 5 × 5 mm image (15,390 × 1590 pixels) using a slide scanner. Images are opened in QuPath for pre-processing and subdivided into 2000 × 2000 pixels tiles. A set of manually annotated images was used to train the AIComet model. A python script applies the AIComet model to each tile, generating annotated images with bounding boxes, text files containing predicted class and boundary box coordinates, and a final.csv file compiling all detected objects classes and coordinates Overview of the comet assay analysis workflow. The assay produces a Gelbond® film mounted on a slide, containing eight deposits. Each deposit is digitalized as a 5 × 5 mm image (15,390 × 1590 pixels) using a slide scanner. Images are opened in QuPath for pre-processing and subdivided into 2000 × 2000 pixels tiles. A set of manually annotated images was used to train the AIComet model. A python script applies the AIComet model to each tile, generating annotated images with bounding boxes, text files containing predicted class and boundary box coordinates, and a final.csv file compiling all detected objects classes and coordinates YOLO evaluation and performance metrics We followed standard YOLO practice in selecting relevant metrics. The performance metrics included the mean average precision (mAP) metrics. AP is a single value encapsulating both the model’s precision and recall performances, corresponding to its capacity to avoid false positives and the ability to detect all instances. mAP50 is the mean average precision calculated at an intersection over union (IoU) threshold of 0.50. It is a measure of the model’s accuracy considering only the “easy” detections, and is given mAP50-95 is the average of the mean average precision calculated at varying IoU thresholds, ranging from 0.50 to 0.95. mAP50 and mAP50:95 values ranges from 0 to 1. It gives a comprehensive view of the model’s performance across different levels of detection difficulty. IoU refers to the overlap between the model-predicted object box and the reference object box. Evaluation of impact of number of nuclei scored AIComet can be used to score all nuclei in the scanned area, i.e. usually around two thousand scored per deposit. The effect of a larger number of nuclei scored on the sensitivity of the results was evaluated. Evaluation was performed on 1031 deposits from our laboratory scored by AIComet. For each deposit, a final reference score (scoref) was defined as the score obtained using all available nuclei. Then for each number (N) of nuclei from one to a thousand, nuclei were selected randomly from the available dataset and a subsampled score (scoren) was computed. For each deposit and each N number of nuclei, the procedure was repeated fifty times to characterize sampling variability. Results are presented using median values with 95% percentile-based confidence intervals (CI). The relative absolute (abs) difference between the subsampled score and the final reference score was calculated as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100* \frac{\mathrm{a}\mathrm{b}\mathrm{s}( {\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_{f} - {Score}_{n})}{{Score}_{f}}$$\end{document}. Moller et al. training datasets We used an open-access image dataset to compare AIComet scoring to results from thirteen external investigators from different laboratories who all had long-term experience with comet assay scoring. Three training sets of digital comet images were generated, originating from HeLa cells (set I) and THP-1 cells (set II and III), coloured by SYBR Gold, YOYO-1 or DAPI for set I, II and III respectively, and obtained from three different laboratories (Møller et al. 2023). The investigators used a five-class scoring system. Since the visual scores are expressed on a 0–100 arbitrary unit (a.u.) scale, the four-class and five-class scoring systems can be compared. Results Model validation Internal YOLO metrics for model validation We first evaluated the AIComet model with YOLO standard metrics on the validation datatset. The model achieved strong discriminative performance for classes 0, 1, 2 and 3, with per-class AP50 of 0.981, 0.815, 0.879 and 0.951 (range: 0–1) respectively, and a mAP50 of 0.854 for all classes (Fig. 3). AP50-95 for class 0, 1, 2 and 3 were 0.713, 0.536, 0.66 and 0.725 respectively, and a mAP50-90 of 0.604 for all classes. mAP values reflecting both the precision and recall performances of the model. Both mAP50 and mAP50-95 for all classes are driven downwards by the undefined class (0.644 AP50 and 0.383 AP50-95), in which classified nuclei overlap nuclei or artefacts varying considerably in size, form and brightness. Overall, these internal metrics confirm the overall high performance of the model for both detection and classification.Fig. 3Precision–recall