Skip to main content
NCBI home page
Scientific Reports logoLink to Scientific Reports
. 2024 Oct 12;14:23872. doi: 10.1038/s41598-024-74953-w

Validation of a blood biomarker panel for machine learning-based radiation biodosimetry in juvenile and adult C57BL/6 mice

Leah Nemzow 1,✉, Michelle A Phillippi 1, Karthik Kanagaraj 1, Igor Shuryak 1, Maria Taveras 1, Xuefeng Wu 1, Helen C Turner 1
PMCID: PMC11470949  PMID: 39396080

Abstract

Following a large-scale radiological event, timely collection of samples from all potentially exposed individuals may be precluded, and high-throughput bioassays capable of rapid and individualized dose assessment several days post-exposure will be essential for population triage and efficient implementation of medical treatment. The objective of this work was to validate the performance of a biomarker panel of radiosensitive intracellular leukocyte proteins (ACTN1, DDB2, and FDXR) and blood cell counts (CD19+ B-cells and CD3+ T-cells) for retrospective classification of exposure and dose estimation up to 7 days post-exposure in an in-vivo C57BL/6 mouse model. Juvenile and adult C57BL/6 mice of both sexes were total body irradiated with 0, 1, 2, 3, or 4 Gy, peripheral blood was collected 1, 4, and 7-days post-exposure, and individual blood biomarkers were quantified by imaging flow cytometry. An ensemble machine learning platform was used to identify the strongest predictor variables and combine them for biodosimetry outputs. This approach generated successful exposure classification (ROC AUC = 0.94, 95% CI: 0.90–0.97) and quantitative dose reconstruction (R2 = 0.79, RMSE = 0.68 Gy, MAE = 0.53 Gy), supporting the potential utility of the proposed biomarker assay for determining exposure and received dose in an individual.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-74953-w.

Subject terms: High-throughput screening, Predictive markers

Introduction

In the event of a radiological/nuclear incident that leads to widespread exposure of ionizing radiation, rapid yet accurate identification of exposed individuals will be necessary to swiftly triage and quantify radiation exposure to guide medical treatment, at a time when emergency resources could be scarce. Biomarkers which measure radiation-induced biological effects in individuals can serve as useful diagnostic tools for biodosimetry and guiding patient-specific medical treatment decisions1–4. The combination of biomarkers in an integrated panel can be especially useful for the development of a bioassay capable of detecting radiation exposure across a range of conditions (such as dose ranges, post-exposure time points, and demographics) with improved accuracy5–9, as compared to the performance of individual biomarkers alone. The practical time considerations for various stages of the emergency response, including deployment of emergency response teams (“boots on the ground”), organization of potentially exposed individuals for the collection of samples, and assay time-to-result, underscore the critical need for a same-day result, high-throughput bioassay that detects radiation exposure in the general population up to a week later. Towards this goal, we previously identified a panel of candidate human intracellular (IC) leukocyte protein biomarkers10, and developed an imaging flow cytometry (IFC)-based platform for the rapid and high-throughput quantification of radiation-induced upregulated protein biomarker expression in peripheral blood samples11,12. More recently, we have shown that the combination of lymphocyte blood cell (BC) counts together with IC protein biomarkers in one panel produced strong biodosimetry results in the C57BL/6 model8.

The objective of this work was to evaluate the in vivo performance of our biomarker panel (IC proteins: ACTN1, DDB2, and FDXR, and BC counts: surface-stained CD19+ B-cells and CD3+ T-cells) in juvenile and adult C57BL/6 mice up to 7 days post-exposure for radiation biodosimetry applications. These IC proteins are known to play roles in DNA Damage Response (DDR) mechanisms. Alpha-actinin 1 (ACTN1) is a cross-linking cytoskeletal protein that has been associated with stress-induced cellular senescence13. Damage-specific DNA-binding Protein 2 (DDB2) is a DNA lesion recognizing subunit for facilitation of nucleotide excision repair, which has also been implicated in mediation of apoptosis and premature senescence14. Ferredoxin Reductase (FDXR) is a mitochondrial flavoprotein essential for electron transport that modulates p53-dependent apoptosis via the production of reactive oxygen species in the mitochondria15,16. Peripheral leukocyte counts are well-known biomarkers of radiation exposure; mature lymphocytes (B-cells and T-cells) in particular are considered among the most radiosensitive leukocyte cell type17 and models of their depletion kinetics may be used for estimating dose18–22.

In this study, juvenile and adult (aged 4 and 12–13 weeks, respectively) C57BL/6 mice were sham or X-irradiated with 1, 2, 3, or 4 Gy total body doses, and peripheral blood was collected 1, 4, and 7 days post-exposure. Levels of ACTN1, DDB2, and FDXR protein expression, and percentages of B-cells (CD19+) and T-cells (CD3+) in the mouse peripheral blood samples were quantified by IFC. Biomarker data were input into an ensemble machine learning (ML) platform for prediction of (1) exposure status and (2) received dose. ML algorithms which integrate many predictor variables and types of data inputs are advantageous with the use of biomarker panels for detection of radiation exposure in a range of doses, time points, demographics, and exposure levels. We employed the Boruta feature selection algorithm to first identify the most influential predictor variables, and subsequently aggregated the predictive power of these variables through a stacking ensemble that incorporates multiple ML methods. Stacking is a powerful ML technique because the use of various models can complement each other on difficult to predict samples, and the ensemble of models can often outperform any individual model23–26. To our knowledge, our group is the first to implement the stacking methodology to perform radiation biodosimetry7.

