Application of the Cytokinesis-Block Micronucleus Assay for High-Dose Exposures Using Imaging Flow Cytometry

In the event of a radiological or nuclear incident, information on the magnitude of the dose of radiation received by people potentially exposed to radiation is essential for medical and psychological care, especially during such incidents when the exposure is unknown. To assess for genetic damage in cases where the dose is not known, several different cytogenetic assays for biodosimetry can be used. One assay is the well-established cytokinesis-block micronucleus (CBMN) assay [Fenech and Morley, 1985; Fenech, 2007; Wang et al., 2019]. This assay examines radiation-induced genetic damage that is present in human cells as small, rounded micronuclei (MN) formed from whole or fragmented chromosomes in the cytoplasm of binucleated cells (BNCs). These MN are surrounded by their own nuclear envelope instead of being incorporated inside one of the two main nuclei after karyokinesis and can persist for up to 1 year after they are formed [Fenech, 2000].

The CBMN assay quantifies the frequency of chromosome damage (MN in BNCs [MN/BNC]) in human peripheral lymphocytes [Fenech and Morley, 1985, 1986; Fenech, 2010] by manual microscopy. While this assay is considered as accurate as other biodosimetry assays (e.g., dicentric chromosome assay) and dose responsive, it requires 2–3 days to process samples even before analysis can begin. This method is labor-intensive and time-consuming. The slide quality varies greatly depending on preparation details, which can have a significant impact on the accuracy of the scoring. Following a radiological/nuclear incident, the delay in receiving accurate dose estimates impedes health care providers from determining the proper treatment course for people potentially exposed, leading to misallocation of resources, late treatment, and poorer outcomes [Medical Management of Radiation Injuries, 2020].

Manual scoring of the slides can also be time-consuming, and attempts have been made to automate or semi-automate image analysis to help increase throughput of samples and reduce the time needed for sample analysis. Automated slide-based image analysis microscopy is one such technological advancement that enables capturing images, higher sample throughput, and increased speed of data collection and removes scorer variability [Romm et al., 2013; Depuydt et al., 2017].

Recently, our team has automated the CBMN assay by adapting it to imaging flow cytometry (IFC) that integrates the statistical power of traditional flow cytometry with the image clarity of microscopy. This improves upon traditional scoring methods by increasing the measurement accuracy of damage within the cells [Rodrigues et al., 2014a, 2014b, 2016a]. IFC also eliminates the need to create microscope slides and can automatically score more BNCs, which increases the speed and statistical power of the assay [Rodrigues, 2018]. Furthermore, images collected on the cytometer can be batch-analyzed with preestablished templates ensuring reproducibility.

The use of small-volume blood samples in combination with the automated CBMN assay has also been recently explored [Rodrigues et al., 2016]. Typically, 1 mL of blood is used for the microscope-based CBMN; however, much less volume is required to adequately analyze samples for dose estimation. The use of smaller volumes facilitates sample collection processes such as finger prick instead of venipuncture. It also reduces the time needed for sample processing and the supplies required such as vials and reagents. This in turn reduces the storage and costs needed for keeping supplies on hand and reduces waste (expired reagents), increasing efficiencies for emergency preparedness. It can, however, increase the sample acquisition time on the IFC as it results in a lower density of cells per sample. While these advances have increased efficiencies in the CBMN assay, the two major economies focused on in this manuscript are as follows:

• 1.

  Use of the autosampler for IFC: Our automated CBMN assay for radiation biodosimetry used manual loading of each sample into the IFC which is time-consuming and requires constant hands-on operation. We sought to reduce the amount of manual manipulation required and increase the number of samples run by using a 96-well plate on an autosampler to allow hands-free, unattended operation. To validate the use of the autosampler, the data acquired with its use and manual acquisition were compared to determine whether they were statistically equivalent so that going forward, the automated system could be used, particularly in situations where hundreds of samples might require analysis. This would allow for unattended analysis and samples to be run through the night.

• 2.