curves showing per-class average precision and mean average precision on the validation dataset (N = 3618 nuclei, distributed as such: 2507, 375,272,402 and 62 nuclei for class 0,1,2,3, and undefined respectively). AP50 for class 0: 0.981; class 1: 0.815; class 2: 0.879; class 3: 0.951; undefined: 0.643; mAP50: 0.854 Precision–recall curves showing per-class average precision and mean average precision on the validation dataset (N = 3618 nuclei, distributed as such: 2507, 375,272,402 and 62 nuclei for class 0,1,2,3, and undefined respectively). AP50 for class 0: 0.981; class 1: 0.815; class 2: 0.879; class 3: 0.951; undefined: 0.643; mAP50: 0.854 AIComet evaluation by nuclei scored AIComet was also evaluated for its capacity to accurately class nuclei in categories regardless of the surface overlap of boxes, based on the testing dataset (excluding the undefined nuclei). The normalized confusion matrix (Fig. 4) showed an excellent object classification, with less than 5% misclassification overall. The percentage of nuclei accurately scored per class was 0.98, 0.95 and 0.94 for classes 0, 2 and 3, respectively. The performance on class 1 nuclei was lower but remained very satisfactory with 0.81 percent accuracy. Misclassifications were equally distributed between true class 0 and class 2 nuclei. The model missed 1.06% of nuclei overall (Table 1), with a tendency to miss more class 3 nuclei (2.65% missed). These internal l metrics confirms an overall high-performance model, for both detection and classification parameters.Fig. 4Normalized confusion matrix showing agreement between expert annotations (ground truth) and AIComet classifications (predicted class) on the testing dataset without undefined nuclei (N = 3572 nuclei distributed as such: 2400, 370, 312 and 490 nuclei for class 0, 1, 2 and 3 respectively. Agreements were 0.98, 0.81, 0.95 and 0.94 for class 0, 1, 2 and 3 respectivelyTable 1Distribution of nuclei and of nuclei missed by AIComet per class in % and absolute values in the testing datasetCategoriesGround-truth nuclei numberNuclei missed by model % (N)Class 02,4000.67% (16)Class 13701.35% (5)Class 23121.28% (4)Class 34902.65% (13)Total3,5721.06% (38) Normalized confusion matrix showing agreement between expert annotations (ground truth) and AIComet classifications (predicted class) on the testing dataset without undefined nuclei (N = 3572 nuclei distributed as such: 2400, 370, 312 and 490 nuclei for class 0, 1, 2 and 3 respectively. Agreements were 0.98, 0.81, 0.95 and 0.94 for class 0, 1, 2 and 3 respectively Table 1Distribution of nuclei and of nuclei missed by AIComet per class in % and absolute values in the testing datasetCategoriesGround-truth nuclei numberNuclei missed by model % (N)Class 02,4000.67% (16)Class 13701.35% (5)Class 23121.28% (4)Class 34902.65% (13)Total3,5721.06% (38) Distribution of nuclei and of nuclei missed by AIComet per class in % and absolute values in the testing dataset AIComet evaluation compared to manual scoring Comparison with internal investigator We compared AIComet to manual scoring by a trained investigator from our laboratory (PE). Scoring was performed on isolated PBMCs of 327 healthy volunteers from a cohort of farmers (Roulland et al. 2004). Visual score from manual scoring was based on 200 nuclei scored per individual. The regression equation of this comparison is y = 1.02x − 0.155, and the determination coefficient R2 = 0.92. The slope of the regression line (1.02) suggests a near one-to-one correspondence between model and manual scores across the full range of values. The intercept of 0.155 indicates minimal systematic bias at low score values (Fig. 5).Fig. 5Correlation between visual scores of DNA migration (0—100 a.u.) in 327 individuals from ESMERALDA-14 sub-cohort, scored by AIComet (Model score) and by an experienced investigator (Manual score) from our laboratory (PE). Manual scoring was assessed on 200 nuclei per individual Correlation between visual scores of DNA migration (0—100 a.u.) in 327 individuals from ESMERALDA-14 sub-cohort, scored by AIComet (Model score) and by an experienced investigator (Manual score) from our laboratory (PE). Manual scoring was assessed on 200 nuclei per individual Comparison with external investigators AIComet performances were also compared to those of external investigators on three sample sets available in the literature. The model returned visual scores of 51.0, 40.2, and 55.0 for datasets I, II, and III, respectively, very closely matching the corresponding median visual scores of 50.4 (range 41–57.5), 41.0 (range 23.8–51.5), and 61.7 (range 41.2–71.1) of the experts (Fig. 6). The model identified every nucleus from dataset I and dataset II, and failed to detect only one of 64 (1.5%) cells in datasets III.Fig. 6Variation