Materials and methods

Animal model

These mouse studies were approved by the Columbia Institutional Animal Care and Use Committee (IACUC approved protocol #AABS1600) and were conducted in facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC). Male and female juvenile (aged 3 weeks) and adult (aged 11–12 weeks) C57BL/6 mice were purchased from Charles River Laboratories (Frederick, MD). Following at least 7 days of acclimation, mice were randomly assigned to the sham irradiation group (0 Gy) or irradiated experimental groups (1, 2, 3, and 4 Gy). This study included 251 mice in total, with 15–18 mice at each dose (approximately equal male and female, and adult and juvenile). All methods were performed in accordance with ARRIVE guidelines (https://arriveguidelines.org) and with other relevant state and federal guidelines.

Irradiation and blood sample collection

Mouse irradiation was performed according to our earlier studies11. Briefly, mice were placed in a commercial mouse irradiation holder (Precision X-ray) inside the X-RAD 320 biological irradiator (Precision X-ray Inc., North Branford, CT), and either sham irradiated (for 5 min) or exposed to 1, 2, 3, or 4 Gy X-rays using a 1.5 mm Al + 0.25 mm Cu + 0.75 mm Sn filter, operating at 320 kVp, 12.5 mA (dose rate = 1 Gy/min). The delivered dose was measured using a Radcal® ion chamber (Monrovia, CA; calibrated annually by Radcal) and the probe was placed in the irradiation holder along with the mice to monitor the dose rate. Mouse weights were recorded before and after irradiation. All mice were euthanized by CO2 asphyxiation at 1, 4 and 7 days after irradiation prior to blood collection. Peripheral whole blood samples (approximately 0.4–0.8 ml) were collected by cardiac puncture using a heparinized 1 ml syringe (BD Precisionglide™; Becton-Dickson, Franklin Lakes, NJ) and transferred into 1.5 ml Eppendorf tube. Across the doses, weight loss did not exceed 5% in either age cohort up to 7 days post-irradiation.

Imaging flow cytometry

Samples were fixed, stained, and acquired by imaging flow cytometry, and analyzed as previously described8. Briefly, 100 µl peripheral whole blood samples in triplicate from each mouse were treated with RBC lysis buffer (eBioscience, #00–4333-57), surface stained with anti-mouse CD19 PE/Dazzle (1:800; Biolegend, #115,554) B-cell marker, and anti-mouse CD3 PE (1:800; eBioscience, #12–0031–82) T-cell marker, and then fixed and permeabilized (BD Biosciences; Cytofix/Cytoperm; #554,714). Each of the triplicate samples were then stained intracellularly with one of the following antibodies: FDXR (1:100; Sigma, #HPA044393), ACTN1 (1:100; Cell Signaling Technology, #3134s), or DDB2-FITC (1:100; Cusabio, #CSB-PA846067LC01HU). Unconjugated primary intracellular antibodies FDXR and ACTN1 were then stained with goat anti-rabbit Alexa Fluor 488 (AF488) secondary antibody (1:1000; Life Technologies, #A11034). Single, focused cells (3,000) were acquired on the ImageStreamx MkII imaging flow cytometer (Luminex, Austin, TX) at 40 × magnification, with the 488 nm excitation laser at 200 mW. To compensate for spectral spillover, cells stained with single fluorescence only were acquired using the 488 nm laser with the brightfield and side scatter inactivated, and the compensation coefficients were determined automatically by the compensation wizard within the IDEAS® software (Luminex ver. 6.2). All captured images were analyzed with the application of a uniform template and each sample file was batch processed using the IDEAS® software. Single, focused, and healthy cells were gated as previously described8. B-cell and T-cell populations were identified and quantified by gating for CD19+ and CD3+, respectively. The Mean Fluorescence Intensity (MFI) of each IC protein biomarker was quantified in “all leukocytes” (focused, single, non-apoptotic leukocytes); B-cell (focused, single, non-apoptotic CD19+), and T-cell (focused, single, non-apoptotic, CD3+) populations. BC count biomarkers were quantified as relative percentages of B-cell and T-cell population present in the acquired samples (“% Gated B-cells” and “% Gated T-cells”, respectively) by calculating [# of CD19+ or CD3+ cells / total single, focused, healthy cells]. Unstained and AF488-stained 0 Gy samples were prepared and acquired in each experimental batch for use as technical controls to account for any variability introduced by changes in antibody lots or sample handling. In this study, MFI values of the unstained/AF488-stained controls were normalized to common values, and biomarker MFI values corresponding with each control were normalized accordingly.

Statistical analyses

IC protein MFI values and BC B-cell / T-cell percentage values were natural log (ln) transformed to bring their distributions closer to the normal distribution. The mean differences of both surface-stained B-cell and T-cell percentages, and ACTN1, DDB2, and FDXR expression in all leukocytes, B-cell, and T-cell populations on Days 1, 4, and 7 post 0–4 Gy X-irradiation were analyzed by two-way ANOVA with Tukey’s multiple comparison test. The baseline differences in surface-stained B-cell and T-cell percentages, in unirradiated adults and juvenile cohorts were analyzed with unpaired Student’s T tests). Two-tailed p values less than 0.05 were considered statistically significant. Simple linear regression was used to test predictive association of X-ray dose with ln-transformed MFI biomarker expression and B-cell and T-cell percentages. ANCOVA was performed to compare the slopes and intercepts of the fitted regressions for biomarker expression in adult and juvenile cohorts. All statistical analyses were performed using GraphPad Prism (version 10.1.2; GraphPad Software, Inc., La Jolla, CA).

