Introduction

Radiation exposure is a major threat to public health whether in the form of an improvised nuclear device (IND), nuclear reactor accidents caused by natural calamities such as what happened in Fukushima, or the loss of radioactive sources. Following a mass-casualty radiological/nuclear incident, there will be a critical need to rapidly evaluate potentially exposed individuals for clinical triage and medical interventions. Modelling studies performed by Lawrence Livermore National Laboratory for a 10 KT (kiloton) nuclear detonation1 and prompt radiation exposures within an urban environment suggest that up to one million individuals could be subjected to triage based on the prediction that there will be significant infrastructure damage from ground zero to approximately a 2 mile radius2,3,4,5. About 0.75–1 Gy is the threshold level of exposure that induces mild radiation illness. Although this level of exposure is not anticipated to pose an immediate threat to life, individuals receiving this dose of radiation may still need medical management and treatment for symptoms or a follow-up evaluation. Individuals exposed to more than 2 Gy are at risk of suffering substantial damage to multiple organ systems and tissues, most notably the hematological and gastrointestinal systems, leading to the development of acute radiation syndrome (ARS). These individuals will benefit considerably from timely medical attention6,7.

Preparedness for and response to such a catastrophic event requires that emergency responders be able to discriminate between different levels of radiation exposure quickly and accurately. Currently, no radiation biodosimetry methods have been cleared or approved by the U.S. Food and Drug Administration (FDA)8. The ability to rapidly triage patients and characterize exposure level is essential to enable effective prioritization of scarce medical resources. Unfortunately, the lack of effective high throughput laboratory-based assays for biodosimetry triage tools are a recognized deficiency in national preparedness. Computer based software diagnostic tools are available for use by health-care providers including the EAST tool9, BAT10 and HemoDose11. These approaches largely rely on data from complete blood counts, leukocyte depletion kinetics and clinical symptoms (i.e., nausea and vomiting) to assist in identifying individuals for radiological triage.

Presently, several biodosimetry tools are in development that use cytogenetics12, proteomics13,14, and genomics15 endpoints. The CytoRADx system developed by ASELL™ employs a high throughput process to perform a standardized micronucleus assay that has been validated in human and NHP blood samples in ex vivo and in vivo models up to a radiation dose of 8 Gy12. Although cytogenetic endpoints such as dicentrics and micronuclei are well-established and validated for biodosimetry16,17, their main drawback is that the assay protocols require 2–3 days of cell culture18,19, whereas proteomic and genomic-based biodosimetry tools offer a much shorter time-to-result. Measurement of γ-H2AX foci formed at or near the vicinity radiation-induced DNA double-strand breaks (DSBs) is also an established rapid and sensitive biodosimetry method20,21 with throughput capability22,23,24. However, the limitation of this biomarker as a biodosimetry method for radiation triage is that the γ-H2AX protein signal is relatively short-lived after ionizing radiation25,26 due to the fast kinetics of DNA DSB repair (~ 24–48 h, depending on dose). SRI International is developing a lateral flow immunoassay point-of-care device for radiological triage that can quantify cytokine markers such as AMY1, FLT3L and MCP-1 in blood plasma from NHPs up to 7 days post exposure to classify samples < or ≥ 2 Gy27. Recent advancements in genomic mRNA markers have led to the development of a PCR-based high throughput ARad biodosimetry test by Arizona State University with MRIGlobal to estimate absorbed dose between 0 and 10 Gy from blood sampled 1–7 days post exposure, generating a risk–benefit analysis for the estimated absorbed dose28. Additionally, the REDI-Dx Biodosimetry test system developed by DxTerity, measures RNA expression in blood using the DxDirect genomic platform to classify absorbed dose at 2 thresholds above 2 Gy and above 6 Gy15.