  Extension of calibration curves to higher doses: For radiation biodosimetry, dose calibration curves are used to convert the radiation-induced biological damage in cells to the dose of radiation delivered to the cells. Typically, the curves are generated for doses between 0 and 4 Gy and are linear-quadratic in shape [McNamee et al., 2009; Vral et al., 2011]. Above 4 Gy, the curve starts to plateau followed by a decrease in MN/BNC, forming a downward parabola such that for every value of MN/BNC, there are two possible doses [Müller and Rode, 2002]. This work will explore the use of additional parameters in IFC and statistical analysis, allowing the discrimination between the two possible dose estimates, thus permitting extension of the dose range up to 10 Gy.

Whole blood was collected from 10 healthy adult (5 males/5 females) donors (between the ages of 20–60 years) with informed consent and approval by Health Canada’s Research Ethics Board (protocol REB 2002-0012H). All donors selected were nonsmokers in relatively good health at the time of donation. The donors had no obvious illnesses and no known exposures to medical ionizing radiation within the last 12 months. Blood was drawn by venipuncture in 10 mL lithium-heparinized Vacutainer^® tubes (BD Vacutainer^™, Mississauga, ON, Canada). Aliquots of blood samples (1.2 mL) were dispensed into 5 mL round bottom Falcon flow tubes (Corning Life Sciences, Corning, NY, USA) and transported to an XRAD 320 X-ray cabinet biological irradiator (Precision X-Ray, N. Branford, CT, USA). The tubes were placed on their side in the middle of the cabinet on a Styrofoam plate and exposed to X-rays at a dose rate of 1.7 Gy/min. Irradiations were performed at 250 kVp and 12.5 mA with a 2-mm Al filter. All doses were calibrated using a PTW TW30010-10 ion chamber and a PTW UNIDOS T10002 electrometer (PTW Dosimetry, Lörracher Strasse 7, 79115 Freiburg) with N_k = 48.3 mGy nC⁻¹ at 250 kV (calibrated by the National Research Council, Ottawa, ON, Canada) assuming air kerma to be equal to dose [Rodrigues, 2018]. Two sets of exposures between 0 and 10 Gy were conducted: (1) for the generation of the dose-response curve (“curve data”), samples from each of the 10 donors were exposed at intervals of 1 Gy, and (2) for validation of the method (“test data”), samples from 6 donors were exposed to doses between 0 and 10 Gy in increments of 0.5 Gy.

Aliquots of 200 μL from the irradiated blood samples were transferred to 1 mL Matrix tubes (Thermo Fisher Scientific, Waltham, MA, USA) containing 800 μL complete media (86.5% RPMI 1640 media (Gibco, Waltham, MA, USA), 10% heat-inactivated FBS (Sigma-Aldrich, St. Louis, MO, USA), 2.5% phytohemagglutinin (Gibco), and 1% l-Glutamine/Pen Strep (Sigma-Aldrich)) and gently mixed. The samples were incubated at 37°C, 5% CO₂, for 24 h. After 24 h, cytochalasin-B (Sigma-Aldrich) was added to cultures at a final concentration of 6.0 μg/mL for the purpose of blocking cytokinesis of proliferating lymphocytes. The samples were incubated for an additional 44 h prior to the fixation process [Wang et al., 2019].

The fixation process was adapted from Qi et al. [Wang et al., 2019]. Briefly, samples were centrifuged and resuspended in a 75-mm KCl solution (Sigma-Aldrich) for a hypotonic soft fixation of cells, and then centrifuged and resuspended in FACS lysing solution (Becton, Dickinson and Company [BD], Franklin Lakes, NJ, USA) to lyse red blood cells and fix lymphocytes. Cell suspensions were then washed one time with FACS lysing solution to complete the lymphocyte fixation and washed three times with PBS pH 7.4 (Teknova; Hollister, CA, USA) before being resuspended in PBS to a total volume of 100 μL per cell suspension. Cells were stored at 4°C for up to 2 weeks until being run on an IFC (ImageStream^®X Mark II (ISXMkII), Cytek Biosciences, Seattle, WA, USA).

