Dicentric Dose Estimates for Patients Undergoing Radiotherapy in the RTGene Study to Assess Blood Dosimetric Models and the New Bayesian Method for Gradient Exposure

Abstract

The RTGene study was focused on the development and validation of new transcriptional biomarkers for prediction of individual radiotherapy patient responses to ionizing radiation. In parallel, for validation purposes, this study incorporated conventional biomarkers of radiation exposure, including the dicentric assay. Peripheral blood samples were taken with ethical approval and informed consent from a total of 20 patients undergoing external beam radiotherapy for breast, lung, gastrointestinal or genitourinary tumors. For the dicentric assay, two samples were taken from each patient: prior to radiotherapy and before the final fraction. Blood samples were set up using standard methods for the dicentric assay. All the baseline samples had dicentric frequencies consistent with the expected background for the normal population. For blood taken before the final fraction, all the samples displayed distributions of aberrations, which are indicative of partial-body exposures. Whole-body and partial-body cytogenetic doses were calculated with reference to a 250-kVp X-ray calibration curve and then compared to the dose to blood derived using two newly developed blood dosimetric models. Initial comparisons indicated that the relationship between these measures of dose appear very promising, with a correlation of 0.88 (P= 0.001). A new Bayesian zero-inflated Poisson finite mixture method was applied to the dicentric data, and partial-body dose estimates showed no significant difference (P > 0.999) from those calculated by the contaminated Poisson technique. The next step will be further development and validation in a larger patient group.

INTRODUCTION

Background: The RTGene Project

Biological markers of radiation exposure play a crucial role in the triage of suspected exposed persons after a radiation accident or incident (1, 2). In recent years it has been shown that the gene expression assay is a sensitive marker of radiation exposure, with the potential to be used for truly individualized biological dosimetry (3–5). Classic cytogenetic techniques, and in particular the gold standard dicentric assay, have two main disadvantages in mass casualty scenarios: 1. lack of high-throughput; and 2. delays of several days between blood sampling and the availability of results (1). With the use of blood samples, gene expression analysis can provide valuable information, as there is a window of time (i.e., 12–48 h) postirradiation where specific radiation-responsive genes have linear dose responses (0–5 Gy) (5). New technology for gene expression analysis allows direct counting of nucleic acid molecules (DNA, mRNA, miRNA and lncRNA) without the need for enzymatic reaction or amplification steps, thereby reducing time for data collection (6); this has been assessed for radiation biodosimetry applications with promising results (7). Linearity of the transcriptional dose response for specific radiation-responsive genes in ex vivo exposed human blood samples has recently been demonstrated for the first time, and interindividual variability in the response after low-dose and high-dose exposures has been newly assessed (3, 5). The logical next stage for biological development of the gene expression assay was to further validate these new techniques with human blood samples exposed to radiation in vivo (8, 9). The RTGene Project was a feasibility study to develop existing knowledge on coding and noncoding transcriptional responses to ionizing radiation into a useable radiation-specific biomarker of exposure and response using blood samples from radiotherapy patients. In parallel, for validation purposes, the study included conventional biomarkers of radiation exposure, i.e., chromosome aberrations using the dicentric assay (DCA) and DNA damage using the γ-H2AX foci assay.

A range of standard radiotherapy schedules was chosen for inclusion in this study to provide a wide range of doses for assessment of the gene expression assay alone and in combination with the DCA, to simulate a wide range of potential exposure scenarios. Conventional cytogenetics was chosen for inclusion in the RTGene project, because the DCA is the most widely used and validated biological dosimetry assay, as well as being a standardized technique (1, 10). Whole-body (WB) and partial-body (PB) doses can be assessed based on the observed yield of dicentric chromosome aberrations with reference to an appropriate calibration curve (1). Not only can cytogenetic dose estimates be used to validate the gene expression assay, but they can be compared to the calculated dose to blood during irradiation from dosimetric models. In addition, the DCA data can be applied to a more sophisticated Bayesian zero-inflated Poisson finite mixture method to calculate PB dose estimates and then compared to the radiotherapy data.

Blood Dosimetric Models

Radiation treatment planning systems produce detailed maps of the predicted radiation dose to be delivered by the treatment units. The radiation dose is focused on an area outlined on a computed tomography (CT) image set by a radiation oncologist. This region is referred to as the target volume and is the site of the primary tumor, tumor bed or region to which the cancer has spread. Radiotherapy is most commonly delivered using photon radiation. Photon radiation is attenuated, but not stopped, by the body; thus, nontarget normal tissues in the path of the beam receive radiation dose. During the planning process, the radiation oncologist also outlines nontarget tissues/organs of particular concern in the vicinity of the target volume to minimize the radiation dose received by them. While radiotherapy dose information is reasonably accurate in the target volume (11), it provides information only on static objects that have been explicitly delineated in the treatment planning system. It can be used to infer dose to the circulating blood but does not give this directly. For this reason, two simple models were set up to test for correlation with the dicentric dose models.