Machine learning analyses

The data from all experiments were compiled into one table (in the R 4.3.2 programming language27). There were 251 samples which had non-missing values for at least one IC protein biomarker (ACTN1, DDB2, FDXR). Missing values for the biomarkers were imputed using the MissForest algorithm in R. The main variables in the resulting data set were:

  • Time after irradiation (Day 1, 4, or 7).

  • Sex (female = 0, male = 1).

  • Age (juvenile = 0, adult = 1).

  • Individual ln-transformed IC protein biomarker signal intensities in healthy “all leukocytes”, B-cell and T-cell populations:

    • ln-transformed ACTN1/DDB2/FDXR in all leukocytes,
    • ln-transformed ACTN1/DDB2/FDXR in B-cells,
    • ln-transformed ACTN1/DDB2/FDXR in T-cells,

  • ln-transformed percentages of surface-stained B-cells and T-cells (% gated B-cells and % gated T-cells, respectively; mean percentage values from replicate samples generated from each mouse were used).

The complete data set can be found in Supplementary Table S1. Exposure Index (a categorical variable; 0 = 0 Gy, 1 = all doses > 0 Gy) and Dose (a continuous variable) were treated as the target variables to be predicted by the ML models, using the other variables as predictors. The ML analyses for predicting these variables were performed in Python 3.10.4, Jupyter notebooks (https://jupyter.org/). Data were split randomly (50% each) into 2 parts: (1) a training set that was input into the ML platform for iterative feature selection and subsequent model stacking using multiple ML algorithms, and (2) a testing set that was initially reserved (hidden from the training models). Thus, all feature selection and model stacking computations were performed on the training set with repeated cross-validation (10-fold, repeated 10 times) without access to the testing set, and the performance of the final stacked model ensemble was evaluated on the testing set.

Feature Selection: Using the training dataset, the Boruta algorithm (implemented by the BorutaPy Python package28) was used to select the most important predictors by creating “shadow features” (randomized copies of original features) and comparing their importances to the original features using Random Forest (RF). A “hit” was assigned to an original feature if its importance (measured using a Z-score) was greater than the Maximum Shadow Z-score Attribute (MSZA). If the Z-score of an original feature was less than or equal to the MSZA, a two-sided test against the MSZA was performed. If the Z-score of the original feature was significantly less than the MSZA, the feature was dropped as unimportant. Original features that significantly outperformed the top score (≥ 100th percentile) of the randomized “shadow” features, using α = 0.05 with Bonferroni correction as the statistical threshold, were retained as important. The Boruta algorithm iterated through this process until all features were either confirmed as important or unimportant.

Model Stacking: The retained predictors (which passed the Boruta screening for importance) were used to train stacked ensemble ML algorithms to classify samples by Exposure Index (0 = 0 Gy, 1 = all doses > 0) or quantitatively reconstruct Dose as a continuous variable. In stacking, several ML methods (level 0 models) were applied to the training data with repeated k-fold cross validation (10-fold, repeated 10 times) generating many predictions to evaluate the performance of each machine learning model, increase robustness of the stacked model, and reduce overfitting. For predictions of Exposure Index, the stacking approach was used to integrate the prediction outputs of the following level 0 models: logistic regression, CatBoost, XGBoost, Light GBM, random forest (RF), K Nearest neighbors, Naïve Bayes, and Support Vector Machines. In this case, best results were found using RF as the level 1 meta-model, based on comparing the mean performance scores (balanced accuracy metric) for all models on the repeated cross validations. Stacking was also used for regression to predict Dose, where the level 0 models used were: Linear Regression, CatBoost, Light GBM, Linear Boost, RF, Elastic Net, and Support Vector Machines regression. The best level 1 model for Dose predictions was also RF, based on comparing mean Root Mean Squared Error (RMSE) scores (square root of the average squared differences between predicted and actual values) for all models on the repeated cross validations.

Stacked ensemble model Level 1 Predictions of Exposure Index or Dose on testing data: Once the level 1 meta-model was trained, it was used to make predictions on the previously reserved testing data in 2 steps: (1) First, each level 0 base model made predictions on the testing data which served as intermediate inputs for the next step and were collected and combined into a new dataset. (2) Next, the level 1 meta-model took these aggregated predictions as inputs to generate the final predictions of the target variables (Exposure Index or Dose). As these testing data had not been seen by the training algorithms, they mimic a real-world scenario where samples from potentially exposed individuals will be presented to the stacked ensemble for predicting dose in a categorical or continuous manner.

Results

IC protein biomarker dose-responses

Adult and juvenile mice were X-irradiated 0–4 Gy and MFI values of IC protein biomarkers ACTN1, DDB2, FDXR were quantified by IFC in all leukocytes, B-cell, and T-cell populations on Days 1, 4, and 7 post-exposure. Figure 1 shows the ln-transformed ACTN1, DDB2, and FDXR MFI values in each leukocyte population at each Dose on each Day post-exposure across all adult and juvenile mice. Two-way ANOVA analyses were performed to analyze the effects of X-ray Dose and time post-irradiation (Day) on IC protein biomarker MFI values measured within all leukocytes, B-cell, and T-cell populations. Dose and Day were both found to significantly affect mean expression of all biomarkers in all cell populations (p < 0.0001, and 0.05–0.0001, respectively), with the one exception of Day affecting ACTN1 expression in the B-cell population. A significant interaction between Dose and Day were seen in most conditions (see Supplementary Table S2 for F and p values), indicating that Day affects the mean biomarker expression measured across the Doses. The exceptions were DDB2 in the all leukocytes population, and FDXR in all leukocytes and T-cell populations, where the effect of Dose on mean biomarker expression is not affected by the Day. The data were also examined by simple linear regression to test the predictive association of Dose on mean biomarker expression on each Day (Supplementary Figure S1 ), and results show positive slopes that significantly differ from 0 across all the biomarkers and populations (p < 0.01–0.0001), on all the Days (with the exception for ACTN1 in T-cells on Day 1, where the mean at 4 Gy is not statistically different from the mean at 0 Gy, and the response in non-linear). Taken together, these results importantly show that dose-dependent upregulation of all three IC protein biomarkers is persistent up to 7 days post-exposure in all the adult and juvenile mice together, although kinetics of each biomarker are not identical and mean biomarker levels at each Dose do differ by Day (in most conditions tested).