At the Columbia University Center for High-Throughput Minimally Invasive Biodosimetry, we have developed the FAST-DOSE (Fluorescent Automated Screening Tool for Dosimetry) bioassay, designed to rapidly quantify radio-responsive intracellular proteins in blood leukocytes by imaging flow cytometry (IFC) for retrospective dose reconstruction up to at least a week after exposure to ionizing radiation13,29,30. In our early work, we identified a panel of top-candidate intracellular protein biomarkers (DDB2, BAX, FDXR, TSPYL2 and ACTN1), using shotgun proteomics to assess proteome-wide changes in human CD45 + blood leukocytes in X-irradiated humanized mice29. Since then, we have evaluated the performance of the FAST-DOSE biomarker system across different models and species including a human blood ex-vivo model31, non-human primates (NHP)13, and both humanized and C57BL/6 mice13,30. In the present work, we have transitioned and integrated two of our top-performing biomarkers BAX (BCL2 associated X, a regulator of apoptosis32,33) and DDB2 (DNA damage specific binding protein, a protein which binds to DNA as part of the cellular response to DNA damage34) into an ELISA-based platform with the goal to simplify the assay, and reduce the time-to-result. Here, we used human and NHP blood ex vivo culture models to evaluate biomarker dose response up to 48 h after exposure to acute dose X rays. We also compared their radiation response in vivo in blood samples collected from four NHPs exposed to a single total body irradiation dose of 2.5 Gy, up to 14 days after exposure. Using our custom-designed machine learning models, we classified radiation exposure (exposed vs. unexposed) and absorbed radiation dose based on the dose-dependent response of the intracellular protein biomarkers together with measurements of leukocyte cell counts/viability.

Results

Cell viability and correlation with dose

Leukocyte viability based on PI/AO staining measured on days 1 and 2 post-irradiation of both human and NHP blood samples exposed ex vivo showed that the percentage of surviving leukocytes decreased as the dose of radiation exposure increased. Figure 1a (human) shows that the average cell viability in the day 1 cultures was 97.7% ± 0.4% in the control group compared to 86.4% ± 2% in samples exposed to 5 Gy X rays (n = 10), whereas by day 2, 92.9% ± 2.7% of the control group leukocytes and 74.5% ± 4.2% in cells exposed to 5 Gy X rays (n = 10) remained viable. As seen in Fig. 1b (NHP), overall leukocyte cell viability was lower in the NHP cultures compared to the human cultures which was likely due to the fact that the NHP blood samples were shipped fresh overnight from Wake Forest. On days 1 and 2, the control NHP samples showed an average viability of 86.2% ± 1.9% and 83.1% ± 2% and in the 5 Gy-irradiated blood samples viability was 52.9% ± 2.6% (n = 13) and 49.4% ± 5% (n = 12), respectively.

Figure 1
figure 1

Percentage of leukocyte viability was measured by AO/PI staining on day 1 and day 2 post-irradiation. (a) Average human and (b) NHP leukocyte viability vs. dose plotted for each day. Error bars represent mean ± SEM.

Full size image

Viability was studied as a function of radiation dose at each timepoint using Pearson’s correlation coefficient: The human samples (Fig. 1a), showed a non-linear correlation between dose and leukocyte viability on days 1 (r = 1, p = 0.003; R2 = 0.97) and 2 (r = 0.955, p = 0.003; R2 = 0.92); Further, there was no apparent sex-based difference in leukocyte viability on either day (p > 0.35). Similarly, in the NHP samples (Fig. 1b), a non-linear correlation was observed between dose and leukocyte viability on day 1 (r = 1, p = 0.006; R2 = 0.91) and day 2 (r = 0.935, p = 0.006; R2 = 0.87). These data indicate that a high percentage of unirradiated leukocytes survived isolation and 2-day culture, whereas cells exposed to X-ray exhibited reduced cell viability and apparent radiation-induced cell death.

Quantification of BAX and DDB2 in X-irradiated human PBMCs

BAX and DDB2 levels negatively correlated with leukocyte viability on both days: BAX on day 1 (r = − 0.822, p = 0.044) and day 2 (r = − 0.964, p = 0.001) and DDB2 on day 1 (r = − 0.953, p = 0.003) and day 2 (r = − 0.989, p = 0.0002). Figure 2a shows the dose response of BAX concentration in human leukocyte cell lysates at 24 h (n = 11) and 48 h (n = 10) after exposure. There was a significant positive correlation between dose and BAX response on day 1 (r = 1.000, p = 0.002) and day 2 (r = 0.961, p = 0.002) which could not be fitted to a linear regression curve (R2 > 0.33). A significant difference was found between the unirradiated and 1 Gy cultures on days 1 (p = 0.0002) and day 2 (p = 0.0003). Overall, BAX yields were significantly (p < 0.025) higher in the day 1 samples exposed to 1–3 Gy compared to day 2. Figure 2b shows that there was no significant difference in BAX concentrations between male and female donors on either day (p > 0.35).