The Bayesian Zero-Inflated Poisson Finite Mixture Method to Assess Partial-Body Exposure

Radiation produces damage at a cellular level in humans and, as mentioned above, the DCA is a well-established cytogenetic biomarker of radiation exposure. To calibrate the effect of radiation, dose-response curves are built, by exposing in vitro blood samples to different doses, simulating homogeneous whole-body exposure. For this kind of exposure, it is typically assumed that the yield of dicentrics per blood cell is a Poisson number whose intensity is a quadratic function of the dose, β₀ + β₁D + β₂D² [further details are reported elsewhere (1)].

Gradient exposures are heterogeneous irradiations where different doses occur in the radiation field within the individual’s body. Partial-body irradiations are those where the dose or doses are not absorbed by the whole body of the individual, i.e., there is a portion of the body that does not receive the radiation dose. Traditionally, PB dose has been estimated using, for example, the contaminated Poisson method (12), but more recently a Bayesian zero-inflated Poisson finite mixture model has been applied to cytogenetic data derived from simulated PB irradiations, by mixing one fraction of homogeneously irradiated blood with a nonirradiated sample (13).

Here, we examine dicentric dose estimates for patients undergoing radiotherapy enrolled in the RTGene study to assess blood dosimetric models and the Bayesian zero-inflated Poisson finite mixture method for estimating partial-body exposure. Additional results from gene expression, radiation-induced γ-H2AX foci and translocation analysis, plus a comparison of the cytogenetic dose estimates with the gene expression data, will be addressed in the future.

MATERIALS AND METHODS

Patient Selection and Blood Sampling

Eligible volunteers who required external beam radiation therapy for breast, lung, gastrointestinal or genitourinary tumors were identified in the Outpatient Department at The Royal Marsden NHS Foundation Trust (RM; Sutton, UK). Patients were included in the study if: 1. they were aged 18 years or older; 2. they had no previous radiotherapy; 3. they were not concurrently receiving chemotherapy or were not less than four weeks before radiotherapy; 4. they were not concurrently receiving hormone therapy or were not less than four weeks before radiotherapy; and 5. written informed consent was given, but could be withdrawn at any time. The study was performed in accordance with the Declaration of Helsinki (1964), the Research Governance Framework 2nd edition (2005) and the Human Tissue Act (2004). The study was approved by the South Central-Hampshire B Research Ethics Committee (16/SC/0307) and registered with ClinicalTrials.gov (NCT02780375). All relevant information was collected from the participants and extracted from their case notes by the research team at RM and only transferred to Public Health England (PHE) for the final analysis. No identifiable information was passed between the two institutions and patient confidentiality was maintained at all times. Heparinized venous blood was collected from a total of 20 volunteers, prior to irradiation and before the final fraction for the DCA. Coded samples were dispatched by express courier overnight to PHE. The 20 volunteers were comprised of patients undergoing radiation treatment for the following tumor types: breast (5 patients); endometrial (4 patients); prostate (3 patients); lung (5 patients); esophageal (2 patients) and colon (1 patient). The radiotherapy schedules and doses are shown in Table 1 for each of the 20 volunteers. Also included in Table 1 is the 95% isodose volume for each patient, to give an indication of the different sizes of the field irradiated.

TABLE 1.

Tumor Type, Radiotherapy Schedule and Prescribed Radiation Doses, Together with the 95% Isodose Volume for Each Patient

		Prescribed target	No. of	Prescribed target	95% Isodose
RTGene ID	Treated area	dose (Gy)	fractions	dose/fraction (Gy)	volume (cm3)
RTG002	Breast (right)	40.05	15	2.67	726
RTG003	Endometrium	45	25	1.80	1,779
RTG004	Breast (left)	40.05	15	2.67	954
RTG005	Breast (left)	40.05	15	2.67	734
RTG006	Breast (right)	40.05	15	2.67	1,044
RTG007	Endometrium	45	25	1.80	1,362
RTG008	Prostate	60	20	3.00	89
RTG009	Lung	55	20	2.75	520
RTG010	Lung	55	20	2.75	281
RTG011	Prostate	60	20	3.00	202
RTG012	Lung	55	20	2.75	624
RTG013	Lung	55	20	2.75	415
RTG014	Lung	55	20	2.75	927
RTG015	Endometrium	45	25	1.80	1,469
RTG016	Endometrium	45	25	1.80	1,134
RTG017	Prostate	60	20	3.00	123
RTG018	Esophagus	36	12	3.00	1,152
RTG019	Breast (both)	40.05	15	2.67	1,729
RTG020	Esophagus	20	5	4.00	2,197
RTG021	Colon	40	15	2.67	599