Fig. 1.

Fig. 1

Open in a new tab

Intracellular protein biomarker panel MFI in peripheral blood samples 1, 4, and 7, days post 0–4 Gy X-irradiation. Data represent ACTN1, DDB2, and FDXR MFI values in all adult and juvenile mice. Data were natural log (ln) transformed (Y = ln(Y)). Bars represent the ln-transformed mean MFI of each group. Two-way ANOVA were performed to analyze the effect of Day and Dose on biomarker levels in all leukocytes, B-cell and T-cell populations; complete results can be found in Supplementary Table S2. n = 15–18 at each Dose per Day.

Simple linear regression analyses with ANCOVA were performed to test for differences in ACTN1, DDB2, and FDXR dose-responses between the adult and juvenile cohorts in all leukocytes, B-cell, and T-cell populations on Days 1, 4, and 7 (Supplementary Figures S2-S4). Overall, no significant differences in regression slopes are seen between the adult and juvenile cohorts, apart from ACTN1 in all leukocytes on Day 4, and DDB2 in all leukocytes and B-cells on Day 4. These different slopes can be explained by the expression of biomarkers in juveniles that continues to increase up to 4 Gy, while expression in adults begins to decrease after 3 Gy. We also observed that Y-intercepts of biomarker levels are statistically different in several conditions, with juvenile cohorts showing generally lower values. These data indicate that while different regressions for adults and juvenile cohorts are generally required to fit the data (due to the varied Y-intercepts), biomarker dose-responses do not generally vary between age cohorts. Whether these different regression intercepts affect biodosimetry will depend upon the mathematical model used to classify exposure or reconstruct dose.

Radiation Dose-dependent depletion kinetics of B-cells and T-cells

As a versatile instrument, the IFC quantified IC biomarkers in specific leukocyte sub-types within the mixed blood cell population (Fig. 1), and in the same samples also quantified populations of surface-stained leukocytes. Figure 2 shows the ln-transformed percentages of surface stained B-cells (CD19+) and T-cells (CD3+) present in the total combined adult and juvenile mice (using the mean percentage value from the triplicate samples generated from each mouse) 1, 4, and 7 days after 0–4 Gy X-irradiation.

Fig. 2.

Fig. 2

Open in a new tab

Relative percentages of surface-stained B-cells and T-cells in C57BL/6 mouse peripheral blood samples 1, 4, and 7, days post 0–4 Gy X-irradiation. Data were natural log (ln) transformed (Y = ln(Y)). Bars represent the ln-transformed mean percentage of each group. Two-way ANOVA were performed to analyze the effect of Day and Dose on B-cell and T-cell percentages; complete results can be found in Table S3. n = 15–18 at each Dose per Day.

General observations of the ln-transformed percentages of B-cells show depletion of B-cells up to 4 Gy on Days 1, 4, and 7. A two-way ANOVA was performed to analyze the effect of Dose and Day on B-cell percentages (results shown in Supplementary Table S3 ). Results show that Dose does have a statistically significant effect (p < 0.0001) on B-cell percentages, and Day does not (p = 0.06). However, the results show a significant interaction between the effects of Dose and Day (F (8,235) = 2.8, p < 0.01), indicating that while B-cell depletion is dose-dependent on Days 1, 4, and 7, the percentage of B-cells at each Dose is affected by which Day the data is measured on. These results suggest that Dose-dependent depletion of B-cells will likely serve as a strong predictor for dose exposure at early time-points, as well as later time points up to Day 7. In the case of the T-cell percentage levels, general observations only show moderate dose-dependent depletion beginning at 3 Gy on Day 1. Two-way ANOVA results show that while Dose does not significantly affect mean T-cell percentages (p = 0.08), Day and interaction of Day and Dose were significant variables (p < 0.0001). Although these results indicate that Dose does not generally affect T-cell percentages, a Dose effect is present on certain Days. Tukey’s multiple comparisons tests were performed to further investigate T-cell Dose response on each Day, and results show significant differences in mean percentages of T-cells between most Doses on Day 1 (see Supplementary Table S3 for p values), with no significant differences between most Doses on Days 4 and 7. These data indicate that Dose-dependent depletion of T-cell percentages does not persist up to Days 4 or 7, suggesting that T-cells will likely not serve as a strong predictor for dose exposure up to 7 days.

Baseline levels of B-cell and T-cell percentages in unirradiated mice (from all the days) were evaluated, and results show no significant differences between the adult and juvenile cohorts (unpaired T test, Supplementary Figure S5a). Simple linear regression analyses with ANCOVA were then performed to test for differences in hematological dose-responses (B-cell and T-cell percentage levels) between the adult and juvenile cohorts on Days 1, 4, and 7 (Supplementary Figure S5b). B-cell percentages in adult and juvenile cohorts on Days 1 and 4 show no differences in slopes (although Y-intercepts are significantly higher in juveniles), and significantly different slopes are seen on Day 7 (with a smaller slope in the juvenile data). These data indicate that different regressions for adults and juvenile cohorts are generally required to fit the data, and biomarker dose-responses do not generally vary between age cohorts until Day 7. Slopes and Y-intercepts are not significantly different between adult and juvenile cohorts in T-cells on any Day.