Figure 2
figure 2

(a) BAX concentration in human peripheral blood samples (after 1 or 2 days of cell culture) vs. dose. All data points plotted with lines connecting mean concentration for each dose on each day. (b) BAX in female (n = 4 dose/day) and male (day 1, n = 7 and day 2, n = 6) was not significantly different (paired t test). Box and whisker plots show minimum, median, quartiles and maximum BAX concentrations for each sex on each day at each dose.

Full size image

Figure 3a shows the dose response for DDB2 at 24 h (n = 8) and 48 h (n = 8) after exposure. There was a significant positive correlation between dose and DDB2 response on day 1 (r = 0.965, p = 0.001) and day 2 (r = 0.974, p = 0.001) which could be fitted reasonably well to a simple linear regression model on days 1 (R2 = 0.830) and 2 (R2 = 0.786). Unlike BAX, there was no significant difference (p > 0.55) in DDB2 expression between the control and 1 Gy cultures, however, there was a significant dose dependent increase in DDB2 yields from 1 to 5 Gy on both days. DDB2 levels were significantly higher (p < 0.013) on day 1 after exposure to 2 and 5 Gy X rays. Figure 3b shows the mean DDB2 levels were significantly higher (p = 0.030) in the male samples on day 1 but not on day 2 when compared to female samples.

Figure 3
figure 3

(a) DDB2 concentration in human peripheral blood samples (after 1 or 2 days of cell culture) vs. dose. All data points plotted with lines connecting mean concentration for each dose on each day. (b) DDB2 concentration presented by sex (on both days, female n = 5 and male n = 3) showed no significant differences. Box and whisker plots show minimum, median, and maximum DDB2 concentration values.

Full size image

Quantification of BAX and DDB2 in X-irradiated NHP PBMCs

Overall, on both days, BAX and DDB2 levels showed a significant negative correlation with leukocyte viability in the NHP cell cultures using Pearson correlation coefficients: for BAX on day 1 (r = − 0.559, p = 0.249) and day 2 (r = − 0.553, p = 0.255); for DDB2 on day 1 (r = − 0.939, p = 0.006) and on day 2 (r = − 0.985, p = 0.0003). Figure 4a,b show the dose responses for BAX and DDB2 in NHP blood samples irradiated ex vivo up to 5 Gy. All NHP samples (n = 13 on both days) were collected from male NHPs. There was no significant correlation between BAX concentration and dose on either day 1 (r = 0.679, p = 0.138) or day 2 (r = 0.472, p = 0.345). There was a significant difference in BAX levels between the control 0 Gy and 1 Gy samples on days 1 (p = 0.014) and 2 (p = 0.011) with no measurable dose dependent increase across the irradiated doses. DDB2 showed a significant positive correlation for concentration and dose on both days (day 1, r = 0.991 and day 2, r = 0.991; p = 0.0001 on both days). There was a significant difference between the control and 1 Gy samples on day 1 (p = 0.0002), whereas at day 2 there was no significant difference (p = 0.057). BAX did not fit linearly with dose on both the time points (R2 < 0.2), whereas the DDB2 response fitted reasonably well with dose on day 1 (R2 = 0.812) and for day 2 it did not fit that well (R2 = 0.619).

Figure 4
figure 4

Dose response curves for BAX (a) and DDB2 (b) in NHP peripheral blood samples exposed to X rays ex vivo. All data points plotted, lines connect means for each day and each dose.

