Raman spectroscopy detects deterioration in biomechanical properties of bone in a glucocorticoid-treated mouse model of rheumatoid arthritis

Arthritis and other rheumatic conditions represent a public health burden that is both significant and growing. In 2008, it was estimated that more than 21% of U.S. adults (46.4 million persons) were diagnosed with some type of arthritis, with rheumatoid arthritis affecting 1.3 million patients.¹

Glucocorticoids are among the most extensively used anti-inflammatory agents in the management of the painful symptoms of rheumatoid arthritis. However, since the glucocorticoid (GC) receptor is expressed in almost every cell in the body, glucocorticoids have many effects, including potentially harmful ones on bone metabolism. These adverse effects include inhibition of bone formation and an enhancement of bone resorption, often leading to osteoporosis, and ultimately an increased risk of fracture. Currently, extended exposure to glucocorticoids is the leading cause of physician-induced osteoporosis, leaving patients susceptible to fractures at reported rates of 30% to 50%.² It has also been shown that GC excess in mice affects bone metabolism in a similar fashion, leading to significant osteoporosis that mimics clinical symptoms in human patients.^2- 4

To reverse GC-induced osteoporosis, several classes of drugs have been developed, including antiresorptive drugs of the bisphosphonate family^5- 7 and anti-tumor necrosis factor (TNF) drugs.⁸ Clinical trials of these drugs typically report increases in bone mineral density (BMD), as measured by dual energy x-ray absorptiometry (DXA) or computed tomography, as the primary outcome of drug efficacy. However, this increase in BMD is not predictive of fracture risk in the nearly 10% of treated patients that experience a fracture.^9- 10 Therefore, new diagnostic and monitoring technologies for detection of bone changes that better correlate with fracture risk could potentially be more definitive than BMD in establishing the efficacy of new generations of these drugs.

Vibrational spectroscopic techniques, including Raman spectroscopy, have been used extensively to study bone and other mineralized tissue.^11- 12 In particular, the bone diseases postmenopausal osteoporosis,¹³ osteogenesis imperfecta,¹⁴ and early-onset osteoarthritis¹⁵ all induce perturbations to cortical bone that can be detected by Raman spectroscopy. In this communication, we report the results of the first Raman spectroscopy study of cortical bone in a transgenic mouse model of rheumatoid arthritis that constitutively overexpresses the inflammatory molecule human tumor necrosis factor-alpha (hTNF-α), which was developed by Keffer et al. ¹⁶ Because of its chronic progressive nature, 100% incidence, and simple etiology, the hTNF-α transgenic (hTNF-Tg) mouse is an excellent model to investigate the pathogenesis and treatment of rheumatoid arthritis and side effects such as GC-induced osteoporosis.

Five month old hTNF-Tg mice (3647 line originally obtained from Dr. G. Kollias and maintained as heterozygotes in a CBA × C57Bl6 background)¹⁷ and wild-type (WT) littermates (C57Bl6 strain) were randomized into placebo (PL) or GC treatment groups. Placebo pellets were implanted subcutaneously in group 1 (WT PL, N = 4) and group 3 (hTNF-Tg PL, N = 3) mice, while group 2 (WT GC, N = 4) and group 4 (hTNF-Tg GC, N = 4) were implanted with a slow-release pellet (5 mg/kg/day) of the glucocorticoid Prednisolone (Innovative Research of America, Sarasota, Florida) as described by Lane et al. ¹⁸ Following a 6-week treatment period, the mice were sacrificed and the left tibiae were excised and evaluated by microcomputed tomography (microCT), Raman spectroscopy, and biomechanical testing. The tibiae were stored in phosphate buffered saline between measurements. A sample size of N = 4 mice/group was chosen for this preliminary study; however, one mouse from group 3 was lost due to premature death.

MicroCT measurements were performed on a Scanco VivaCT 40 (Scanco Medical AG, Bassersdorf, Switzerland) with a 10.5-μm isotropic resolution. In addition to the whole BMD, 4 cortical parameters related to torsional biomechanics (cortical BMD, cortical thickness, polar moment of inertia, and bone area) were measured. The whole BMD was measured in the proximal half of the tibia. The cortical BMD and thickness were measured over 100 axial slices from the center of the diaphysis. Each parameter was calculated from three-dimensional reconstructions of the microCT scans generated by the provided complete image solution software (Scanco Medical, Bassersdorf, Switzerland).