Immediately before analysis by IFC, DRAQ5 (eBioscience, San Diego, CA, USA) was added to the sample at a final concentration of 5 μm in 100 μL of fixed cell suspension and incubated at room temperature for 20 min. Samples were introduced into the ISXMkII either manually one at a time in 1.5 mL Eppendorf tubes or using a 96-well plate. All samples were run on the ISXMkII system (Millipore Sigma, Seattle, WA, USA) at ×40 magnification with the 642-nm laser set at approximately 30 mW, adjusted as necessary so that the fluorescence intensity of the desired population was between 1,000 and 3,000 raw max pixels in channel 5. The Brightfield LED was set to 800 mW, and the side scatter (785 nm) laser was set to 2.5 mW. Images were collected on a single camera system with Brightfield images collected in channel 1 and DRAQ5 images collected in channel 5. Data were collected using the ISX INSPIRE^® software (version 200.1.388.0) with only the area feature applied. Events with an area smaller than 100 pixels (25 μm²) were gated out to minimize the collection of small debris. Typically, at least 50,000 events were collected in each sample.

All data were processed using IDEAS (version 6.2); BNC, mononucleated (MONO), trinucleated (TRI), quadranucleated (QUAD), and non-apoptotic (NAP) cells were identified along with BNCs containing 0, 1, 2, 3, or 4 MN (0 MN, 1 MN, 2 MN, 3 MN, 4 MN) using the gating strategies outlined in [Rodrigues et al. 2016b, 2018].

The following ratios were included for analysis: TotalMN/BNC (TotalMN = MN + 2*2 MN + 3*3 MN + 4*4 MN), BNC/NAP, MONO/NAP, TRI/NAP, and QUAD/NAP. The ratio TotalMN/BNC was treated as a rate following a Poisson distribution, whereas the other ratios were treated as proportional outcomes less than 1 following a binomial distribution. Counts of each population were determined using both manual and autoplate acquisition, allowing these results to be combined and compared.

The first objective was to determine if the data from the two acquisition methods, “autoplate” and “manual,” were similar with respect to the rate of TotalMN/BNC. Simultaneous random coefficient model (RCM) with Poisson errors was fit to the autoplate and manual data relating TotalMN/BNC to radiation dose. The fixed portions of the simultaneous RCM were used to compare if the autoplate and manual counting methods were similar. See online supplementary material “Comparison of autoplate and manual counting methods” for more details (for all online suppl. material, see https://doi.org/10.1159/000532124).

The random portion of Eq. 1 was not included in predicting the dose of exposure for a new observation of TotalMN/BNC (see “Solutions to the quadratic model” in online suppl. material). The fixed portion of the quadratic RCM reduces to a generalized linear model and allows one to predict dose in closed form. The variance-covariance structure of the fixed parameter estimates was used in the delta method to determine an approximate 95% CI about the predicted dose estimate. The fixed parameter estimates of Eq. 1 were used to determine an estimate of dose (d_e) based on the observed TotalMN_o/BNC_o by using the solutions of the quadratic formula (see “Solutions to the quadratic model”in online suppl. material)

If the observed TotalMN/BNC is greater than p_m, the model is inadequate for predicting dose. In this situation, a series of successive quadratic Poisson regression models were developed to be used as calibration curves. The upper 99% CI to Eq. 1, a quadratic Poisson regression fit to the maximum TotalMN/BNC in each dose group, and the upper 99% CI to this “maximum” quadratic Poisson regression can all be used as calibration curves. See “Successive calibration curves” in online supplementary material for a detailed discussion of the approach.

For the fixed portion of the quadratic RCM and the model based on the maximum values, the CI of the dose value was estimated using the delta method with the variance-covariance estimates of the parameter estimates. When extrapolating to the upper 99% CI of these models, bootstrap methods were used to determine the CI of the estimated dose with 1,000 replicates.

The inverse prediction of dose based on the observed rate of TotalMN/BNC produced two possible doses of exposure in the dose range of 0–10 Gy. A logistic discriminant function (LDF) was applied to determine if the dose of exposure was below or above d_m. The LDF predicted the probability of observing an outcome being below or above d_m based on the remaining endpoints collected in the assay (BNC/NAP, MONO/NAP, TRI/NAP, and QUAD/NAP). To this end, the data were recoded to indicate if the observation was exposed to a dose less than or equal to d_m (coded as 1) or greater than d_m (coded as 2). The LDF was then used to compute the probability for each observation of belonging to class 1 (less than or equal to d_m). If the calculated probability based on LDF was less than or equal to a predetermined cut-off probability, then the observation was classified as belonging to class 1 (less than or equal to d_m). The predetermined cut-off probability was chosen to maximize the proportion of correct classifications through a sensitivity analysis.