Full size image

Dose reconstruction

The dose classification (exposed vs. unexposed) and reconstruction (quantitative dose predictions) results for the human and NHP data (on the testing data subset) are shown in Figs. 5 and 6 which also include the performance metrics on each testing data set. The combination of the three markers of radiation exposure, including leukocyte cell counts/viability (Supplementary Table 1) for all the donors and animals, at both the time points has been measured and intracellular concentration of BAX and DDB2 were successfully used to classify samples as exposed or non-exposed (Fig. 5). In humans, the assay and stacking machine learning analysis methods were 97.92% accurate in predicting the exposure status of a sample on testing data, with a true positive rate of 100% and a true negative rate of 88.89%. Of the 39 samples exposed to radiation of 1 Gy or more that were used for testing in the binary dose classification model, all 39 were correctly predicted to be exposed samples based on their viability, BAX, and DDB2 data. Of the 9 non-irradiated samples that were used for testing in this model, 8 were correctly predicted to be non-exposed, and only 1 was mis-classified as radiation-exposed (Fig. 5a). In NHPs, the assay and analysis were slightly less successful in classifying radiation exposure, and similar to humans, the model struggled more with identifying non-irradiated samples compared to identifying samples that had been irradiated. The binary classification model was 96% accurate in NHPs, with a true positive rate of 100% and a true negative rate of 72.73%. All 64 irradiated samples were classified as such, while 8 of 11 non-irradiated samples used for testing were correctly classified as non-irradiated (Fig. 5b).

Figure 5
figure 5

Performance of the stacking ensemble for classifying samples as irradiated or not. (a) human data, (b) NHP data. Shown are the ROC curve (95% CIs were generated by 10,000 bootstrap replicates) and confusion matrix statistics on testing data portion for each species.

Full size image
Figure 6
figure 6

Performance of the stacking ensemble for reconstructing dose quantitatively. (a) human data, (b) NHP data. The observed and reconstructed dose values are compared using scatter plots (left) and violin plots (right) for each species.

Full size image

The proposed approach also achieved good performance for reconstructing dose in a quantitative manner. In humans, plotting dose reconstruction values generated by modeling against the true dose for each sample in the testing set (Fig. 6a) produced R2 = 0.7914, RMSE (Root Mean Square Error) = 0.8007 Gy, and MAE (Mean Absolute Difference) = 0.6304 Gy. Here, BAX and DDB2 concentration, and leukocyte cell counts were used to build this dose reconstruction model. Leukocyte viability did not pass Boruta testing and so was not included in the model. In NHPs, the relationship between reconstructed dose and actual dose of the testing set produced R2 = 0.7980, RMSE = 0.7816 Gy, and MAE = 0.6099 Gy (Fig. 6b). Here, BAX and DDB2 levels, leukocyte cell counts, and cell viability passed Boruta testing and were used in the dose reconstruction model.

In vivo biomarker validation

BAX and DDB2 protein biomarker levels were also measured in peripheral blood samples collected from healthy NHPs (n = 4) exposed to 2.5 Gy total-body radiation in vivo with blood sampling on days 2, 5 and 14. Figure 7a shows that intracellular BAX levels significantly increased on day 2 (p = 0.021) and day 5 (p = 0.010) when compared to that of pre-irradiated samples with also a significant increase between days 2 and 5 (p = 0.004). Figure 7b shows a similar response for DDB2, whereby protein expression is significantly increased on day 2 (p = 0.015) and day 5 (p = 0.014) compared to the pre-irradiated samples, and between days 2 and 5 (p = 0.020).

Figure 7
figure 7

Concentration of (a) BAX and (b) DDB2 measured in NHPs exposed in vivo to whole-body irradiation of 2.5 Gy, *p < 0.01 and **p < 0.005.

