Discovering Clinical Biomarkers of Ionizing Radiation Exposure with Serum Proteomic Analysis

Decades of research have been dedicated to the discovery of inherent biomarkers of radiation exposure. Prior studies have highlighted the importance of this research in the epidemiologic field of occupational exposure (1), exposure assessments in cases of environmental or industrial accidents (2), astronaut exposure in manned space missions (3–5), and for early markers of clinical response to radiation therapy for cancer (6–8). Recently, fears of radiological terror incidents have renewed the search for rapid and simple methods to identify exposed individuals within large populations and to predict health effects (9, 10).

The long-established gold standard biomarker of radiation exposure, cytogenetic analysis of peripheral blood lymphocytes, requires complex analysis by skilled workers (5, 11, 12). A number of alternative minimally invasive strategies have been explored, including the measurement of specific cytokines and metabolites in tissues and body fluids (4), micronucleus assays (13), and germ line radiation–induced apoptosis rates (2). More recently, a need for comprehensive profiling in the field of biomarker discovery has been recognized (10, 14). Clinical studies to date have focused on genomic profiling of circulating peripheral blood lymphocytes using microarray technology (15–17).

Although the ability to distinguish cancers from the profile of circulating protein constituents of serum was first shown in 1978 (18), only recent technological advances have led to robust techniques that permit high-throughput and comprehensive analysis of serum proteins (19). These techniques, including surface-enhanced laser desorption and ionization time of flight (SELDI-TOF) mass spectrometry (MS; ref. 20) and complementary mass spectroscopic fractionation (21), enable the study of global expression changes following pathologic events and physiologic stresses. As such, they hold promise as a comprehensive technique for biomarker discovery. Recently, it has been shown that most of the low-molecular-weight molecules that underpin the SELDI-TOF MS profiles exist bound to larger highly abundant proteins, such as albumin (22).

Thus, we sought to explore the merit of proteomic profiling strategies in patients with cancer before and during radiotherapy in an effort to discover clinical biomarkers of radiation exposure. We also sought to further characterize the information content of the spectral components using an albumin enrichment strategy of the same serum samples used in the high-throughput discovery series. There are three hypotheses tested in this study: First, proteomic profiling methods can discover a discriminating set of low-molecular-weight biomarkers that, in turn, can be used as a pattern-based diagnostic of human exposure to ionizing radiation. Second, a serum proteomic pattern diagnostic can discriminate a low dose-volume radiation exposure from a high dose-volume exposure. Last, MS identification of the same information archive that underpins the observed MS peaks will reveal important clues to the identity and nature of the discriminatory entities.

Specimen collection. Sixty-eight patients with a diagnosis of cancer provided informed consent for enrollment on one of two Institutional Review Board–approved studies permitting the collection of serum immediately before and during a course of radiation therapy between 2001 and 2003. The interval between the first and second sample collection varied according to patient preference, with the second sample typically acquired in the latter part of the course of treatment. Approximately 8 mL of blood was drawn by venipuncture and placed on ice. The samples were centrifuged within 2 hours of collection, and serum was aliquoted into Eppendorf tubes and stored at −80°C.

Specimen processing—SELDI-TOF. Aliquots of fresh frozen sera were sent to the National Cancer Institute (NCI)-Food and Drug Administration Clinical Proteomics laboratory. The pretreatment and posttreatment samples all had been collected and handled in an exact identical manner, not previously thawed to minimize any potential for collection and sample handling bias. Samples were thawed and 10 μL aliquots were obtained and used immediately for high-resolution SELDI-TOF analysis. Ciphergen (Fremont, CA) immobilized metal ion-affinity chromatography 3 protein arrays were used for this study. All steps were carried out on a Tecan Genesis 2000 robotic processor (Tecan, Research Triangle Park, NC) equipped with a 96-position Temo unit to reduce any potential for bias due to variation in sample preparation. The Ciphergen Bioprocessor unit was used to hold the protein arrays in place for all steps in the process. The patient-matched preexposure and postexposure samples were randomized and comingled on the same chip surface to reduce any bias that may occur due to run order and daily spectrometer fluctuations. The surfaces were activated with two applications of 100 μL per spot of 50 mmol/L nickel sulfate (Sigma, St. Louis, MO) incubated for 5 minutes per application. After aspiration, the surfaces were washed twice with 100 μL of deionized water. Dulbecco's PBS without calcium and magnesium chloride (Invitrogen, Grand Island, NY) with 0.1% Triton X-100 (Sigma) was added to each spot with a volume of 100 μL per spot and incubated for 5 minutes, aspirated, and discarded. This was repeated a second time. The chip surfaces were then dried using vacuum aspiration to remove any excess liquid. A total of 5 μL of undiluted serum was added to each spot and incubated for 20 minutes. The wash steps were carried out on the Tecan processor as follows: 100 μL of PBS was added and mixed by aspiration and dispensing. This was done for three cycles. Sterile deionized water was added and mixed by aspiration and dispensing. Remaining liquid was removed by vacuum aspiration and the chips were allowed to dry completely before the bioprocessor gasket was removed. Once removed, 1 μL matrix consisting of 12 mg/mL cinnamic acid (Fluka, St. Louis, MO) in 50% acetonitrile (Sigma) and 0.5% trifluoracetic acid (Sigma) was applied by the Tecan processor using a fixed pipette tip and a serpentine path to each spot on the arrays. The arrays were then dried at room temperature for 5 minutes and the matrix was reapplied. After the chips were completely dried, they were read on an ABI Q-Star mass spectrometer equipped with a Ciphergen PCI 1000 Protein Chip Interface as previously described (23).