Calibration curves and LDF were based on the curve data and not on the test data for which we were trying to estimate dose. A similar approach was used whether we based the results on the autoplate data, manual or combined data. In this paper, we describe the method for the combined autoplate and manual data.

Step 1: The appropriate calibration curve was determined by the observed value of TotalMN/BNC for each of the test data results. If the observed TotalMN/BNC was less than or equal to 0.23, then the mean mode (model 1) was used (online suppl. Fig. S1 black curve). If the observed TotalMN/BNC was greater than 0.23 but less than or equal to 0.29, then the model based on the upper 99% CI of the mean model (model 2) was used (online suppl. Fig. S1 green curve). If the observed TotalMN/BNC was greater than 0.29 but less than or equal to 0.34, then maximum model (model 3) was used (online suppl. Fig. S1 red curve). If the observed TotalMN/BNC was greater than 0.34 but less than or equal to 0.37, then upper 99% CI of the maximum model (model 4) was used (online suppl. Fig. S1 blue curve). Finally, if the observed rate of TotalMN/BNC was greater than 0.37, then the dose of exposure was taken to be equal to where model 4 achieves its maximum, at a dose of 5.9 Gy.

Step 2: The LDF was used to fit variables BNC/NAP, MONO/NAP, TRI/NAP, and QUAD/NAP from the curve data set (see plots of these variables in online suppl. Fig. S3). This model was used to classify if an observation from the test data was exposed to a dose less than or equal to d_m or above d_m. A probability was determined that maximizes the proportion of correct classifications based on the LDF (P_c). If the probability of the new observation, based on the LDF, was less than or equal to P_c, then it was assumed that the observation was exposed to a dose less than or equal to d_m. If the probability of the new observation, based on the LDF, was greater than P_c, then it was assumed that the observation was exposed to a dose greater than d_m.

Step 3: The information from Step 2 indicated whether the expected dose was less/greater than d_m. This in turn indicated which root of the calibration curve in Step 1 was the estimated dose of exposure. In the case where the upper 99% CI of the regression models 1 or 3 (i.e., for models 2 and 4, respectively) was extrapolated, then the R library rootSolve was used to solve for the roots of Eq. S3 in “Successive calibration curves” in online supplementary material.

Step 4: The delta method with the fixed parameters based on the quadratic RCM (model 1) or parameter estimates (model 3) as well as the variance-covariance structure of these parameter estimates was used to get an approximate 95% CI about the dose estimate. If extrapolating to models 2 or 4, bootstrap methods were used to determine an upper and lower bound on the estimated dose.

It can happen that the dose estimate based on the calibration curve is outside the limits of the acceptable dose range. For example, the dose could be estimated as less than 0 Gy or higher than 10 Gy. In these cases, where the dose estimate was not within 0–10 Gy, the dose estimate was set to the nearest limit.

All analysis was carried out in R version 4.1.0 [R Core Team, 2021]. Functions from the R libraries include nlme [Pinheiro et al., 2021], lme4 [Bates et al., 2015], rootSolve [Soetaert, 2009], msm [Jackson, 2011], and boot [Canty and Ripley, 2021].

Online supplementary Figure S2, panels A–C (Comparison of autoplate and manual counting methods) represent the RCM with Poisson errors fit to the autoplate data, manual data, and combined data, respectively. Online supplementary Figure S2, panel D represents the fixed portion of the quadratic RCM for the three data sets graphed together. The quadratic RCM for the autoplate and manual methods is considered statistically similar (likelihood ratio $χ32=2.1,p=0.558$⁠). There was insufficient evidence to suggest that the models for the autoplate and manual methods differed. Nevertheless, in the analysis that follows, the results for the combined data are still compared to the results for autoplate and manual data to determine the effect on the dose estimates.