Dicentric Assay

On arrival at the laboratory whole blood was mixed with minimal essential medium (MEM) for the DCA (Sigma-Aldrich, Dorset, UK), supplemented with 10% heat-inactivated fetal bovine serum, 1% phytohemagglutinin, 100 units/ml penicillin plus 100 μg/ml streptomycin and 2 mM L-glutamine (all from Invitrogen, Paisley, UK). In addition, 5-bromo-2-deoxyuridine (Sigma-Aldrich) was added to the DCA cultures at a final concentration of 10 μg/ml. All samples were cultured at 37°C in a 5% CO₂ humidified atmosphere. After 45 h Colcemid (Sigma-Aldrich) was added to each culture to give a final concentration of 0.2 μg/ml. At 50 h metaphases were harvested by a standard hypotonic treatment in 0.075 M potassium chloride for 7 min at 37°C followed by three changes of 3:1 methanol:acetic acid fixative. Fixed cells were dropped onto clean microscope slides, air dried and stained using the fluorescence plus Giemsa technique. The culture, fixation and staining procedures followed the standard protocol recommended by the International Atomic Energy Agency (1). A maximum of 1,000 first-division metaphases per donor for the preirradiation sample and 500 cells or 100 dicentrics for the final sample were scored manually for chromosome aberrations. Dose estimates, based on the number of dicentrics per cell were calculated using Dose Estimate_v5.1 (14) and PHEs standard 250 kVp X-ray calibration curve, with the following coefficients: C = 0.0005 ± 0.0005, α = 0.046 ± 0.005, β = 0.065 ± 0.003 (15). In addition, the standard “contaminated Poisson” method to calculate the most likely partial-body dose, percentage of lymphocytes exposed and percentage of the body exposed was applied (1).

Simple Blood Dosimetry Models

The first model, model 1 (EDD1), uses typical values for the circulation time of blood in humans combined with the time taken for radiotherapy to be delivered. It uses the high dose (volume within 95% isodose curve on the radiation treatment plan) as a fraction of an assumed 6 liters of blood for a human with a blood flow rate of 6 l/ min. Irradiation duration of 1 min is assumed:

$${\text{D}}_{\text{B}}=\text{ dose to blood per fraction }={\text{D}}_{\text{F}}\times \left({\text{V}}_{95\%}÷{\text{V}}_{\text{B}}\right)(\text{Gy}),$$ (1)

where: V_95% = high-dose volume, which is specific to each patient (cc); V_B = total blood volume, assumed to be 6 liters; and D_F = radiotherapy dose per fraction, which is specific to each patient (Gy). The main limitations of model 1 are a lack of patient-specific blood volume, circulation time and no knowledge of volumes of blood in different organs. The uncertainty of V_95% is of the order of 2–3%; that of V_B is of the order of 20%; D_F is a set number (the specified dose prescription) and as such does not have an uncertainty; the blood flow rate and irradiation times will have variation of at least 20%. It is likely that D_B calculated using this method will have a minimum uncertainty of at least 20%.

The second model, model 2 (EDD2), uses patient-specific data. Model 2 estimates a whole-body mean dose and assumes the blood receives this. The whole-body dose is calculated using the mean dose for the volume of the body covered by the CT planning scan. This is scaled assuming the total-body volume is 2.5 times this volume:

$${\text{D}}_{\text{B}}=\text{ dose to blood per fraction }=\left({\text{D}}_{\text{PB}}÷{\text{D}}_{\text{F}}\right)\times 2.5(\text{Gy}),$$ (2)

where: D_PB = mean dose (Gy) of body volume covered by CT scan (specific to each patient); D_F = radiotherapy dose (Gy) per fraction which is specific to each patient. The main limitations of model 2 are the use of an estimate of 2.5 for the scaling factor from partial- to whole-body volume and a lack of knowledge of the amount of blood volume in specific organs. The uncertainty of D_PB is of the order of 2–3%; D_F is a set parameter and as such there is no uncertainty associated with it; the factor 2.5 is estimated to have an uncertainty of 30–40%. It is likely that D_B calculated using this method will have an uncertainty of approximately 40%. It is nontrivial to determine the volume of blood and blood flow through specific organs relevant to radiation and similarly, partial-body dose (EDD2) without whole-body imaging information. Investigations are underway into more sophisticated estimates using virtual body phantoms.