Raman spectra were acquired from the medial side of the proximal, distal, and mid-diaphysis regions of the excised tibiae with an exposure time of 300 s per region. These measurements were performed on a locally constructed Raman spectroscopy system that utilized an 830-nm semiconductor laser (Model PI-ECL-830-500-FS, Process Instruments Inc., Salt Lake City, Utah) for excitation. The laser light was filtered with a bandpass filter (Chroma Technology Corp., Bellows Falls, Vermont) and focused to a 1.5-mm (1/e² full-width) spot on the surface of the bone. The total power delivered to the specimen was approximately 80 mW and the illumination numerical aperture was 0.05.

In order to collect the Raman scattered light, the illuminated surface region was imaged onto the face of an optical fiber bundle. The fiber bundle consisted of 61 multimode optical fibers with 100/120-μm core/cladding diameters in a circular array at the collection end mapped to a linear array at the delivery end. Due to the system magnification, the diameter of the fiber bundle image was 1.5 mm and each fiber viewed a 285-μm diameter spot at the surface of the bone. Following collection, the scattered light was delivered to a spectrograph (Holospec™ VPT System, Kaiser Optical Systems Inc., Ann Arbor, Michigan) where it was dispersed onto a 1024 × 256 array back-illuminated deep-depletion charge-coupled device (CCD) camera (Model DU 420-BR-DD, Andor Technology, Belfast, Northern Ireland), achieving a spectral resolution of approximately 5 cm⁻¹. A dichroic beam splitter (Chroma Technology Corp., Bellows Falls, Vermont) and a holographic notch filter (Kaiser Optical Systems Inc., Ann Arbor, Michigan) were used to reject the elastically scattered light and pass only the Stokes-shifted light to the spectrograph. During acquisitions, the CCD was thermoelectrically cooled to −60 °C to suppress electronic noise due to thermal fluctuations. Finally, because of the limited height of the CCD array, only spectra from the 40 central fibers of the fiber bundle were imaged onto the CCD.

Full CCD spectral images were imported into MATLAB^® (version 7.8, The MathWorks™, Natick, Massachusetts) for analysis. Spectral processing included cosmic ray removal, readout and dark-current subtraction, and correction for the frequency-dependent response of the system. A nonlinear spatial transformation was applied to each image in order to correct for optical aberrations in the spectrograph system in a manner similar to that described by Esmonde-White et al. ¹⁹ The spectrum of Raman scattered light collected by each fiber was extracted from the full image by fitting to the separately measured spectral response of each fiber.²⁰

All spectra were truncated to the 825 to 1750 cm⁻¹ region, which includes many relevant mineral and organic matrix bands.¹² The lower spectral cutoff was chosen due to the presence of a strong uncorrelated fluorescence contribution below 825 cm⁻¹. Following spectral truncation, the background fluorescence lineshapes were modeled and removed through the use of an iterative polynomial fitting routine.²¹ Next, the spectra were normalized to the area under the amide I peak in the 1625 to 1725 cm⁻¹ region and smoothed with a Savitzky–Golay filter²² over a 3 pixel (approximately 5 cm⁻¹) window, chosen to match the resolution of the spectrograph. The spectra from each bone [⩽120 total; ⩽40 spectra per region (1 per fiber) × 3 regions] were averaged, yielding a single Raman spectrum per tibia. Spectra with a large amount of high frequency noise relative to Raman signal, due to either a large fluorescence-to-Raman ratio or a small amount of absolute Raman signal, were discarded before averaging. The amount of high frequency noise in a spectrum was quantified by calculating

Following microCT and Raman measurements, the tibiae were tested in torsion, as previously described.²³ Briefly, the ends of the tibiae were cemented into square 6.35-mm² aluminum tube holders using polymethylmethacrylate bone cement (DePuy Orthopaedics, Inc., Warsaw, Indiana) in order to maintain a 4-mm gauge length over the proximal 1/3 of the tibial diaphysis. Following this procedure, the specimens were bathed in phosphate buffered saline at room temperature for at least 2 h in order to allow for rehydration of the tissue.²⁴ Specimens were then mounted on an EnduraTec TestBench™ system (200 N·mm torque cell; Bose Corporation, Minnetonka, Minnesota) and loaded monotonically in torsion at a rate of 1°/s until failure.²⁵ The torque data were plotted against the rotational deformation (normalized by the gauge length and expressed as rad/mm) to determine the ultimate torque (maximum applied torque prior to fracture). The root-mean-squared error in the ultimate torque measurement was 0.2 N·mm.