Model 1 as defined by Eq 1 is fit to the combined data. The maximum rate TotalMN/BNC (or p_m) of the quadratic function is attained at 0.23 when the dose (d_m) is equal to 5.3 Gy. Therefore, the data were classified as class 1 if the dose of exposure is less than or equal to 5.3 Gy and class 2 if the dose of exposure was greater than 5.3 Gy. The LDF was used to calculate the probability of belonging to class 1 (less than or equal to 5.3 Gy) or 2 (greater than 5.3 Gy). If the probability (as determined by the LDF) was less than or equal to 60%, then the observation was classified as belonging to class 1. If the probability was greater than 60%, then the observation was classified as belonging to class 2. The 60% cut-off was based on a sensitivity analysis to determine the greatest proportion of data being correctly classified.

Based on the combined data, the LDF was able to correctly classify 89.1% of the data as being above or below 5.3 Gy exposure. For the autoplate data alone, the LDF was able to correctly classify 87.9%, and for the manual data alone, the LDF was able to correctly classify 88.6%. Figure 1, which demonstrates how the methodology is applied, presents the calibration function based on the combined data set. For example, suppose we have a given observation TotalMN/BNC of 0.10 (green horizontal line), model 1 calculates two possible doses at 2.1 Gy and 8.5 Gy (light blue vertical lines). Based on the results from the LDF, it was determined that the dose should be in the 0–5.3 Gy dose range. An estimated dose that is below 5.3 Gy with TotalMN/BNC equal to 0.10 is approximately 2.1 Gy, with corresponding 95% CI (1.5–2.7) Gy (CI indicated with purple vertical lines).

Preliminary validation of the calibration curve was conducted using data from the samples used to generate the dose-response curve (data not shown).

Full validation was performed with a test data file that was independent of the curve data sets. There were 288 samples in total, collected either manually or with the 96-well autoplate, with irradiated blood, where the dose of exposure was blinded. To estimate dose from our developed calibration models, we applied all three models (based on combined, autoplate, and manual data) to determine which model was most appropriate. Our first step was to use the LDF based on the original CBMN data set to classify whether the dose was less than or equal to 5.3 Gy or above 5.3 Gy. Based on the combined data LDF, 86.8% of the test data were classified correctly. Similar results were attained when we used the LDF for autoplate data (87.2% correct classification) and manual LDF (87.2% correct classification).

The second step was to then use the classification from the LDF and the observed rate of TotalMN/BNC to use the appropriate calibration curve to estimate dose. The method was carried out using the calibration curve of the combined data, autoplate data, and manual data.

Figures 2 and 3 present the probability that the estimated dose is within 0.5 Gy and 1 Gy of the true dose, respectively. From Figure 2, the probability that the estimated dose was within 0.5 Gy of the true dose was greater than 55% when the true dose was less than or equal to 2.5 Gy, and then dropped with increasing true dose for both manual and autoplate acquisitions. In Figure 3, the probability that the estimated dose was within 1 Gy increased substantially to greater than 90% regardless of acquisition method, when the true dose groups were less than or equal to 2.5 Gy. The lowest probability of being within 1 Gy of the true dose was observed at 5, 5.5, and 6 Gy, likely due to these values being close to the cut-off dose for classification in the LDF. It should be noted that the number of samples for each dose varied with half-Gy values having only 6 samples. For these doses, large differences in the percentages between methods correspond to a change in only 1 or 2 samples.

Figure 4 presents the probability that the true dose fell inside the 95% CI of the estimated dose. Again, there was greater success at lower doses, with decreasing success at the higher doses.

Figure 5 presents the average estimated dose at each dose level versus the true dose. In this figure, we can also observe how often the estimated dose is within 0.5 Gy of the true dose (light gray band) and within 1 Gy of the true dose (dark gray band). Since the points represent the average estimated dose at each true dose level, an estimate of the standard error (SE) of the mean is also presented to give an indication of the spread of the estimated doses at each true dose level. Below 3 Gy, the average estimated dose was always within 0.5 Gy of the true dose. As the true dose increased, this accuracy decreased and, above 7 Gy, the estimated dose more often underestimated the true dose, sometimes with a difference greater than 1 Gy. It is observed that at increased true dose the SE of the average estimated dose also increased indicating poorer reproducibility.