Machine learning-based biodosimetry: exposure classification and dose reconstructions

This study examines the ln-transformed expression of three individual IC protein biomarkers (ACTN1, DDB2, FDXR) in three cell populations (all leukocytes, B-cells, T-cells), as well as ln-transformed percentages of B-cells and T- cells, at 5 doses (0, 1, 2, 3, 4 Gy) on 3 days post-exposure (1, 4, 7) in adult and juvenile cohorts comprised of males and females. Each of these variables in this expansive study serves as a potential predictor for radiation dose classification (Exposure Index) or quantitative reconstruction (Dose), as detailed fully in the Materials and Methods. To identify the strongest predictor variables and combine them for exposure classification and dose reconstruction, we used an ensemble ML method, which includes Boruta feature selection and integrated stacked algorithms.

Boruta feature selection, which duplicates the data into a shuffled dataset (shadow features) and then tests whether the predictors outperform their shadows in the context of an ML algorithm which produces variable importance measures (VIMs), such as RF, was used to determine which predictor variables are the most important for predicting Exposure Index or Dose and should be retained for each ML analysis endpoint. Boruta testing did not retain Sex, Age, Day, ACTN1 in B-cells and T-cells, DDB2 in T-cells, and % Gated T-cells variables for predicting Exposure Index (exposure classification) (Fig. 3a), and the resulting ensemble algorithm with the retained predictors successfully discriminated between exposed and unexposed samples with classification accuracy of 87% and ROC curve AUC = 0.94 (95% CI 0.90–0.97) (Fig. 3b, complete table of predicted values on the testing dataset can be found in Supplementary Table S4 ). For predicting Dose (dose reconstruction), Boruta testing only retained the % Gated B-cells, ACTN1 in B-cells, DDB2 in B-cells, and FDXR in B-cells and T-cells variables (Fig. 4a). The stacking ensemble algorithm with these retained predictors achieved strong results for quantitative dose reconstruction (regression task): a comparison of true dose values with reconstructed values on testing data produced coefficient of determination R2 = 0.79, root mean squared error (RMSE) = 0.68 Gy, mean absolute error (MAE) = 0.53 Gy, and comparisons of mean reconstructed dose with actual dose remained within 0.5 Gy for all doses tested (Fig. 4b, complete table of predicted values on the testing dataset can be found in Supplementary Table S4).

Fig. 3.

Fig. 3

Open in a new tab

Machine learning predictions of Exposure Index. (a) Boruta feature selection results on training data. (b) Discrimination between unexposed (0 Gy) and exposed (1–4 Gy) on testing data by the stacking ensemble (several level 0 models providing input for the RF level 1 model). Full testing dataset and Exposure Index predictions can be found in Supplementary Table S4.

Fig. 4.

Fig. 4

Open in a new tab

Machine Learning predictions of Dose. (a) Boruta feature selection results on training data. (b) Actual versus reconstructed Dose on testing data by the stacking ensemble (several level 0 models providing input for the RF level 1 model). Full testing dataset and Dose predictions can be found in Supplementary Table S4.

Discussion

In a widespread radiological/nuclear emergency, bioassays for rapid and accurate mass-screening of individuals for radiation exposure will be a critical component of an efficient emergency response. Large-scale emergencies may preclude timely access to blood samples from all potentially exposed individuals. Therefore, the identification and validation of radiation biomarkers that persist for several days after radiation exposure are necessary for the development of bioassays that remain clinically relevant throughout the duration of the triage stage of the emergency response. Additionally, radiological/nuclear emergencies will likely include understudied or vulnerable populations such as juveniles, seniors, or those with pre-existing medical conditions, and it is important to develop inclusive bioassays that can effectively inform triage and medical treatment decisions across demographics. Previously, we developed a same-day-result, high-throughput IFC-based multiparametric blood bioassay for quantifying radioresponsive IC protein (ACTN1, DDB2, FDXR) and BC count (% Gated B-cells and T-cells) biomarkers 24 h after exposure in adult C57BL/6 mice8, though no validation of this bioassay for extended time points or demographics had yet been performed. Thus, the work presented here tests the in vivo performance of the IC protein and BC count biomarker panel across 7 days post-exposure in adult and juvenile C57BL/6 mice to validate the utility of the bioassay across these conditions. Each of the blood bioassay components tested across the days, doses, and demographics were input as variables into an ensemble ML platform to identify the strongest predictor variables and combine them for predictions of exposure classification and absorbed dose.

Evaluation of individual biomarker responses across the tested temporal and demographic conditions (Day, Age, Sex) together with consideration of which of these variables have been identified as strong predictors for biodosimetry, can be used to inform the conditional range of application and limitations of the bioassay:

Kinetics and temporality: Quantifications of IC protein biomarker expression levels and B-cell and T-cell percentages across all the ages and sexes overall show Dose-dependent biomarker responses in the 0–4 Gy range that persist up to Day 7 (except for T-cell percentages which only show Dose-dependent response at Day 1). Although biomarker responses at each Dose generally differed by Day (Figs. 1 and 2), the Boruta feature selection component of the ML platform rejected the Day variable as a predictor for both exposure classification and dose reconstruction (Figs. 3 and 4), highlighting the temporal versatility of this bioassay for triage use up to a week after radiation exposure. However, persistence of biomarker expression across the time points may vary in higher dose ranges, and therefore, the inclusion of extended doses in future studies will further investigate the kinetic and temporal limits of this bioassay.