Data analysis—SELDI-TOF. High-resolution MS data were derived by the unexposed and exposed blood draws from 68 cancer patients for an initial total of 136 spectra. Regions of spectra known to be dominated by matrix effect (<1,000 Da) or terminal measurement effects (>11,000 Da) were excluded a priori. Quality assurance/quality control (QA/QC) analyses resulted in the elimination of one spectrum from the unexposed group due to low average amplitude and total ion counts. The QA/QC tools and procedures have been previously described (23). Quartile values were computed for each spectra. Initial aggregate pattern analyses using visual data mining techniques showed interesting groupings involving the most abundant ions (third quartile or greater). Spectral data below this threshold were excluded. High-resolution mass spectral data were then binned as a function of the nominal mass values, while generating a variety of statistical measures for the captured elements (mass and amplitude values), including, means, SDs, variances, and sums.

An amplitude-based concept hierarchy was constructed for each spectra based on percentage contributions. Normalization was achieved through transforming the summed amplitude values through the concept hierarchy mapping scheme (24). Aggregate pattern analyses were done by three-dimensional visual data mining techniques to explore for regions of interest. An example of this is illustrated in Fig. 1A. Given the heterogeneity present in the data due to large variety of diseases/cancers (18), as well as a wide range of both anatomic sites (9) and degree of radiation received, the visual data mining activities aided greatly in the identification of interesting ion patterns for additional pursuit (25). Custom T-SQL stored procedures were written to further aid in feature selection through the implementation of a minimax algorithmic procedure to maximize the differences between candidate ion groups while minimizing their coefficient of variation.

The unexposed (n = 67) and exposed (n = 68) case-controlled spectra were then analyzed by a variety of classification methods derived from Clementine, version 8.5 (SPSS, Chicago, IL). The three methods of classification tested were as follows: (a) C5.0 decision tree, (b) a hybrid classifier constructed from a multilayer perceptron neural network using an exhaustive prune strategy linked into a decision tree, and (c) a hybrid classifier constructed from a Classification and Regression Tree (CART) coupled to a C5.0 decision tree. Cross-validation was used to assess the classification error rates of the various models. As previously stated, due to the inherent heterogeneity in this data set and relatively small size, it was felt a validation strategy using a rotation method would better address biological relevance as well as credence in classifier accuracies. Construction of larger training data sets permitted more accurate predictive classification for this study set. The data set was partitioned into N equal-sized subgroups and models were fitted N times. N was chosen to be 10. Each model used N − 1 of the subgroups for training, and then applied the resulting model to the remaining subgroup. Accuracy values were averaged over the N holdout subgroups.

These computational techniques were also used to construct supervised predictive classifiers to discriminate high (n = 32) versus low (n = 36) dose-volume radiation exposure from the exposed group. The method for dividing exposure dose-volumes into high and low was done by a statistical stratification method using quartile analysis of dose-volume measurements coupled to aggregation of both cancer type and site of radiation. This revealed a natural cut in the dose-volume measurements at 190 Gy/m²; therefore, measurements >190 Gy/m² were recorded as a high exposure and the remainder as low. In the low dose-volume classification, the mean dose-volume was 126 Gy/m² (range 44-190); in the high dose-volume classification, the mean dose-volume was 350 Gy/m² (range 200-1,050). Attention to disease/cancer type and anatomic site of radiation was taken to avoid bias. For instance, there are 18 prostate cancer patients in the low dose-volume group, and 15 in the high dose-volume group. In this study, the units for dose-volume measurements are Gy × fraction body volume exposed to >50% radiation dose. This definition was chosen a priori and assumes that both radiation dose and irradiated volume contribute equivalently to changes in the serum proteome.