Bayesian Zero-Inflated Poisson Finite Mixture Method

The goal is to estimate the absorbed doses and the irradiated fractions for each irradiated component. In a scenario of partial-body gradient exposure with k irradiated components, given a sample y = {y₁, …, y_n} of n chromosome aberration counts within blood cells, the yield of chromosome aberrations can be represented by a zero-inflated Poisson finite mixture model whose probability mass function has the form

$$p\left(y_i|\omega ,\lambda \right)={\omega }_0{1}_{\left(y_i=0\right)}+{\displaystyle \sum _{j=1}^k{\omega }_j}p\left(y_i|{\lambda }_j\right),$$ (3)

where ω denotes the proportions (with ${\displaystyle {\sum }_{j=0}^{\text{k}}{\omega }_j}=1$ ), λ is the vector of Poisson intensities,$p\left(y_i|{\lambda }_j\right)=e^{-\lambda }{\lambda }^{y_i}/y_i!$ is the Poisson probability of observing y_i for expectation λ_j(>0), i is the index of observations and 1_(yi=0) takes the value 1 if y_i = 0 and 0 otherwise. Values λ_j, and ω_j, represent the yield of chromosome aberrations and the proportion of scored cells at component j, respectively. Value ω₀ is the proportion of extra zeroes, over and above those expected from a purely Poisson process, and represents the proportion of non-irradiated scored cells.

The doses for each component, D_j, are estimated by matching the yield of aberrations to the fitted dose-response curve,${\beta }_0+{\beta }_1{D}_j+{\beta }_2{D}_j^2={\lambda }_j$ To calculate the number of irradiated fractions within the body, F_j, it is necessary to rescale the proportion of scored cells by adding to each component the proportion of cells that died because of the irradiation, i.e.,

$$F_0=\frac{{\omega }_0}{{\omega }_0+{\displaystyle {\sum }_{l=1}^k{\omega }_1{e}^{D_1/d_0}}},F_j=\frac{{\omega }_j{\text{e}}^{D_j/d_0}}{{\omega }_0+{\displaystyle {\sum }_{l=1}^k{\omega }_1{e}^{D_1/d_0}}},$$ (4)

where d₀ is the 37% cell survival dose, with experimental evidence to be between 2.7 and 3.5 Gy (1). F₀ represents the fraction of the body nonirradiated and F_j represents the fraction of the body irradiated by dose D_j.

A Bayesian model is proposed to estimate both the doses and the fractions, assuming prior distribution densities for each of the parameters. The technique proposed here consists of two steps. The first step is to infer the yields and the proportions and the second is to obtain the estimation of the doses and the fractions using Eqs. (3) and (4), shown above.

Given a sample y and assuming the observations are independent, the likelihood is the product of the probability of the observations,$L(\omega ,\lambda |y_i)={\Pi }_{i=1}^{n}p\left(y_i|\omega ,\lambda \right)$ Assuming ω and all λ_j, values are independent the following prior structure is defined as:

$$\omega ~\text{ Dirichlet }\left({\overrightarrow{1}}_{k+1}\right),$$ (5)

$${\lambda }_j~\mathcal{U}(0,M),{\lambda }_{j+1}>{\lambda }_j,M=\max (y).$$ (6)

The prior for the proportions of scored cells, ω, is a flat Dirichlet distribution of k + 1 elements. The ordering constraint of the yields prior is to ensure identifiability. By the Bayes’ theorem, the joint posterior distribution of {ω,λ} is

$$p(\omega ,\lambda |y_i)=\frac{L(\omega ,\lambda |y_i)p(\omega ,\lambda )}{{\displaystyle \int L}(\omega ,\lambda |y_i)p(\omega ,\lambda )d\omega d\lambda },$$ (7)

where p(ω,λ) is the product of the prior densities of all λ_j and ω. The above joint posterior density has a non-tractable form, so acceptance-rejection sampling is used to simulate it. Let $\widehat{L}$ be the maximum value of $L(\omega ,\lambda |y_i)$ then the following steps are taken to sample the joint posterior distribution:

1. Generate u from u(0, M).

2. Generate one random variate for each prior, ω* and λ*, all of them independent of u.

3. Compute $L^{*}=L\left({\omega }^{*},\lambda *|y_i\right)$. If $u<\widehat{L}/L^{*}$, then set {ω*,λ*} to the joint posterior sampling.

4. When the size of the sample is lower than the desired value, go to step 1.

To get the joint posterior distribution of the doses and the fractions, a prior is defined for the calibration coefficients {β₀,β₁,β₂} based on the dose-response curve maximum likelihood estimation. Another prior is defined for the cell survival dose, which is uniform between 2.7 and 3.5 Gy. Keeping independency for all priors, the additional prior structure is defined as:

$$\beta ~N(\widehat{\beta },\widehat{\Sigma }),$$

$$d_0~\mathcal{U}(2.7,3.5).$$

Thus, the following steps are included in the previous algorithm after step 3 if the condition is met:

1. Generate one random variable for the new priors: β* and $d_0^{*}$.

2. Calculate a new sample for the doses by solving ${\beta }_0+{\beta }_1{D}_j^{*}+{\beta }_2{D}_j^{*2}={\lambda }_j^{*}$

3. Calculate the fractions from

$$F_0^{*}=\frac{{\omega }_0^{*}}{{\omega }_0^{*}+{\displaystyle {\sum }_{l=1}^k{\omega }_l^{*}}{e}^{D_1^{*}/d_0^{*}}},F_j^{*}=\frac{{\omega }_1^{*}{e}^{D_j^{*}/d_0^{*}}}{{\omega }_0^{*}+{\displaystyle {\sum }_{l=1}^k{\omega }_l^{*}}{e}^{D_l^{*}/d_0^{*}}}.$$ (8)

After this process, samples {F, D} represent the joint posterior densities and the posterior marginal densities and are represented by each F_j and D_j, for the joint sample.