Demographics: Biomarker performance in the in-vivo C57BL/6 mouse model was tested up to Day 7 post-exposure in male and female adult and juvenile cohorts. We first evaluated whether biomarker dose-responses differed between the age cohorts up to Day 7 post-exposure. Y-intercepts of biomarker expression are generally lower in the juvenile cohorts (especially in the case of FDXR), and biomarker dose-responses are generally equivalent between the two age cohorts (except in select cases of ACTN1 and DDB2 on Day 4 where expression in adults decreases after 3 Gy but continues to increase in the juvenile data) (Supplementary Figures S2-S4). While unirradiated B-cell percentage levels at baseline do not differ between the age cohorts (Supplementary Figure S5a), B-cell percentage levels in the irradiated mice are overall higher in the juvenile mouse cohort as compared to the adult (Supplementary Figure S5b) and show less B-cell depletion at Day 7. These data suggest that in juveniles, B-cell lymphocytes may be less sensitive to radiation exposure or there is increased proliferation and replenishment compared to the adults. Measuring individual biomarker performances in different age cohorts across the Doses and Days is important for evaluating radiobiological responses across demographics, yet differences detected in this capacity do not necessarily affect biodosimetry outcomes. Boruta feature selection did not retain Age or Sex as an important predictor for either exposure classification of dose reconstruction (Figs. 3 and 4). The results at this preliminary stage, therefore, support the use of this biomarker panel for biodosimetry across age (adult and juvenile) and sex demographics towards a universal bioassay in a radiological emergency. However, continued studies with increased sample sizes, age and dose ranges will be needed to validate these findings beyond the currently tested conditions. Future studies that include additional vulnerable demographics (such as seniors or those with pre-existing medical condition radiation treatments) will be important for comprehensive identification of any potential confounding variables.

The evaluation of the blood biomarker components that were used as input predictor variables and retained by Boruta feature selection for successful exposure classification (Fig. 3) or dose reconstruction (Fig. 4), can inform which blood biomarker components are important to include in the final bioassay design. IC proteins in the “all leukocytes” population passed Boruta testing for exposure classification but did not pass testing for dose reconstruction, suggesting that the more demanding dose reconstruction task requires more stringent selection of predictor variables. IC protein biomarker quantifications in B-cells and B-cell percentages were important predictors retained for both exposure classification and dose reconstruction, whereas biomarker quantifications in T-cells and T-cell percentages were generally discarded from the prediction algorithms. In the context of the in-vivo C57BL/6 mouse model studied here, these data signify B-cells as a driving cell type for radiation biodosimetry, and consideration of T-cells for biomarker quantification can likely be ignored. Yet, it is essential to note the biomarker panel is validated here in C57BL/6 mice as a preclinical model, and the results observed in this study may be model-specific with a need for continued validations of blood biomarker performance in non-rodent models. Towards translation into the human model, the IC protein antibodies used in this study are anti-human with good cross-reactivity in mouse cells, showing conserved homology of protein structure between these species.

In this study, we tested the in-vivo performance of an IC protein and BC count biomarker panel which effectively combine the strengths and capabilities of each biomarker type. Lymphocytes are known to be highly sensitive to radiation and lymphocyte counts are already an established biomarker for radiation biodosimetry1,29, yet alone, their usefulness for guiding medical countermeasures in a large-scale radiological emergency is limited. Lymphocyte depletion kinetic assays require early establishment of baseline blood counts, which vary greatly within the healthy population and may be further confounded by other conditions such as infection, trauma, burns, and cancer30–34. Use of biomarker panels has been shown to reduce baseline variations of single biomarkers35,36, and the integration of hematological and proteomic biomarkers together as one panel in this study adequately accounted for inter-mouse variability, generating successful exposure classifications (Fig. 3) and dose reconstructions (Fig. 4). The inclusion of IC protein biomarkers together with BC counts in this biodosimetry platform importantly also removes the need for serial blood collection, as a blood sample at one post-exposure time point (up to seven days) is sufficient for accurate radiation biodosimetry. Additionally, lymphopenia is a main feature of acute radiation syndrome (ARS)37,38, and the inclusion of BC counts in this all-in-one bioassay platform also importantly helps correlate this radiation biomarker panel with clinical outcomes.

In summary, we have developed a multiparametric bioassay that uses a radiosensitive biomarker panel (comprising IC proteins relating to DDR together with lymphocyte quantifications) and ensemble ML methods to successfully classify radiation exposure and estimate dose. The results in this work showed dose-dependent biomarker changes in the 0–4 Gy range persisted up to Day 7 (with the exception of T-cell percentages). Distinctions between age (adults/juvenile), sex (male/female) or time (Day 1, 4, 7) did not affect the ML prediction of the target variables, suggesting a universal bioassay that can perform biodosimetry across days and demographics. Importantly, the IFC + ML biodosimetry platform presented in this work is adaptable for validation of these biomarkers in other species and conditions. Future work with increased samples sizes, along with expansion of the tested parameters, will continue to validate the findings presented here and determine the full range of utility towards the development of a comprehensive bioassay for rapid triage of the general population in a radiological emergency.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Table S1 (99.6KB, xlsx)
Supplementary Table S2 (14.8KB, xlsx)
Supplementary Table S3 (19.5KB, xlsx)
Supplementary Table S4 (31.3KB, xlsx)
Supplementary Figures S1-5 (516.9KB, pdf)

Author contributions

H.C.T. conceived the overall study. M.A.P., K.K., and H.C.T. designed the experimental study plan. M.A.P., K.K., and X.W. performed mouse handling, irradiation, and blood draws. M.A.P., K.K., M.T., and L.N. performed sample preparation and IFC acquisition. L.N. and M.A.P. performed biomarker analyses and statistical analyses. I.S. performed ML analyses. L.N., K.K., I.S., and H.C.T., prepared the manuscript. All authors contributed to reviewing the manuscript.

