Since the early application of microarrays to radiation responses in human WBC first suggested the potential of gene expression for biodosimetry (13
), studies have continued to support this application. Gene expression signatures in peripheral WBC can predict radiation dose using both ex vivo
and in vivo
). In the present study we tested the ability of gene expression profiles to predict radiation doses received in vivo
in a heterogeneous population of TBI patients and to compare the predictive ability of signatures derived from in vivo
versus ex vivo
exposures. Classifiers built from in vivo
responding genes were able to predict radiation exposure levels of the in vivo
samples with 96% accuracy. Classifiers built from the gene signature identified previously in ex vivo
) were similarly effective, predicting the dose to TBI samples with 98% accuracy, indicating that the ex vivo
signature is robust across different sample sets and radiation models.
A potential concern in the use of cancer patients for these studies is the possibility that their clinical conditions or prior treatments could cause alterations in preirradiation gene expression that could confound radiation signatures. The heterogeneity of diagnoses in this study should buffer somewhat against selecting gene expression signatures that would apply only to a subset of individuals (those either predisposed to or having a specific disease). We added to our analysis 14 control samples from healthy donors. Adding these samples did not dramatically affect the predictive ability of either the TBI or ex vivo derived classifiers. Using the 3-NN algorithm, the signature derived from the TBI data predicted the dose category of 94% of the samples (including the healthy controls) correctly, and the ex vivo derived signature predicted 98% of the samples correctly (). Although we cannot directly test samples from healthy people irradiated in vivo, this study demonstrates that there is no appreciable confounding between unirradiated healthy people and cancer patients.
Performance of 3-NN Classifiers on Postirradiation Samples Only
Comparing the current findings with our first analysis of in vivo
radiation response (16
), only a subset of the genes responding significantly in the current study were identified in that earlier study. The first study surveyed only 6485 genes, used different protocols, and included microarray data from only one patient. Studies of radiation responses in the skin of radiotherapy patients have found large variations in the specific genes responding in individual patients, but in that case gene set enrichment analysis indicated conserved functions among the responding genes (24
Gene ontology analysis of our earlier in vivo
study had identified immune and inflammatory responses as the most significant processes represented among the radiation-responsive genes, a finding consistent with the gene ontology analysis of the present study (). After updating the gene annotations using the Array Clone Information Database (26
) we analyzed the older TBI gene set using PANTHER (23
). Despite the smaller number of genes in the earlier study, we found similar enrichment of immune- and cytokine-related responses in the two studies, consistent with the profound immunomodulatory effects of TBI observed in hematopoietic stem cell transplantation (27
These immune-related categories were also enriched in the analysis of data sets from our ex vivo
). In contrast to the TBI study, the ex vivo
response was dominated by genes associated with apoptosis, cell cycle and the p53 pathway, rather than with immune functions, consistent with the responses in cell lines irradiated in vitro
). Because our TBI study spanned only 0–3.75 Gy, while the ex vivo
study was based on genes with dose-dependent responses up to 8 Gy, the use of higher doses may also contribute to relative differences in gene ontology enrichment. It is important to note that the p53-related functions were also represented among the significant TBI responses; they were just not the most highly enriched categories. The broad patterns of gene categories responding in vivo
and ex vivo
highlight the similarity between whole organism stress responses and cellular responses in vitro
. They suggest a “core” of gene expression responses, many of which are conserved ex vivo
, with additional responses and layers of signaling present in the whole organism. The greater magnitude of the in vivo
response was reflected in the larger numbers of genes responding significantly in vivo
in all the categories examined. This supports our earlier suggestion that ex vivo
selection of radiation signature genes may actually help reduce the potential for confounding by excluding a number of immune response genes that respond robustly to radiation in vivo
but that may represent more generalized stress responses not specific for radiation exposure.
While it should be kept in mind that the response of a signature is likely to be less prone to confounding than the response of an individual gene from that signature (29
), extensive testing for potential confounding of dosimetric signatures will ultimately be needed. Variation is known to exist in the peripheral blood gene expression profiles of healthy individuals (30
), and subtle variations may be attributed to factors such as body mass index or smoking status (32
). Acute or chronic disease states including infection (33
), sepsis (34
), immune or inflammatory conditions (36
), stroke (37
), cancer (38
) and transplant rejection (39
) can also alter the peripheral blood gene expression profile.