Deliberate steps were taken to produce the most general predictive models and to avoid overtraining. To this end, strategies used included the selection of overtraining prevention options, the use of global pruning with a pruning severity of 75%, requiring minimum records per child branch to be greater than or equal to a threshold value (i.e., 4), and winnowing of attributes for a more parsimonious solution. To enhance predictive accuracy of the decision tree models, boosting was used.

Visual data mining techniques were also used in an iterative fashion as candidate ions were meticulously winnowed. In many cases, candidate ion sets, whether they were considered for promotion or elimination, were further analyzed by high-resolution visual data mining methods. These techniques use a high-end PC graphics card (NVIDIA Quardro FX 3000, NVIDIA Corp., Sunnyvale, CA) coupled to 9-megapixel high-resolution IBM T221 computer graphics monitor (IBM, Armonk, NY). Thus, high-resolution mining of extremely complex data sets may be addressed. Figure 1B illustrates aspects of a three-dimensional visual data mining exercise involving an intermediate set of candidate ions.

Both a high-end computer workstation and server were used to accomplish the aforementioned computational tasks. These are located at the NCI, Center for Cancer Research, Clinical Proteomics Reference Laboratory (CPRL, Gaithersburg, MD). The server consists of dual processor AMD Opertron CPU module (AMD Corporation, Sunnyvale, CA) housed on a Newisys motherboard (Newisys Corp, Austin, TX) fitted with a total of 16 GB of main memory. Server disc capacity consists of an Adaptec RAID housing assembly containing twelve 146 GB IBM disc drives in a RAID 5 configuration, allowing one disc for parity checking and one disc for a hot spare. Thus, the effective storage capacity for this server at the CPRL is ∼1.3 TB. Following CPRL specification, the server system was custom built by Colfax International (Sunnyvale, CA).

The server computer operating system is the Windows 2003 Advanced Server (Microsoft Corp, Redmond, WA). Database activities are done through the use of the Microsoft SQL Server 2000 Enterprise Edition. Computer workstations at the CPRL consist of one to two Xeon processors (Intel Corporation, Santa Clara, CA), with 2 to 4 GB of memory, running the Windows XP Professional operating system. Visual data mining activities are done on workstations outfitted with high-end graphics cards and displays. Data intensive computational tasks are off-loaded to the server.

MS-MS protein identification of pooled sera. From the original study set, a subset of 25 patients was randomly selected and their serum was pooled for MS-based sequencing and peptide identification from samples obtained before and during radiation therapy.

Albumin and associated low-molecular-weight bound peptide purification was done in the following fashion and has been further described previously (22). Forty microliters of preradiation or postradiation treatment pooled serum (∼5 mg protein) were diluted to 200 μL with an equilibration buffer (Millipore) and run twice through a Montage (Millipore, Billerica, MA) albumin-specific affinity column. The bound protein was washed thoroughly and eluted from the column by equilibrating with 70% ACN/30% H₂O/0.2% trifluoroacetic acid for 30 minutes, followed by a slow spin-through of the elution mixture. The eluate was lyophilized to <10 μL in a HetoVac rotofor (CT 110) and reconstituted in a 95% H₂O/5% ACN/0.1% formic acid buffer (buffer A). Samples were sometimes desalted with a ZipTip cleanup or with Vivaspin 500 centrifugal membranes and always reconstituted in a 1:1 mixture of water and SDS sample buffer (20 μL total volume).

Analysis then proceeded with one-dimensional gel separation and digestion. Twenty microliters of sample in SDS sample buffer (40 μL of original serum) were boiled for 5 minutes at 95°C and run on a one-dimensional precast gel (4-20% Tris-glycine or 4-12% bis-Tris) to separate albumin from the low abundant proteins/peptides/fragments of interest. The gel was stained with Coomassie blue for 1 hour and destained overnight in 30% methanol/10% acetic acid solution. The entire lane containing stage-specific serum proteins was excised from the gel and finely sliced into very small molecular weight regions (∼40 slices/lane). Gel bands were reduced and alkylated with 10 mmol/L DTT and 55 mmol/L iodoacetamide, incubated at 4°C for 1 hour in porcine trypsin (20 ng/μL; Promega) and allowed to digest overnight at 37°C in 25 mmol/L NH₄HCO₃. The following morning, peptides were extracted from the gel with a 70% ACN/5% formic acid solution.