Funding

This work was supported by the National Institute of Allergy and Infectious Diseases (NIAID) U01 #AI148309 and Administrative Supplements (awarded to H.C.T.).

Data availability

All datasets analyzed during this presented study are available in Supplementary Table S1 online and also available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

The original online version of this Article was revised: The original version of this Article contained an error in Figure 4, where the dose numbers in the axes of the graph were incorrect.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Change history

2/10/2025

A Correction to this paper has been published: 10.1038/s41598-025-87067-8

References

  • 1.Sproull, M. T., Camphausen, K. A. & Koblentz, G. D.  Biodosimetry: A future tool for medical management of radiological emergencies. Health Secur.15, 599–610. 10.1089/hs.2017.0050 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Swartz, H. M., Williams, B. B. & Flood, A. B. Overview of the principles and practice of biodosimetry. Radiat. Environ. Biophys.53, 221–232. 10.1007/s00411-014-0522-0 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Singh, V. K., Newman, V. L., Romaine, P. L., Hauer-Jensen, M. & Pollard, H. B. Use of biomarkers for assessing radiation injury and efficacy of countermeasures. Expert Rev. Mol. Diagn.16, 65–81. 10.1586/14737159.2016.1121102 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Satyamitra, M. M., DiCarlo, A. L., Hollingsworth, B. A., Winters, T. A. & Taliaferro, L. P. Development of biomarkers for Radiation Biodosimetry and Medical Countermeasures Research: current status, Utility, and Regulatory pathways. Radiat. Res.197, 514–532. 10.1667/RADE-21-00157.1 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Li, S. et al. Developing gender-specific gene expression Biodosimetry using a panel of Radiation-responsive genes for determining Radiation Dose in Human Peripheral blood. Radiat. Res.192, 399–409. 10.1667/RR15355.1 (2019). [DOI] [PubMed] [Google Scholar]
  • 6.Shuryak, I. et al. A machine learning method for improving the accuracy of radiation biodosimetry by combining data from the dicentric chromosomes and micronucleus assays. Sci. Rep.12, 21077. 10.1038/s41598-022-25453-2 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Shuryak, I. et al. Biomarker integration for improved biodosimetry of mixed neutron + photon exposures. Sci. Rep.13, 10936. 10.1038/s41598-023-37906-3 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Shuryak, I. et al. Machine learning approach for quantitative biodosimetry of partial-body or total-body radiation exposures by combining radiation-responsive biomarkers. Sci. Rep.13, 949. 10.1038/s41598-023-28130-0 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Blakely, W. F. et al. Multiple parameter radiation injury assessment using a nonhuman primate radiation model-biodosimetry applications. Health Phys.98, 153–159. 10.1097/HP.0b013e3181b0306d (2010). [DOI] [PubMed] [Google Scholar]
  • 10.Lee, Y. et al. Candidate protein markers for radiation biodosimetry in the hematopoietically humanized mouse model. Sci. Rep.8, 13557. 10.1038/s41598-018-31740-8 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Wang, Q. et al. Development of the FAST-DOSE assay system for high-throughput biodosimetry and radiation triage. Sci. Rep.10, 12716. 10.1038/s41598-020-69460-7 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Nemzow, L., Boehringer, T., Bacon, B. & Turner, H. C. Development of a human peripheral blood ex vivo model for rapid protein biomarker detection and applications to radiation biodosimetry. PLoS One. 18, e0289634. 10.1371/journal.pone.0289634 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Dierick, J. F. et al. Identification of 30 protein species involved in replicative senescence and stress-induced premature senescence. FEBS Lett.531, 499–504. 10.1016/s0014-5793(02)03604-9 (2002). [DOI] [PubMed] [Google Scholar]
  • 14.Bao, N., Han, J. & Zhou, H. A protein with broad functions: damage-specific DNA-binding protein 2. Mol. Biol. Rep.49, 12181–12192. 10.1007/s11033-022-07963-4 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Druck, T. et al. Fhit-Fdxr interaction in the mitochondria: modulation of reactive oxygen species generation and apoptosis in cancer cells. Cell. Death Dis.10, 147. 10.1038/s41419-019-1414-7 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Imamichi, Y. et al. Transcriptional regulation of human ferredoxin reductase through an intronic enhancer in steroidogenic cells. Biochim. Biophys. Acta. 1839, 33–42. 10.1016/j.bbagrm.2013.11.005 (2014). [DOI] [PubMed] [Google Scholar]
  • 17.Heylmann, D., Rodel, F., Kindler, T. & Kaina, B. Radiation sensitivity of human and murine peripheral blood lymphocytes, stem and progenitor cells. Biochim. Biophys. Acta. 1846, 121–129. 10.1016/j.bbcan.2014.04.009 (2014). [DOI] [PubMed] [Google Scholar]