Injuries such as burns or physical trauma also produce specific gene expression signatures in peripheral blood, and these would be especially relevant in the context of a radiological or nuclear event. To get some preliminary gauge of the amount of overlap that might be expected in responses to such trauma and the response to ionizing radiation exposure, we compared our ex vivo
radiation signature with the genes reported to respond significantly at 1 day after treatment in a recent study comparing gene expression changes in the blood of mice undergoing traumatic blood loss or scald burns over 25% of their bodies (40
). Of 4850 genes reported as responding in the burn model at 1 day after injury, 9 genes (BTG3, CDKN1A, GADD45A, LY9, MDM2, MGAT3, PLK2, PTP4A1 and SESN1
) changed in the same direction in our radiation signature. A further 6 genes from our signature (ARHGEF3, DDB2, IL21R, LIG1, PHPT1
) were upregulated by radiation but downregulated in the burn study. In the trauma-hemorrhage study 1164 genes responded 1 day after blood loss. Of these, five genes from our ex vivo
signature (BTG3, CCNG1, CDKN1A, DCP1B
) were upregulated both by blood loss and by radiation exposure, and one gene (DDB2
) was downregulated by blood loss but upregulated by radiation. The results of this comparison illustrate some of the complexities of gene expression responses to different stresses and suggest that although some genes in our current radiation signature will respond similarly to traumatic injuries, the majority of the signature is not likely to overlap with the response to a different specific stressor, and most importantly, the overall patterns of response are likely to be distinct. Further direct testing, not only comparisons with published studies, will obviously be required to fully assess the extent to which conditions of trauma, disease or lifestyle factors may affect the actual estimation of radiation dose using gene expression.
We also compared our TBI results with the predictive signature reported by Dressman et al.
). These authors identified 25 probe sequences representing 18 human genes that predicted radiation exposure between pre-TBI and post-TBI samples with 90% accuracy. Of these genes, 16 were significantly regulated in our current TBI study and 13 were included in our ex vivo
signature. Although we had focused on prediction between three sample classes (unexposed, 1.25 Gy or 3.75 Gy), it was of interest to compare the ability of our approach to predict samples as either pre- or postexposure, either after the first fraction as reported (17
) or for any unexposed versus any postexposed. Focusing on the first dose alone also avoids complications in interpretation that may arise due to the three-fraction nature of the 3.75-Gy total-body dose. Classifiers built either from our TBI data or using the previously defined ex vivo
signature predicted the samples (including the healthy controls) as exposed or unexposed with 100% accuracy by all methods tested.
The separation of control and in vivo
irradiated samples using the ex vivo
signature is illustrated by multi-dimensional scaling (MDS) in . MDS is a visualization tool that preserves pairwise similarities between samples while allowing graphical representation of high-dimensional data in two or three dimensions (41
). Each point represents a sample and the first three principal components of the gene expression matrix are used as axes. The distance between any two points represents the overall similarity of the expression levels of all genes in the signature. The similarity between controls (from healthy donors or TBI patients) further illustrates that the disease status of TBI patients does not detrimentally affect the baseline expression of the genes in our ex vivo
signature. The 6-h control and 2-Gy samples or 24-h control and 5-Gy samples from our previous ex vivo
) are included in the MDS for comparison. A similar separation is seen between control and irradiated samples both in vivo
and ex vivo
, although a lateral shift is also evident between in vivo
and ex vivo
samples in the MDS (). This lateral shift suggests that ex vivo
incubation did affect expression of some genes in the signature independent of radiation exposure, as previously reported for increasing time in culture (14
). The“time in culture” effect did not interfere with the prediction of radiation dose in either study, however. Previous studies have also demonstrated some gene expression changes in WBC during ex vivo
FIG. 2 MDS plots of pre- and post-TBI samples using the ex vivo gene signature. Each point represents an individual sample; the axes represent the first three principal components of the MDS. Dark blue: patients prior to TBI; cyan: healthy unirradiated donors (more ...)
Because our previous signatures have all incorporated information from more than the 25 probe sets used in ref. (17)
, we were also interested to see if we could achieve good classification of samples using smaller gene sets. To this end, we applied a support vector machine recursive feature elimination algorithm in BRB-Array-Tools (20
) to select 25 features for predicting the status of the unexposed or post-first-fraction samples, analogous to the Dressman study. This smaller feature set (Supplementary Table 6
) also yielded 100% correct prediction of exposure status.
Inter- and intraindividual variations in gene expression do occur among healthy individuals (31
), and some genes respond to radiation with great variability in immortalized blood cells derived from different donors (44
), suggesting a potential problem for accurate prediction of radiation exposure by gene expression profiling. In practice, however, we find that appropriate signature selection can provide accurate prediction of both exposure status and dose. Using a signature of multiple genes also “buffers” the predictive ability by reducing the contribution of individual variations that may occur in a single gene. This has been reported with other biodosimetry approaches, where combining measurement of up to three serum proteins (45
) or adding C-reactive protein measurements to time-dependent blood cell counts (29
) improved dose prediction or identification of irradiated samples and provided robustness against confounding. Nonetheless, more extensive testing on large heterogeneous populations will still be needed prior to final implementation of any gene expression signature for radiation biodosimetry.
The performance of our ex vivo signature in predicting dose classes of samples from TBI patients irradiated in vivo further supports the robust nature of this signature, because it performs well not only under different irradiation conditions but also when applied to a completely different population with different health status from that in which it was selected. This signature predicted exposure with 100% accuracy and dose level with up to 98% accuracy in a heterogeneous human population undergoing TBI with variable gender, age and prior treatment history without the need for a pre-exposure control. These findings strongly support the utility of further developing this approach to provide biodosimetry in the event of a deliberate or accidental radiological or nuclear event.