Samples were lyophilized to near dryness and reconstituted in 6.5 μL of buffer A for mass spectrometric analysis. Microcapillary reverse phase liquid chromatography (LC)/MS/MS analysis was done with Dionex's LC Packings LC system (Dionex, Sunnyvale, CA) coupled online to a ThermoFinnigan LCQ Classic ion trap mass spectrometer (ThermoFinnigan, San Jose, CA) with a nanospray source. Reverse phase separations were done with an in-house, slurry-packed capillary column. The C₁₈ silica-bonded column is 75 μm ID, 360 μm OD, 10-cm-long fused silica packed with 5 μm beads with 300 Å pores (Vydac, Hesperia, CA). A μ-precolumn PepMap C₁₈ cartridge (Dionex) acts as a desalting column. Sample is injected in microliter pick-up mode and washed with buffer A for 5 minutes before a linear gradient elution with buffer B (95% ACN/5% H₂O/0.1% formic acid) up to 85% over 95 minutes at a flow rate of 200 nL/min. Full MS scans are followed by three MS/MS scans of the most abundant peptide ions (in a data-dependent mode) and collision-induced dissociation is done at a collision energy of 38%.

Data analysis was done by searching MS/MS spectra against the European Bioinformatics Institute of the nonredundant proteome set of Swiss-Prot, TrEMBL, and Ensembl entries through the Sequest Bioworks Browser (ThermoFinnigan). Peptides were considered legitimate hits after filtering the correlation scores and manual inspection of the MS/MS data (26–30).

Accepted peptide hits are required to have an X_corr ranking = 1 relative to all other peptides in the database. The albumin extraction, gel electrophoresis, protein digestion/extraction, and LC/MS/MS analysis was repeated in six distinct trials (three for preradiation treatment and three for postradiation treatment)—each time yielding diminishing returns of new identifications for low abundance peptide hits. Repetitive sequencing of identical peptides in multiple trials further validates our experimental procedure—both within and between radiation treatment stages.

Patient characteristics for the SELDI and pooled MS study sets are described in Table 1. Due to referral patterns in our clinic, a disproportionately high number of patients with prostate cancer and/or receiving pelvic radiotherapy are represented in this study. Overall, our study population consists of a wide range of diagnoses, body site, elapse time interval, volume, and dose of radiotherapy. All patients were exposed to ionizing radiation between the first and second blood sample acquisitions.

SELDI-TOF analysis. A summary of the classification techniques used in this study, the data sets to which they were applied, and results are contained in Tables 2 and 3. Table 2 addresses the first stated hypothesis that proteomic profiles can distinguish unexposed from radiation-exposed patient serum samples. This involves the analyses of 135 spectra, obtained by randomized and comingled placement on the SELDI chip surfaces, with the unexposed group having an n = 67, and the exposed group with an n = 68. A variety of classification methods were used for consensus and concordance purposes. As is seen by the reported sensitivity and specificity values, all classification methods did well. Different sets of key ions were discovered and used by each method. Ions common to more than one pattern-based diagnostic include ions 7,264, 2,551, 7,554, 2,698, and 10,157.

Table 3 addresses the second hypothesis that proteomic profiles distinguish high from low dose-volume radiation exposures. These analyses involve only the radiation-exposed group of samples (n = 68; high dose-volume = 32, low = 36). To facilitate accord, a variety of classification methods were again used to assess the overall predicative accuracy of a proteomic pattern-based diagnostic. The reported predictive results show good performance in distinguishing the degree of ionizing radiation exposure (low or high dose-volume). Common discriminating ions for these models are 7,691 and 7,889.

MS-MS protein identification of pooled sera. A total of 182 protein fragments were identified in this study, 82 of which were uniquely identified in pooled serum before patients were exposed to ionizing radiation, and 23 of which were uniquely identified in pooled serum while patients were being exposed to ionizing radiation (Table 4—ions listed in alphabetical order). The remaining 77 protein fragments were identified both before and during radiation exposure (data not shown).

Following radiation accidents or intentional exposure, it is critically important to estimate rapidly the level of exposure in persons at risk. In the moderate dose range of 1 to 10 Gy whole body exposure, gold standard lymphocyte assays have restricted value as this cell population is depleted as a function of both dose and time after exposure (31). Microarray studies of exposed cell cultures have led to the discovery of GADD45 and CDKN1a as potential biomarkers of radiation exposure. A linear gene induction was found between 2 and 50 cGy, supporting the idea that it may be possible to develop gene induction profiles that are useful markers of human radiation exposure or dose (32). However, this approach is more challenging for clinical samples given the low abundance of RNA in isolated peripheral blood lymphocytes from venipuncture (17). In the clinically critical moderate exposure dose range where lymphocytes are depleted, alternative biomarkers need to be considered and developed further.