in visual scores of samples sets of digitalized comet images by 13 investigators (black, based on a 5-class scoring system) and AIComet model (red, based on a 4-class scoring system). The score from each sample set have been normalized to 0–100 arbitrary unit (a.u.). Images comes from Hela cells (set I) and monocytic THP-1 cells (set II and III). Results: 51.0, 40.2, and 55.0 for AIComet, 50.4, 41.0 and 61.7 median value for experts on datasets I, II and III respectively Variation in visual scores of samples sets of digitalized comet images by 13 investigators (black, based on a 5-class scoring system) and AIComet model (red, based on a 4-class scoring system). The score from each sample set have been normalized to 0–100 arbitrary unit (a.u.). Images comes from Hela cells (set I) and monocytic THP-1 cells (set II and III). Results: 51.0, 40.2, and 55.0 for AIComet, 50.4, 41.0 and 61.7 median value for experts on datasets I, II and III respectively Increased sensitivity with number of nuclei scored We evaluated the effect of the number of nuclei scored from the images of 1031 deposits. The relative difference between the reference score and the score for a given number of nuclei decreased as the number of nuclei scored increased (Fig. 7). For 100 nuclei scored, i.e. a conventional number in research articles, the median relative difference was 19.4%, with a [0.8–79.4] bootstrap 95% confidence interval, reflecting a large dispersion. Scoring a minimum of 310 nuclei was needed to obtain a median relative difference of 10%, with a [0.4–42.1] 95% CI. To obtain a higher limit of the confidence interval smaller than 20%, a minimum of 900 nuclei was scored, with a median of 4.4.Fig. 7Evolution of the relative difference from the final score (%) with increasing number of nuclei scored per individual (median and 95% interval). Thresholds indicate that at least 90, 310, and 900 nuclei are needed to reach relative differences below 20%, 10%, and a 95% upper confidence limit below 20%, respectively Evolution of the relative difference from the final score (%) with increasing number of nuclei scored per individual (median and 95% interval). Thresholds indicate that at least 90, 310, and 900 nuclei are needed to reach relative differences below 20%, 10%, and a 95% upper confidence limit below 20%, respectively Discussion The primary objective was to develop a reliable open-access automated scoring model that can be broadly used by the scientific community to considerably reduce manual analysis time, improve inter-laboratory comparability, and enhance precision and sensitivity for detecting DNA damage. The model showed excellent overall performances. The confusion matrix compared AIComet classification to manual scoring with overall less than 5% missclassification, which can be partly explained by the assignment of discrete categories to a biologically continuous spectrum of DNA damage. Regarding the capacity of the model to detect all instances, only 1% of the nuclei were not detected by AIComet on our dataset. Misdetection was not homogenous among comet classes, with 2.6% of missed class 3 objects (Fig. 1). This difference may be explained by the very diffuse DNA fragments of some very highly damaged nuclei, sometimes leading to overlapping with other nuclei that is easily distinguishable manually but for which the model was not trained. This limitation can be addressed by a slight decrease in cell concentration to avoid overlapping of heavily damaged cells (i.e. comet described as hedgehogs or ghosts), the main mis-detected nuclei (Azqueta et al. 2019). On the other hand, the model showed remarkable ability to accurately score very highly damaged class 3 nuclei, with diffuse DNA that were not detected manually by any or most of the experts, depending on their screen brightness setups and scoring conditions. The ability of the model to score thousands of nuclei per sample represents a major methodological breakthrough regarding the sensitivity of comet assay results. Figure 7 highlights the value of scoring more than the typical 100 nuclei scored to increase the sensitivity of the results, in accordance with a study on the influence of the number of cells scored by image analysis (Sharma et al. 2012). That study reported that the scoring of 600 cells increased sensitivity compared with the scoring of 100 cells, especially for low levels of DNA damage. This is particularly critical in biomonitoring studies evaluating chronic genotoxic exposures in healthy populations, where small differences in DNA-damage levels could be relevant. The manual scoring of 600 or more nuclei per individual seems unrealistic, whereas similar numbers can be scored easily with our algorithm. AIComet require about 10 s to score 3500 nuclei, corresponding to the