Full size image

Discussion

A radiological/nuclear (R/N) accident would result in exposure of thousands of individuals to radiation, who will be needing medical intervention in a resource-constrained environment but would also result in a sense of radiophobia35,36 in millions of people wondering whether they are exposed or not. A similar mass anxiety around exposure was witnessed recently in the Covid-19 pandemic. In preparation for an R/N emergency there is a critical need for the development of high throughput biodosimetry tests14 that can accurately provide information on biological absorbed dose or provide triage capability to distinguish between individuals exposed to doses above and below 2 Gy37,38, thereby prioritizing victims who will benefit most from prompt medical attention and treatment. Previously, we have used our FAST-DOSE bioassay system to quantify the upregulation of a panel of intracellular radio-responsive proteins in blood leukocytes by imaging flow cytometry platform for dose assessment up to a week after radiation exposure. The goals of this work were to (1) transition two of our top performing FAST-DOSE biomarkers BAX and DDB213,31 to an ELISA-based platform and quantify their radiation response in peripheral blood leukocytes, and (2) develop machine learning (ML) models designed to integrate multiple biomarkers (including blood counts/viability) types into a comprehensive quantitative modeling framework to classify exposure and predict absorbed dose.

In the present work, we used human and NHP ex vivo culture models to evaluate BAX and DDB2 concentration levels at 24 h and 48 h after exposure of 0–5 Gy X rays. Both biomarkers showed a strong correlation with cell viability (coefficiency > − 0.82) and dose (coefficiency > 0.96). BAX exhibited significant sensitivity in blood samples exposed to 1 Gy, that remained persistently elevated up to 5 Gy, observed as a flattening of the curve with increasing dose, whereas DDB2 levels showed a dose-dependent response that could be fitted to a linear regression curve (R2 > 0.79) on both days (Figs. 2, 3). Recently, using the wild-type C57BL/6 mouse model and ML methods, we determined that a combination of intracellular biomarkers DDB2, FDXR, ACTN1, peripheral blood B and T cell counts and percentages can successfully be used to reconstruct dose and distinguish between PBI and TBI exposures30. In this study, we also applied a similar machine learning approach and used a combination of intracellular biomarkers BAX and DDB2, leukocyte cell counts and viability measurements (Supplementary Table 1) for prediction of radiation exposure classification and dose reconstruction.

The prediction accuracy of our ELISA-based proteomic biomarker assay to discriminate the unirradiated from the irradiated samples post exposure was determined using raw biomarker values and AUROC (area under the ROC curve) performance on each testing dataset (Fig. 5). The median AUC and their confidence intervals (CI) for the human and NHP ex-vivo samples were 0.9914 (95% CI 0.978–1.0) and 0.987 (95% CI 0.965–1.0), respectively. The binary classification model was 97.92% accurate in predicting the exposure status in humans and 96% accurate in NHPs. For dose reconstruction, a combination of BAX and DDB2 and leukocyte cell counts, and viability was used to quantitatively predict radiation dose in both the human and NHP samples (Fig. 6). The performance metrics showed an adequate correlation between predicted and actual dose in the human samples (R2 = 0.79, RMSE = 0.80 Gy, and MAE = 0.63 Gy) and NHP (R2 = 0.80, RMSE = 0.78 Gy, and MAE = 0.61 Gy). The Boruta feature selection component of the machine learning workflow also rejected Day and Sex (NHPs were all male) variables as predictors for both exposure classification and dose reconstruction. A limitation of these studies is that we have used relatively small sample sizes and a relatively conservative dose range (0–5 Gy). In future work, we plan to extend the human and NHP blood ex vivo models to include higher doses (up to 8–10 Gy) and evaluate the effect of biological variables such as age, sex, immune status and specific health conditions that could potentially confound dose predictions.

Measurements of BAX and DDB2 in vivo in four NHPs exposed to a single total body irradiation dose of 2.5 Gy showed a persistent upregulation in blood samples collected on days 2 and 5 after exposure (Fig. 7). These results are comparable with our previous study in the NHP in vivo model13, where both biomarkers showed a persistently increased expression up to day 8 after 2–5 Gy of total body exposures. It is well known that both BAX and DDB2 are involved in apoptosis and DNA repair, which are part of the DNA damage response cellular response to ionizing radiation. Previous studies have shown that these biomarkers can act as early predictors of individual radiosensitivity in patients undergoing radiotherapy to monitor risk and biomarker response32,39 as well as predictive markers for therapeutic response40,41,42. This work highlights the need for continued in vivo performance testing of our FAST-DOSE protein panel13,30 and longitudinal measurements for biomarker dose response up to 10–14 days after radiation exposure. Although measurements of cell count/viability only marginally improved the dose prediction in the ex vivo studies (Supplementary Fig. 2), our previous work in humanized mice43 and C57BL/6 mice30 show that T and B cell counts significantly contributed to the accuracy of exposure classification and dose prediction. Integration of complete blood counts with differential into our machine learning algorithm could also potentially help to improve the accuracy of dose classification and dose estimation as well as provide useful diagnostic information for severity of hematopoietic acute radiation syndrome.