This method was applied to the dicentric data to estimate partial-body doses, assuming 2, 3, 4, 5 and 6 irradiated fractions. Due to computational intensity, the number of simulated draws of the joint posterior densities is decreased as the assumption of the number of irradiated components increases. The simulation size for each scenario was as follows: 10,000 for 2; 1,000 for 3, 4 and 5; 100 for 6 irradiated fractions. The Bayesian Information Criterion (BIC) value was also calculated for the different scenarios.

Other Data Analysis

The distribution of dicentric aberrations among the scored cells for each sample was tested for conformity with the Poisson distribution by calculating the dispersion index (the ratio of variance to mean) using Dose Estimate_v 5.1 software (PHE, Chilton, UK) (14). Over-dispersion is indicated by a value > 1.0, thus pointing to a partial-body irradiation (1). To determine whether there was a statistically significant difference in dose response with cancer type, general linear model analysis of variance (GLM ANOVA) was performed, with post hoc testing using Tukey’s pairwise comparisons within factors, using Minitab^® 17 (State College, PA). For comparison of the Bayesian and standard PB method, the doses calculated by each technique were normalized and compared using the standard Student’s t test.

RESULTS

All baseline samples have dicentric frequencies consistent with the expected background for the normal population (0–2 in 1,000), with no cell containing more than 1 dicentric. The dicentric distributions for all the samples, preirradiation and prior to the final fraction, were tested for conformity to the Poisson distribution. As expected, there is no indication of departure from the Poisson distribution, so there is no evidence of recent whole- or partial-body exposures in the preirradiation samples. For the samples taken prior to the final fraction, all samples display distributions of aberrations which are indicative of partial body exposures to some degree, as shown in Table 2.

TABLE 2.

Dicentric Chromosome Aberrations in Samples Taken before Radiation Therapy and prior to the Final Radiation Fraction

RTGene ID	Preirradiation treatment sample		Sample taken prior to the final RT fraction
	Cells	Dics	Cells	Dics	Y	SE	Var:mean	SE
RTG002	1,000	1	500	20	0.040	0.009	1.360	0.062
RTG003	1,000	2	171	99	0.579	0.058	1.790	0.108
RTG004	1,000	1	500	19	0.038	0.009	1.390	0.062
RTG005	1,000	2	500	23	0.460	0.010	1.300	0.062
RTG006	1,000	0	500	19	0.038	0.009	1.070	0.062
RTG007	1,000	0	233	100	0.429	0.043	1.100	0.092
RTG008	1,000	1	500	36	0.072	0.012	1.150	0.062
RTG009	919	1	202	100	0.500	0.050	2.280	0.099
RTG010	514	1	488	101	0.207	0.039	1.990	0.064
RTG011	1,000	2	500	60	0.120	0.028	1.350	0.063
RTG012	1,000	0	203	100	0.493	0.091	2.020	0.099
RTG013	1,000	1	309	100	0.323	0.059	1.660	0.080
RTG014	1,000	2	132	103	0.780	0.143	1.830	0.123
RTG015	1,000	0	181	100	0.552	0.102	2.080	0.105
RTG016	1,000	1	264	100	0.379	0.070	1.270	0.087
RTG017	1,000	1	500	90	0.180	0.035	1.820	0.063
RTG018	1,000	2	181	100	0.552	0.102	1.680	0.105
RTG019	1,000	1	500	91	0.182	0.035	1.680	0.063
RTG020	1,000	1	500	99	0.198	0.037	3.010	0.063
RTG021	1,000	0	276	100	0.362	0.050	1.720	0.085

Notes. Cells = number of peripheral blood lymphocytes scored; Dies = number of dicentric chromosome aberrations identified; Y = yield of dicentrics; var:mean = variance:mean ratio, an indication of departure from Poisson and thus partial-body exposure (var:mean for Poisson = 1); SE = standard error of the measurement in the previous column. In all the preirradiation samples no cell contained more than 1 dicentric, thus the var:mean = 1.0 in all cases.

The BIC values for the different exposure scenarios, assuming PB irradiation, were calculated (data not shown). Lower BIC values indicate a better fit. Following this criterion, a PB irradiation with 2 irradiated components was the best fit for the dicentric data for all patients. The results of the cytogenetic dose estimates (standard and Bayesian methods) and dose to blood calculated from the two models are given in Table 3.

TABLE 3.