mean number of nuclei on two replicate deposits for an individual in our assay conditions. Regarding the variability of comet assay results, Moller et al. reported an overall mean coefficient of variation of 15% on the scoring of the three open access datasets by 13 trained investigators, 64% of the variation explained by intra-investigator variability and 36% by inter-investigator variability (P. Møller et al. 2023). By design, the AIComet model eliminates inter-investigator variability. Nor is there any intra-operator variability, as repeated analysis of the same image yields identical results (data not shown). However, sampling-related variability was not assessed by Møller et al., as all investigators evaluated exactly the same predefined cells from selected sub-images (Møller et al. 2023). This study design focuses on inter-investigator variability on the same objects, but does not reflect the variability introduced by randomly selecting different nuclei across the entire slide, which is closer to laboratory practice when scoring only a pre-defined number of nuclei. It is therefore likely that inter-investigator differences and coefficients of variation would have been greater if each investigator had independently scored selected nuclei from the full image. At low levels of DNA damage, scoring only a few additional damaged nuclei can substantially influence the final score. Scoring all nuclei present on the slide circumvents this sampling-related variability. Differences in image sources always constitute a challenge for automated image analysis (Barbé et al. 2023). Furthermore, the ground truth classification was established by a single trained expert from our laboratory and could lead to misleading results for other users of the model if the expert is not fully familiar with the scoring systems in the literature and remains consistent in scoring. We therefore compared our model to open-access image datasets produced under experimental conditions that were different from those for model training and compared our model scores to other investigators’ scores. The scoring of these images was assessed twice. First on the images accessible online in a PowerPoint file. The images obtained through extraction from presentation software show a noticeable degradation of signal quality, likely related to compression and format conversion processes leading to information loss. Analysis of the intensity histograms reveals a strongly reduced grayscale dynamic range. The quality of images provided was adequate for manual scoring but prevented our algorithm from scoring some of the cells, resulting in 8% of nuclei not detected over the 3 datasets, and visual scores of 43.6, 39.0, and 55 a.u for dataset I, II and III respectively. The scoring was assessed again on the original images send by the corresponding author of the article, with the results presented above (Møller et al. 2023). AIcomet accurately scored all three open-access image datasets, with results consistently falling within the range of manual scoring by the 11 expert investigators, very close to the reported median values (Møller et al. 2023). This supports the robustness and reliability of AIComet for all users applying the hCOMET guidelines. The training datasets of Moller et.al were generated with material differing from our own, THP-1 and Hela cells instead of human PBMCs, and different staining procedures, comet assay setups and slide scanning systems. The consistent performance of AIComet across these heterogeneous datasets despite the low number of nuclei considered therefore suggests its good generalizability and highlights its potential for broader applications in diverse laboratory settings following the hCOMET recommendation (Collins et al. 2023). Moreover, when the model was run on images from organoids and compared to manual scoring, the results were similar to those obtained with PBMCs (results not shown). The code used to develop the algorithm is publicly available online, allowing researchers to extend the existing model by incorporating their own images or to construct new training datasets. This flexibility enables users to retrain or adapt the model to better suit their specific assay conditions or image characteristics if certain comet morphologies or imaging settings are not optimally scored by the current model. Overall, these results confirm that AIComet is a reliable, open-access high-throughput tool for the automated scoring of comet assay images. It provides the scientific community with a standardized method for analysing comet assay data that avoids inter- and intra-investigator variation and sampling-related variability, and allows for more precise and time-saving quantification of DNA damage, while allowing comparability between studies.