In this study, we opted to analyze serum samples in an aggregate manner from a population of patients with known exposure to a wide range of therapeutic doses of ionization radiation, over varying time intervals, and at different body sites and volumes of irradiation. This study design was chosen with the overarching goal of identifying serum protein markers of radiation exposure that are independent of volume and body site of exposure at dose levels that may have relevance to moderate dose whole body exposures.

Challenges unique to this particular study include a large degree of inherent heterogeneity due to two factors. First, the 68 patients exhibit 18 different medical diagnoses, which are predominately different forms of cancer. Second, there are nine different anatomic sites of radiation exposure/delivery. Nonetheless, deliberate computational strategies were used to attempt to model and understand the innate heterogeneity, while also maximizing robustness of classification methods. This was accomplished mainly through visual data mining activities that allowed for aggregate pattern analysis and discovery of key ion groupings, including an initial observation indicating a discriminatory ability in the most abundant ions (likely influenced by injury response and immune mediated events), as well as the use of n-fold cross-validation strategies for the testing of classifier error statistics. Novel visualization techniques facilitated aggregate spectra pattern analyses as well as reduced data set strategies, thus enabling study of cohort heterogeneity.

As a secondary objective, we provide preliminary evidence of a dose-response relationship to the proteomic changes induced with radiation exposure. This early data must be interpreted with caution given the small sample size and lack of a patient control study design for this end point. Future studies in larger groups of patients are warranted to determine the linearity of dose-volume effects and the time course of serum proteome changes after exposure.

To further characterize the nature and identify of the SELDI-TOF profiles, we used a new method of analyzing the albumin-bound peptide archive (22). The molecules that comprise the SELDI portraits have been recently shown to emanate from carrier protein binding, which serves to amplify and enrich low-molecular-weight peptides that would otherwise be efficiently cleared by glomerular filtration (22). Two pooled sera sets, representing a randomly obtained subset from patients before and during exposure, were used for this study. The lower abundance proteins uniquely identified in pooled serum may be secreted or shed by cells as a result of signaling, necrosis, and/or apoptosis of cells in the high-dose region. In fact, one of the entities identified by this approach and that may be a component of the SELDI-TOF spectra is a fragment of interleukin (IL)-6 that was uniquely measured only in the postexposure pools. This finding is consistent with known cellular effects of ionizing radiation at the transcriptional level. A major transcription factor known to be activated by ionizing radiation is nuclear transcription factor-κB (NF-κB; refs. 33, 34). Genes induced by NF-κB following irradiation include intercellular adhesion molecule-1, galectin-3, tumor necrosis factor-α (TNF-α), IL-1β, and IL-6. In a study by Fedorocko et al. (35), mice exposed to a single 9 Gy dose of radiation were found to have elevated serum levels of IL-6 and TNF-α peaking at 5 days following exposure. This is further supported by a clinical study of patients receiving radiotherapy to their liver, where IL-6 was found to be variably elevated in serum at 24 and 48 hours postexposure (7).

The SELDI-TOF method described in this study has been optimized and adapted to a robot for the purpose of reproducibility and enabling of a high-throughput test. This gives precision of pipetting, reproducibility of pipette dispensing position, and ensures accurate timing. In addition, the use of barcodes on both samples and Ciphergen arrays allows for traceability of sample location. The time for processing a serum sample is ∼2.5 hours. Finally, once a classifier is formulated and validated, the computational or wall time required to classify an unknown spectrum is a few milliseconds.

The low-molecular-weight serum proteome is a newly discovered information archive reflective of metabolism and pathologic events. In the recent past, it has been successfully exploited for early diagnosis of ovarian cancer as well as diagnosis of other cancers. This study adapted these established principles for the successful discrimination of both radiation exposure as well as the degree of radiation exposure. Applications of such a high-throughput blood test may include personalized proteomic measures of clinical response to radiation therapy and rapid response measures for populations exposed to moderate doses of radiation. Further refinement of the methods reported in this study may lead to personalized molecular medicine approaches directly applicable to adjuvant cancer care (36). Instead of categorical radiation dosing techniques, a strategy using a serum proteomic test, with the aim to better quantify the biological effect of serial exposures, may better avoid toxicity while maximizing therapeutic efficacy.

The composition of proteins in serum changes with ionizing radiation exposure. Proteomic analysis for the discovery of clinical biomarkers of radiation exposure warrants further study.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.