Dose after the Penultimate Fraction for each Participant Calculated Using Blood Dosimetric Models and the Cytogenetic Dose Estimates Calculated by Standard and Bayesian Methods

			Blood cytogenetic dose estimates
	Blood model doses		Whole body		Partial body				Bayesian partial body assuming two irradiated fractions (mean values)
RTGene ID	EDD1 (Gy)	EDD2 (Gy)	Dose (Gy)	SE	Dose (Gy)	SE	% Cells irradiated	% Body irradiated	% Body not irradiated	Dose 1 (Gy)	95% HPDCI	Fraction 1 (%)	Dose 2 (Gy)	95% HPDCI	Fraction 2 (%)
RTG002	2.26	1.08	0.50	0.08	2.34	0.42	8.6	18.3	67.7	1.28	0.00–2.77	21.5	2.96	1.45–4.89	10.7
RTG003	12.81	5.49	2.65	0.16	4.01	0.84	47.1	79.7	17.9	2.87	0.94–4.39	42.9	5.23	3.56–7.85	39.1
RTG004	5.94	1.26	0.48	0.08	2.01	0.38	10.7	20.2	58.8	1.04	0.01–2.32	32.1	3.27	1.17–6.05	9.2
RTG005	4.57	1.28	0.56	0.08	2.14	0.39	11.6	22.5	63.1	1.22	0.001–2.57	24.4	2.78	1.14–4.62	12.5
RTG006	6.50	1.64	0.48	0.08	0.99	0.24	34.5	43.3	45.9	0.78	0.002–1.65	36.7	1.91	0.51–4.15	17.5
RTG007	9.81	4.30	2.24	0.14	2.50	0.57	82.5	92.3	8.3	2.14	1.01–3.08	49.5	3.24	2.23–5.09	42.1
RTG008	0.85	1.75	0.75	0.01	1.60	0.33	29.9	43.6	42.8	1.07	0.03–2.02	36.4	2.28	0.97–4.25	20.8
RTG009	4.53	3.19	2.43	0.15	4.53	0.86	32.1	71.7	26.8	3.12	0.88–4.82	37.1	5.95	4.14–8.44	36.1
RTG010	2.45	1.77	1.46	0.09	3.36	0.53	23.4	51.4	41.2	2.35	0.99–3.35	45.5	6.48	3.90–9.18	13.4
RTG011	1.92	1.98	1.05	0.09	2.34	0.42	25.9	45.4	43.9	1.48	0.18–2.72	34.1	3.11	1.75–5.11	22.0
RTG012	2.26	3.17	2.42	0.15	3.90	0.78	42.2	75.6	12.3	2.14	1.24–3.25	53.7	5.99	4.25–7.69	34.1
RTG013	0.91	2.35	1.90	0.12	3.16	0.60	40.7	68.9	17.2	1.89	1.02–3.04	57.8	5.05	3.07–6.99	25.0
RTG014	5.11	3.00	3.13	0.18	4.33	0.96	55.0	85.9	11.3	2.94	1.30–4.59	42.4	5.54	4.05–7.49	46.3
RTG015	4.06	4.80	2.58	0.06	4.21	0.85	41.1	76.8	18.3	2.78	1.29–4.24	47.0	6.09	4.11–8.27	34.8
RTG016	5.33	4.94	2.08	0.13	2.65	0.57	65.6	83.6	12.5	2.06	0.87–3.03	53.7	3.80	2.26–6.36	33.8
RTG017	0.38	1.78	1.35	0.09	2.68	0.46	30.5	54.2	36.6	2.12	1.38–2.82	56.6	8.48	5.16–10.85	6.8
RTG018	3.56	1.95	2.58	0.06	3.60	0.78	54.8	82.1	13.0	2.66	1.43–3.85	57.4	5.68	3.37–8.51	29.6
RTG019	10.77	2.58	1.35	0.09	3.22	0.51	22.1	48.3	45.0	2.02	0.28–3.53	30.0	4.03	2.68–5.92	25.0
RTG020	7.32	1.08	1.43	0.09	4.82	0.66	11.4	43.5	45.4	1.98	0.51–3.59	29.4	6.67	5.31–8.26	25.2
RTG021	3.73	2.14	2.03	0.10	3.60	0.67	35.9	68.0	27.4	2.35	0.62–3.94	37.0	4.54	3.13–6.59	35.7

Note. 95% HPDCI = 95% highest posterior density credible interval.