Finally, Figures 6 and 7 present the RBIAS and RMSE results. The absolute value of the RBIAS ranged between 0 and 0.52. The highest RBIAS was observed at 0.5 and 3 Gy, based on the manual and autoplate calibration models, respectively. RMSE was considered low at all true dose groups except at 0.5 Gy where the RMSE was 1 unit for the combined calibration model and 1.2 units for the autoplate calibration model. Recall that RBIAS and RMSE are both divided by the true dose value; therefore at dose 0.5 Gy, the RBIAS is multiplied by a factor of 2 and RMSE is multiplied by a factor of 4, increasing these statistics. The results are comparable between the three-calibration models based on the combined original CBMN data, autoplate or manual, although there are instances when one model performed better than the others at specific dose groups.

The methodology for analysis of the CBMN assay by IFC has been previously reported, highlighting its advantages [Rodrigues et al., 2014a, 2014b, 2016a; Rodrigues, 2018]. The purpose of this manuscript was to validate the use of the autoplate as compared to manual sample acquisition and to demonstrate a novel data analysis approach that extends the useable range of the assay from 0–4 Gy to 0–10 Gy.

To address our first goal, calibration curves were generated for data acquired using manual and automated sampling and the combined data set to determine whether there were differences in the response. Each curve was fit with a quadratic RCM with Poisson errors. There was insufficient evidence to suggest a significant difference between the two acquisition methods. Nevertheless, the analysis of the three data sets was continued through the validation of the calibration curve with blind irradiated samples. This allowed us to compare the dose estimates generated by each data set to determine whether data from the manual and autoplate acquisition could be combined.

Using the extensive CBMN data that are generated from an IFC, we were able to improve the analysis to extend the region of applicability of the assay to doses higher than 4 Gy. Traditional CBMN analysis uses only the MN/BNC rate as a function of dose to estimate doses received by individuals. This response increases up to about 4 Gy but then starts to drop off due to the increase in damage in the cells that causes a cell cycle delay. The reduction in BNC frequency results from a decrease in cell proliferation and an increase in blocking of cells in the G2 phase of the cell cycle following high-dose exposure [Rodrigues et al., 2014b]. Furthermore, at high doses, highly damaged cells are unable to survive mitosis and proceed to cytokinesis. This decrease has been reported in other studies when using automated methods to score the CBMN assay [Rodrigues et al., 2014b; Schunck et al., 2004; Varga et al., 2004]. Even with manual scoring of CBMN past 4 Gy, there is a decrease in the slope of the dose-response curve away from the expected linear-quadratic shape; however, the decrease is smaller than for automated analysis [Müller and Rode, 2002; Beinke et al., 2016]. Other studies have compared the use of manual and automated scoring using microscope-based analysis on slides showing that automated analysis results in lower MN/BNC at high doses and a higher false-positive rate [Rossnerova et al., 2011; Selvan et al., 2015]. With automated scoring, this decrease in MN/BNC at higher doses results in a dose-response curve with an inverse parabolic shape. This means that any 1 MN/BNC rate could be translated into two doses, one below and one above the peak dose value.

To mitigate this effect, a multi-parametric approach was examined to consider other parameters that are measured with IFC such as mono-, bi-, tri-, and quadra-nuclear cells. These parameters are also affected by dose, most notably the monotonic decrease in the number of BNC. While these parameters alone would not provide accurate dose estimates, they do let the data be classified as high or low dose. This allows the appropriate side of the dose-response parabola to be selected. From the modeling of the dose-response curve data, the maximum point of the parabola was determined to occur at 5.3 Gy and was used as the classification dose between low and high doses in the LDF. To continue the comparison between the sample acquisition methods, classification as a high or low dose was performed using all three models (manual, auto, and combined) and it was determined that they all performed similarly for this initial step of dose estimation with about 87% correct classification of dose range. Those misclassified tended to be close to the cut-off dose, as would be expected, but also at the highest doses (online suppl. Fig. S4). This could be due to deterioration in the health and morphology of the cells that would greatly affect image analysis at the high doses. This is further supported by the fact that the LDF was most effective at classifying the doses below 5 Gy.

Once the dose range of each sample was identified, the appropriate calibration curve was applied to estimate the dose. When comparing between the sample acquisition methods, again there was little difference in the dose estimates from each method confirming that acquisition by autoplate is equivalent to manual acquisition and the two methods are interchangeable.