In future studies, we plan to also expand our FAST-DOSE protein panel to include established radio-responsive plasma proteins which have successfully and used for biodosimetry in mouse44 and NHP models27,45. Towards the development of our custom-designed ML platform, we will test the importance of individual biomarkers from a panel of 3 biomarker types (leukocyte proteins, plasma proteins, and differential blood cells counts) and train robust ensemble models for predicting definitive dose (in Gy as a continuous variable) and categorical dose, binned into the following relevant categories for acute exposure: low (< 2 Gy), moderate (2–6 Gy) and high (> 6 Gy). Rapid identification of individuals (human estimated LD50/60 is ~ 4 Gy)46 who have received moderate to high doses of irradiation are most likely to benefit from treatment options for the mitigation of hematopoietic (2–6 Gy dose range) and gastrointestinal (> 6 Gy) injuries15 following a mass-casualty R/N incident. There is also a high value in a negative test to relieve concerned citizens and to identify individuals who have been exposed to doses < 1 Gy. We envision that the advanced optimization and development of our ELISA-based biodosimetry tool can provide a rapid and accurate determination of absorbed radiation dose and in conjunction with clinical data, guide early medical management decisions.

In summary, we have successfully transitioned two radiation responsive protein blood biomarkers, DDB2 and BAX into a high throughput ELISA platform and developed a custom-designed ML algorithms towards the development of a rapid and sensitive (≥ 1 Gy) bioassay tool with a time-to-result < 4 h. We highlight the importance of integrating multiple biomarker types into a comprehensive quantitative ML framework that can adapt and potentially enhance biodosimetry based on a fully integrated biomarker signature.

Materials and methods

Peripheral blood sample collection

Human peripheral blood samples (~ 10 mL) were collected from 11 healthy donors (6 male and 5 female), aged 22–66 years old, with no previous radiation exposure in the 6 months prior to the day of blood draw. Written informed consent was obtained according to approved Columbia University Irving Medical Center (CUIMC) IRB Protocol AAAU2786. Peripheral blood samples were collected by venipuncture in BD Vacutainers® with sodium-heparin (Becton, Dickinson and Company, Franklin Lakes, NJ; #367878).

Non-human primate (NHP; Macaca mulatta) blood samples (~ 6 mL) were collected from 13 male NHPs (age ranged from 8 to 21 years) at the Wake Forest Radiation Late Effects Cohort (RLEC) at Wake Forest University School of Medicine (WFUSM). Blood was collected in BD Vacutainers® with sodium heparin (Becton, Dickinson and Company, Franklin Lakes, NJ; #367878) and transported in temperature-controlled shipping boxes (Credo Cube, Series 22-248, Minnesota Thermal Science, Plymouth, MN) to CUIMC by FedEx priority overnight shipping. For the in vivo NHP irradiation studies performed at CUIMC (approved IACUC protocol #AABE5558), peripheral blood samples (2–3 mL) were similarly collected in BD sodium-heparin tubes from four healthy adult male rhesus monkeys, ages ranging from 11 to 18 years old on days 2, 5, and 14 days after 2.5 Gy irradiation. All NHP studies were conducted under all relevant federal and state guidelines and approved by the WFUSM and Columbia University IACUC.