Figure 1 compares the doses to blood during radiotherapy calculated using Institute of Cancer Research/Royal Marsden (ICR/RM) blood dose models 1 (EDD1) and 2 (EDD2) and the dicentric doses, estimated using Dose Estimate_v5.1 and the standard contaminated Poisson method, to the WB and PB. As shown, the relationship between WB dose and EDD2 gives a regression coefficient (± standard error) of 0.607 (±0.029), with an R² value of 0.88 and 95% confidence limits (CLs) of 0.84–0.94. The corresponding values for PB dose and EDD2 are 1.010 (±0.079) and 0.72 (95% CLs 0.60–0.84), respectively. An F test P value of 0.001 for the significance of the relationships and the 95% CLs indicate no substantial overlap. For EDD1, there was no significant linear relationship between the model dose and either whole- or partial-body dose, but the R² correlations for the plotted relationships were 0.04 and 0.03, respectively. As the models were only initial indications, equal weighting of each point was applied in this case.

FIG. 1.

Dose after the penultimate fraction to blood during radiotherapy, calculated using models 1 (EDD1) and 2 (EDD2), and whole- and partial-body cytogenetic doses, calculated using the standard contaminated Poisson methodology to separate exposed and unexposed fractions in partial-body exposures. WBD = whole-body dose; PBD = partial-body dose.

Partial-body dose estimates calculated by the standard contaminated Poisson method and the new Bayesian technique were compared using the average body doses. These were calculated as the product of the irradiated fraction and dose for the standard method and the sum of the product of the respective doses and fractions for the Bayesian technique. A Student’s t test on these normalized values showed no significant difference (P > 0.999) between doses calculated by the Bayesian and the standard method.

We have grouped the results by cancer type in Table 3, to determine whether there is a difference in dose response, calculated using standard methods for the dicentric data, then applying GLM ANOVA for this factor with post hoc testing, showed that the type of cancer had a significant effect on the WB and PB dose (P < 0.001). The ANOVA comparisons revealed that radiation treatment for some cancer types resulted in either significantly lower or higher WB or PB doses; these are presented in Table 4. No other significant differences were observed.

TABLE 4.

General Linear Model ANOVA Comparisons

	Cancer type
Calculated dose		Endometrial	Lung	Esophageal	Colon	Prostate
Whole body	Breasta	P < 0.001	P < 0.001	P < 0.001	ns	ns
	Prostatea	P < 0.001	P < 0.001	ns	P = 0.026
Partial body	Breasta	ns	P = 0.005	P = 0.013	ns	ns
	Lungb	ns	ns	ns	ns	P = 0.029
	Esophagealb	ns	ns	ns	ns	P = 0.037

Notes. These GLM ANOVA comparisons show that radiation treatment for some cancer types resulted in either significantly lower or higher whole-body or partial-body cytogenetic doses. ns = not significant.

^aCalculated dose significantly lower than for the other cancer types shown.

^bCalculated dose significantly higher than for the other cancer types shown.

DISCUSSION

Biomarkers of radiation exposure have been used for biological dose estimation for many years; in particular, the DCA has been in use since the mid-1960s (1). Biodosimetry methods have the potential to contribute to epidemiological studies of ionizing radiation effects (16–18). With the goal of improving the application of radiotherapy, the significance of predictive and prognostic biomarkers of response to radiation has also been demonstrated (19–21). In addition, some studies using cytogenetic biodosimetry assays have shown that they may be considered as a predictor of radiosensitivity to identify patients likely to develop acute/chronic adverse effects from radiation (22–24), although these studies have been small in scale and not prospectively validated. Gene expression analysis has shown possible potential as a marker of radiosensitivity (25–27) and as a sensitive biological marker for biological dosimetry (3, 5, 28, 29). Despite the development of modern techniques, the DCA remains the most specific and standardized method for biological dosimetry (30) and thus, it is the assay best suited to validate the gene expression technique for dose estimation.

RTGene was a feasibility study to develop and further validate the gene expression assay for biodosimetry with human blood samples exposed in vivo (8, 9) and included conventional biomarkers for additional validation. This has allowed dose estimates based on the dicentric assay to be calculated. As Table 1 shows, the radiotherapy schedules and doses for the patients are different, and ANOVA results indicate that the cancer site has a significant effect on the WB and PB dicentric dose estimates. When the cancer sites were compared further, significant differences were observed, with treatment for breast and prostate cancer resulting in significantly lower cytogenetic dose estimates than other groups. With breast and prostate irradiation, the high dose volumes are generally smaller than those in other tumor sites, resulting in lower WB and PB doses. Breast in particular, if treated with tangential fields only (as is the case in this study), spares the lung and heart, with most of the dose going through less vascular tissue within the breast. Lung and esophagus irradiation would invariably result in doses to highly vascular organs such as lung and heart and this is reflected in the DCA and blood dosimetric models applied here.