  • 18.Goans, R. E. & Waselenko, J. K. Medical management of radiological casualties. Health Phys.89, 505–512. 10.1097/01.hp.0000172144.94491.84 (2005). [DOI] [PubMed] [Google Scholar]
  • 19.Goans, R. E., Holloway, E. C., Berger, M. E. & Ricks, R. C. Early dose assessment following severe radiation accidents. Health Phys.72, 513–518. 10.1097/00004032-199704000-00001 (1997). [DOI] [PubMed] [Google Scholar]
  • 20.Fliedner, T. M., Graessle, D., Meineke, V. & Dorr, H. Pathophysiological principles underlying the blood cell concentration responses used to assess the severity of effect after accidental whole-body radiation exposure: an essential basis for an evidence-based clinical triage. Exp. Hematol.35, 8–16. 10.1016/j.exphem.2007.01.006 (2007). [DOI] [PubMed] [Google Scholar]
  • 21.Goans, R. E., Holloway, E. C., Berger, M. E. & Ricks, R. C. Early dose assessment in criticality accidents. Health Phys.81, 446–449. 10.1097/00004032-200110000-00009 (2001). [DOI] [PubMed] [Google Scholar]
  • 22.Baranov, A. E., Guskova, A. K., Nadejina, N. M. & Nugis, V. Chernobyl experience: biological indicators of exposure to ionizing radiation. Stem Cells. 13 (Suppl 1), 69–77 (1995). [PubMed] [Google Scholar]
  • 23.Park, U., Kang, Y., Lee, H. & Yun, S. A. Stacking heterogeneous ensemble learning method for the prediction of Building Construction Project costs. Appl. Sci.12, 9729 (2022). [Google Scholar]
  • 24.Dey, R. & Mathur, R. 159–169 (Springer Nature Singapore).
  • 25.Proskura, P. & Zaytsev, A. Effective Training-Time Stacking for Ensembling of Deep Neural Networks. Proceedings of the 5th International Conference on Artificial Intelligence and Pattern Recognition (2022).
  • 26.Gupta, A., Jain, V. & Singh, A. Stacking ensemble-based Intelligent Machine Learning Model for Predicting Post-COVID-19 complications. New. Generation Comput.40, 987–1007. 10.1007/s00354-021-00144-0 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing Vienna, Austria, (2022).
  • 28.Kursa, M. B. & Rudnicki, W. R. Feature selection with the Boruta Package. J. Stat. Softw.36, 1–13. 10.18637/jss.v036.i11 (2010). [Google Scholar]
  • 29.Sullivan, J. M. et al. Assessment of biodosimetry methods for a mass-casualty radiological incident: medical response and management considerations. Health Phys.105, 540–554. 10.1097/HP.0b013e31829cf221 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Golob, J. F. Jr. et al. Fever and leukocytosis in critically ill trauma patients: it’s not the urine. Surg. Infect. (Larchmt). 9, 49–56. 10.1089/sur.2007.023 (2008). [DOI] [PubMed] [Google Scholar]
  • 31.Capaccio, C. et al. CytoRADx: a High-Throughput, standardized Biodosimetry Diagnostic System based on the Cytokinesis-Block Micronucleus Assay. Radiat. Res.196, 523–534. 10.1667/RADE-20-00030.1 (2021). [DOI] [PubMed] [Google Scholar]
  • 32.de Pablo, R., Monserrat, J., Prieto, A. & Alvarez-Mon, M. Role of circulating lymphocytes in patients with sepsis. Biomed. Res. Int.2014 (671087). 10.1155/2014/671087 (2014). [DOI] [PMC free article] [PubMed]
  • 33.Mulder, P. P. G. et al. Burn-Induced Local and Systemic Immune Response: Systematic Review and Meta-Analysis of Animal Studies. J Invest Dermatol 142, 3093–3109 e3015, doi: (2022). 10.1016/j.jid.2022.05.004 [DOI] [PubMed]
  • 34.Sen, S. et al. Early clinical complete blood count changes in severe burn injuries. Burns. 45, 97–102. 10.1016/j.burns.2018.09.004 (2019). [DOI] [PubMed] [Google Scholar]
  • 35.Tucker, J. D. et al. Accurate gene expression-based biodosimetry using a minimal set of human gene transcripts. Int. J. Radiat. Oncol. Biol. Phys.88, 933–939. 10.1016/j.ijrobp.2013.11.248 (2014). [DOI] [PubMed] [Google Scholar]
  • 36.Paul, S. & Amundson, S. A. Development of gene expression signatures for practical radiation biodosimetry. Int. J. Radiat. Oncol. Biol. Phys.71, 1236–1244. 10.1016/j.ijrobp.2008.03.043 (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Abend, M., Blakely, W. F., Ostheim, P., Schuele, S. & Port, M. Early molecular markers for retrospective biodosimetry and prediction of acute health effects. J. Radiol. Prot.4210.1088/1361-6498/ac2434 (2022). [DOI] [PubMed]
  • 38.Port, M. et al. Rapid Prediction of Hematologic Acute Radiation Syndrome in Radiation Injury patients using peripheral blood cell counts. Radiat. Res.188, 156–168. 10.1667/RR14612.1 (2017). [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Table S1 (99.6KB, xlsx)
Supplementary Table S2 (14.8KB, xlsx)
Supplementary Table S3 (19.5KB, xlsx)
Supplementary Table S4 (31.3KB, xlsx)
Supplementary Figures S1-5 (516.9KB, pdf)

Data Availability Statement

All datasets analyzed during this presented study are available in Supplementary Table S1 online and also available from the corresponding author upon reasonable request.

Articles from Scientific Reports are provided here courtesy of Nature Publishing Group

RESOURCES