Blood sample and animal irradiation

For the ex vivo studies, blood samples from both adult humans and NHPs were aliquoted in 15 mL polypropylene tubes (Corning, Glendale, AZ; #352095) and irradiated ex vivo using an X-RAD 320 biological irradiator (Precision X-Ray Inc., North Branford, CT) up to total doses of 0 (mock irradiated), 1, 2, 3, 4, or 5 Gy X-rays under the following conditions: 1.5 mm Al, 0.25 Cu, 1.25 Sn and 320 kVp, 12.5 mA, FSD 40, 0.95 Gy/min, custom-made home filter. Before each sample was exposed to radiation, the dose rate was verified using a Radcal 10X6-6 ion chamber (Monrovia, CA; calibrated annually by Radcal).

For the in vivo irradiations, NHPs were whole body irradiated with 2.5 Gy using a Varian Trilogy linear accelerator (LINAC) (Varian Medical Systems, Palo Alto, CA) at 0.2–0.3 Gy/min while enclosed in a specially designed clear acrylic box. An on-site medical physicist performed all the dose calculations for radiation exposure. Prior to transport to the LINAC, while still in animal housing, each animal was anesthetized with a ketamine-containing anesthetic mixture (ketamine hydrochloride 100 mg/mL and acepromazine 10 mg/mL at a dose of 0.1 mL/kg) and placed into the acrylic box which was housed inside a high efficiency particulate air (HEPA) filtered transport cage on a wheeled cart for transport to the LINAC. This transport system is designed with a viewing window to permit continuous monitoring and observation during transport and partial pressure of oxygen (PO2) was monitored continuously. The transport cart was always escorted by veterinary staff with a two-way radio link to the animal facility in case of any emergency. After irradiation, the animals were transported back to housing, placed back into their cages, and continuously monitored until recovered from anesthesia. As a control for any potential effects of transport and preparation for TBI, roughly 2–4 weeks prior to irradiation, blood samples were drawn 24 h after animals were similarly anesthetized, prepared for irradiation, and transported to the LINAC, but not exposed to X rays.

Sample preparation: Isolation of leukocytes and cell lysate preparation

Leukocytes were isolated from the whole blood using density gradient centrifugation. Ficoll Histopaque medium (Sigma Aldrich, St. Louis, MO, #10771-human and #10831-NHP) was added first to 15 mL SepMate™ tubes (STEMCELL™ Technologies; Vancouver, BC, #85415), and 1 mL of blood sample gently poured down the side of the tube. Samples were centrifuged at 1200×g for 10 min and the top layer containing peripheral blood mononuclear cells (PBMC) was transferred to a fresh 15 mL polypropylene tube and washed with 1XPBS (Gibco, Grand Island, NY). The washed PBMCs were aliquoted (~ 1 × 106/mL) into two Matrix™ 1.0 mL microtubes (Thermo Fisher Scientific™, Waltham, MA, #3740TS) per sample with complete RPMI (15% FBS, 1% Pen-Strep) and cultured at 37 °C, 5% CO2 for 1 and 2 days. After the culture time, for ELISA cell lysate preparation the cells were spun down and washed with 1XPBS, and then chilled 1X Cell Extraction Buffer PTR (Abcam, Waltham, MA, #ab193970 for BAX; and 1 X PBS with 1X protease inhibitor cocktail HALT (Thermo Fisher Scientific™, Waltham, MA, #87785 for DDB2) was added to the cell pellet and incubated on ice for 20 min. After the incubation, cells were centrifuged at 18,000×g for 20 min at 4 °C and stored at − 80 °C until use.

Cell count and viability

Cell count and viability staining were performed after 24 and 48 h PBMC culture. Cells were stained with Acridine Orange/Propidium Iodide (AO/PI) viability dye (Logos Biosystems, Annandale, VA, #F23001) and loaded into PhotonSlide™ (Logos Bio-systems, #L12005). A LUNA-FL™ Dual Fluorescence Cell Counter (Logos Biosystems, #L20001) was used to automatically count and determine the viability percentage of the cells, as per manufacturer’s instructions.