To our knowledge, there are currently no recommended methods to calculate the dose to circulating blood for radiotherapy. EDD1 and EDD2 are relatively simple blood dosimetry models, but the assumptions contained in the models are close to reality. The blood volume of an adult is considered to be approximately 6 liters, although this varies from individual to individual. The irradiation time for the model of 1 min was a typical value, as radiotherapy is planned to the requirements of each individual and the time will depend on these factors, the prescribed dose, the dose rate which the treatment machine can deliver and the type of delivery (static or rotating). It is difficult to give a range of irradiation times that encompasses all possibilities, however, appropriate values for the patients in the RTGene study are between 30 to 120 s. Modeling dose to blood during irradiation is a relatively small field, but a recently published study (31) modeled dose to circulating lymphocytes for patients receiving radiotherapy for malignant gliomas. A similar approach was taken to the models described here in that assumptions were made about blood flow and blood volume in the body, for example. Results reported elsewhere (31) indicate that after the total course of radiation treatment, most of the blood received >0.5 Gy, results which are supported by the two simple models (EDD1 and EDD2) and even the dicentric dose estimates in this study. However, the difficulty of not having whole- body, and specifically accurate whole-blood volumes, means that both that reported study (31) and the current work have limitations that are clearly acknowledged.

Dicentric doses, estimated using standard methods (1), have been compared to the calculated dose to blood derived using two newly developed ICR/RM dosimetric models. Dicentric chromosome aberrations are the result of misrepair of radiation-induced DNA double-strand breaks (1) with most being formed quickly, within 2 h postirradiation (32). The DCA evaluates damage in PHA-responsive T lymphocytes, predominantly the CD4⁺ and CD8⁺ subtypes. There are large uncertainties on the lifespan of lymphocytes, however, chromosomal damage after radio-therapy in CD4⁺ and CD8⁺ lymphocytes, which express naïve/memory markers (CD45RA⁺/RO⁺), has been studied. CD45RA⁺ cells have been shown to divide every 3.5 years and those expressing CD45RO⁺ every 22 weeks, on average (33). Generally, it is thought that approximately 80% of circulating lymphocytes survive for approximately 4 years (34, 35) and the biological half-life of dicentric chromosomes is approximately 3 years (36), albeit with some uncertainty. The patients in the RTGene study received radiation treatment over a time period of 7–37 days, blood samples were taken after irradiation, so that repopulation of lymphocytes during treatment would not be a major influence on the comparison of the cytogenetic and model doses.

Despite the limitations and uncertainties of the physical models, which are large and difficult to quantify fully (e.g., no knowledge of the blood volume in specific organs), the relationship between the cytogenetic and the model doses, which are independent of each other, is very promising, especially for EDD2, as shown in Fig. 1. This suggests that despite the models’ crude nature they may be useful. Both physical models have the capacity to be individualized further, which may mean that the uncertainties are reduced, and this initial success will allow further development to occur, e.g., to take account of lymph nodes in the radiation field. The comparison of cytogenetic and model doses reveals an apparent sensitivity to the choice of model, however, this may be related to the small sample size (20 patients).

Partial-body dose estimation, termed the contaminated Poisson method, was first proposed for dicentric data in the late 1960s (12) and is still one of the standard methods recommended for biological dosimetry (1). More recently, it has been suggested that Bayesian statistical analysis may be more suitable for dicentric data (37). A new Bayesian zero-inflated Poisson finite mixture method for estimating PB exposure has been developed with test data from simulated PB irradiations, where nonirradiated and ex vivo irradiated blood samples were mixed in different proportions (13). Cytogenetic data from the RTGene study has allowed the Bayesian zero-inflated Poisson finite mixture method to be used after in vivo irradiation and for a comparison of PB dose estimates calculated by this new approach and the standard contaminated Poisson technique. The Bayesian method has shown that the distribution of the radiation-induced damage at a cellular level can be expressed in terms of a gradient exposure, but the number of irradiated fractions is lower than the number of radiation procedures. In part this difference may be the result of the fractionated nature of the exposure, with a different subset of lymphocytes being irradiated during each fraction. However, the good agreement between the Bayesian and standard technique indicates this new method to calculate PB dose has the potential to provide additional information regarding dose estimates and irradiated fraction for biological dosimetry. In all situations where the DCA is used to estimate PB dose, the Bayesian method is more accurate when information about the priors is well documented. However, if cytogenetic triage dose assessment is used, for example during a mass casualty event where the objective is to place patients into broad dose categories, the more complex Bayesian approach may not be necessary. It is applicable when information about the priors is good and can more accurately characterize the dose, fractions and uncertainty, as all of these are included in the outcome.

In summary, the results from the RTGene study using a conventional biomarker, the DCA, indicate they can be used to validate future gene expression data. The comparisons are very encouraging between the cytogenetic dose estimates and: 1. blood dosimetric models; and 2. the new Bayesian method for gradient exposure. This will allow further development of the dosimetric models and demonstrates that the new Bayesian method can be applied after in vivo irradiation. Much more work is needed, but the next step will be further development and validation in a larger patient group. The RTGene partners will also explore the possibility of combining the cytogenetic, DNA damage and gene expression data to form a multi-assay panel of biomarkers to inform on individual radiation exposure and effects.