Overall, dose estimates were more accurate in the low-dose region (≤4 Gy) and tended to be underestimated at higher doses. This can be explained by a higher number of samples that were initially misclassified with the LDF (online suppl. Fig. S5). If a high-dose sample is misclassified, it will be assigned a low-dose value. This will result in the average dose estimate over all samples being lower than the true dose. In online supplementary Figure S5, it can be seen that the correctly classified samples in the high-dose region were much closer to the true dose than the misclassified samples. Fortunately, only a small percentage of samples were misclassified at the higher doses, with most occurring in the midrange where misclassification has a smaller effect on the final dose estimate. Another issue around the peak of the MN/BNC curve is that there is a region between about 4 and 6 Gy where the MN/BNC changes little with dose. This can lead to inaccuracies in this region. In the dose range above 7 Gy, the dose estimates appear to plateau, likely due to the inability of the highly damaged cells to progress into cytokinesis. This same trend was demonstrated with automated scoring of microscope slides [Capaccio et al., 2021]. It was also shown that there is greater variability in the high-dose samples, even within the correctly classified samples, which could be due to a greater difficulty analyzing the images of highly damaged cells that could be changing in morphology. Kacprzak also observed a large interindividual variation that is more pronounced at high doses [Kacprzak et al., 2013]. These results highlight that, when extending the dose range for the CBMN assay, the initial classification of the sample as a high- or low-dose exposure is paramount. It should also be noted that in the high-dose range, above 7 Gy, the exact dose is less important than identifying that these individuals have been irradiated with lethal doses. This is the information that will guide the medical intervention of these individuals.

We have demonstrated that the method of sample acquisition used had no measurable effect on the calibration curve or the dose estimates on unknown samples. This demonstrates that the IFC-CBMN method can be used in high-throughput, hands-free mode to determine the dose of potentially exposed individuals. The combination of a 96-well plate acquisition and automated data analysis through preestablished templates and batch analysis makes this method desirable for large-scale population screening.

The novel multi-parametric analysis is an accurate way to make dose estimates in the low-dose region, below 5.3 Gy. In the high-dose region, above 5.3 Gy, the dose estimates were less accurate but generally sufficient to identify the samples as highly irradiated. For a mass casualty situation, this novel method of analysis would be sufficient to identify those who require intensive medical intervention. It could also be used as a quick triage tool to identify those exposed to low doses with more accurate biodosimetry being performed later. Future studies will include interlaboratory comparisons to ensure that this methodology is easily transferable.

This study protocol was reviewed and approved by Research Ethics Board, Health Canada, approval number REB 2002-0012H. Written informed consent was obtained from all participants to participate in the study.

The authors have no conflicts of interest to declare.

This work was supported by a pilot grant from the Opportunity Funds Management Core of the Centers for Medical Countermeasures against Radiation Consortium (CMCRC), National Institute of Allergy and Infectious Diseases, National Institutes of Health (Grant No. U19AI067773).

Lindsay A. Beaton-Green conceived and planned the experiments, supervised the project, and contributed to the design and implementation of the research, the analysis of the results, and the writing of the manuscript. Jessica M. Mayenburg and Eman M. Hassan contributed to the analysis of the results and the writing of the manuscript. Leonora Marro conducted the statistical analysis and contributed to the writing of the manuscript. Sarita Cuadros Sanchez contributed to the design of the research, performance of the experiments, analysis of the results, and writing of the manuscript. Riham Darwish and Sylvie Lachapelle contributed to the performance of the experiments. Nadine Adam contributed to the performance of the experiments and writing of the manuscript. Julie J. Burtt contributed to the supervision of the project, the design and implementation of the research, performance of the experiments, the analysis of the results, and the writing of the manuscript. Cyndi Van Den Hanenberg contributed to the writing of the manuscript. Matthew A. Rodrigues provided advice and guidance to the analysis of the data and contributed to the writing of the manuscript. Qi Wang and Helen C. Turner contributed to the development of the methods and the writing of the manuscript. David J. Brenner contributed to the development of the collaboration. Ruth C. Wilkins contributed to the supervision of the project, conception and planning of the experiments, analysis of the data, and writing of the manuscript.

All data generated or analyzed during this study are included in this article and its supplementary material files. Further inquiries can be directed to the corresponding author.