Protein analyses of cell lysates by enzyme-linked immunosorbent assays (ELISA)

Total protein in the cell lysates was quantified using Pierce™ BCA protein assay kits (Thermo Fisher Scientific, Rockford, IL, #23225) as per the manufacturer’s instruction to develop and interpolate concentrations from a standard curve. The human and NHP immunoassays were performed in duplicate with a conventional ELISA sandwich format for two different protein targets, BAX and DDB2, using commercially available kits from Abcam (Waltham, MA, #ab199080) and AFG bioscience (Northbrook, IL, #EK712088), respectively. The absorbance readings were read at 450 nm, with reference to the standard curve and used average difference data between control and test samples as readout (see Supplementary Fig. 1 for BAX and DDB2 standard curves). The plates were read using BioTek Synergy H1 Multimode Microplate Reader (Agilent Technologies, Santa Clara, CA) and analyzed using the built-in Gen5 software. Optical density readings were then interpreted using GainData®, Arigo Biolaboratories’ online calculator to plot standard curves and interpolate unknown concentrations.

Statistical analysis

All statistical analyses were performed, and graphs produced using GraphPad Prism (version 10; GraphPad Software, Inc., La Jolla, CA). Human and NHP leukocyte viability were analyzed as functions of dose separately for each day via Pearson’s correlation and linear least squares regression. Separately, both BAX and DDB2 concentrations were correlated with leukocyte viability using Pearson’s (DDB2) and Spearman (BAX) correlations. BAX and DDB2 in the human and NHP samples were compared to each other across different doses using 2-, 3-way, and repeated measures ANOVA tests and by calculating Pearson’s correlations between the average concentration of the biomarker and dose. For humans, a further ANOVA test was run to determine if there was a significant difference in biomarker expression in male and female subjects. For the in vivo NHP study, data points from in vivo irradiated NHPs were compared to confirm differences in biomarker levels across timepoints using a repeated-measures ANOVA.

Dose reconstruction

Dose reconstruction calculations for the ex-vivo studies were performed in Python 3.10 on a Jupyter notebook platform. Only those samples which had both BAX and DDB2 measurements in the sample were used for the NHP or human data. We performed 50:50 splitting of the data set (48 samples for humans and 75 samples for NHP) into training and testing parts (see Supplementary Table 2 for detailed breakdown of the variable distributions in splits). The Boruta algorithm was used as an initial screening step (on training data) to discard the least important variables for distinguishing between unirradiated and irradiated samples (labeled by the Exposure index variable, where 1 = irradiated, 0 = unirradiated). Separately, a regression analysis was done to reconstruct dose quantitatively for humans or NHPs. Boruta screening was also used.

Boruta creates “shadow features” (randomized copies of original features) and compares their importance using Random Forest regressor or classifier models. If an original feature's importance is significantly higher than the maximum importance of the shadow features using z scores, it is kept; otherwise, it is dropped. This iterative process continues until all features are either confirmed important or unimportant using a pre-defined significance threshold.

Using the retained predictors, multiple machine learning (ML) algorithms (linear regression, random forest, XGBoost, LightGBM, CatBoost, elastic net and support vector machines for regression tasks, and logistic regression, CatBoost, XGBoost, random forest, K-nearest neighbors, and naïve Bayes for classification tasks) were fitted to the training data with repeated cross-validation and evaluated on testing data. For the regression task, root mean squared error (RMSE) was used as the main metric to assess performance and mean absolute error (MAE) and coefficient of determination (R2) were also calculated. For the classification task, balanced accuracy was used.

The stacking approach was used to integrate the outputs of these different ML models to generate an ensemble. It was performed separately for each task. In stacking, several ML methods (level0 models) are applied to the training data with repeated k-fold cross validation. Predictions of each level0 model on out of sample data instances (those withheld during cross validation) are recorded. These predictions serve as inputs to train a meta-model (level1) which learns how to best combine the predictions of the level0 models to predict the outcome variable. Then the whole ensemble (level0 and level1) makes predictions on testing data. Often (but not always) this approach performs better than a single best level0 model, for example, certain samples may be difficult to predict for some models, but easier for other models, so information from several models can be complementary and improve overall predictions. Achieving an improvement in performance depends on the complexity of the problem and whether it is sufficiently well represented by the training data and complex enough that there is more to learn by combining predictions. Data for days 1 and 2 were combined since the Boruta algorithm considered the Day variable to be unimportant for both humans and NHPs for the purposes of exposure classification and dose reconstruction.