It should be noted that torsional testing was chosen for this study since it represents a mode of loading that commonly leads to long bone fractures in humans. However, qualitatively similar results would be expected from bending tests since similarities in the results of these tests have been reported during skeletal growth²⁵ and in a model of skeletal fragility.²⁶

The effects of treatment regimen (PL versus GC) and mouse genotype (WT versus hTNF-Tg) on ultimate torque, BMD, cortical thickness, and Raman band area ratios were examined using a two-way analysis of variance (ANOVA) model. All statistical analyses were performed in MATLAB^® and significance was set at p ⩽ 0.05.

Following these tests on the ensemble-averaged properties of each group, correlations were sought at the single-bone level between ultimate torque and various microCT- and Raman-derived parameters. Because of the small sample size, a leave-one-out cross-validation (LOOCV) approach was used to generate microCT- and Raman-based predictions of ultimate torque for each tibia, based upon calibration using the other 14. First, a univariate prediction model was built using BMD as a predictor variable. A more complex model was built using multiple microCT-derived parameters (cortical BMD, cortical thickness, polar moment of inertia, and bone area), via partial least squares regression (PLSR).²⁷ PLSR was also used to generate predictions based upon spectral data, either from selected Raman peaks (discussed below) or from the entire Raman spectrum.

The rank of the PLSR model used for each prediction was selected based upon the method described by Haaland and Thomas²⁷ applied to an inner cross-validation performed upon each calibration set. In addition, in order to further prevent overfitting the calibration data, the number of samples in the calibration set was required to be at least 3 times larger than the rank of the selected model.²⁸ Following prediction, the accuracy of each technique was assessed by calculating the root-mean-squared error of cross-validation (RMSECV).

The average Raman spectra acquired from each of the 15 mice are overlaid in Fig. 1. Major Raman bands associated with mineral (

The ensemble-averaged ultimate torque, BMD, cortical thickness, MTMR, and PTCR for each group of mice are presented in Fig. 2. Trends are readily visible for all five parameters: they decrease with GC treatment and also in transgenic mice. For ultimate torque and the Raman-based parameters, both of these trends were significant (as determined by two-way ANOVA); for the microCT parameters, only the mouse genotype was found to be significant. The magnitudes of the standardized ultimate torque effect sizes (along with 95% confidence intervals) were 1.20 [0.07,2.34], 2.25 [1.11,3.39], and 0.09 [0,1.22] for the treatment, genotype, and interaction effects, respectively. Only the main effects of treatment regimen and mouse genotype were included in the final ANOVA models since the interaction effects were not statistically significant.

Figure 3 provides scatter plots of ultimate torque as predicted by regression against BMD alone (a), multiple microCT parameters (b), and the full Raman spectrum (c). All of the plots show significant correlation between the measured and predicted values of ultimate torque, with the BMD plot being noticeably noisier. The accuracy of each prediction model is quantified in Table 1.

In order to isolate the Raman spectral features most associated with compromised bone strength, a Raman peak-based prediction algorithm was also sought. The Raman spectral features that were most predictive of ultimate torque were identified by analyzing the PLSR coefficients as described by Beier et al. ²⁹ Briefly, the Hadamard, or entrywise, product of the mean-centered spectrum of the tibia with the largest ultimate torque,

The trend in the ensemble-averaged measurements of the microCT-derived BMD and Raman spectroscopy derived MTMR closely matches the biomechanical ultimate torque measurements as seen in Fig. 2. The BMD and MTMR measurements suggest that both the total mineral content and phosphate mineral content relative to the protein matrix are reduced in the bones of GC-treated and hTNF-Tg mice. Hypomineralization is expected in each of these cases since both GC-treatment and arthritic inflammation decrease bone formation and increase bone resorption.³⁰ The concomitant decrease in the PTCR implies that the degree of carbonate substitution in the apatite crystal lattice is larger in the bones of GC-treated and hTNF-Tg mice compared to the wild-type controls. Others have reported a similar decrease in the PTCR in women who had experienced an osteoporotic fracture.¹³

The fact that statistically significant interaction effects were not observed in this study may be due to the relatively small sample size (15 total specimens). Potentially relevant ultimate torque effect sizes are those greater than about 2.6 N·mm (the prediction accuracy of the full-spectrum Raman PLSR model), which corresponds to a standardized effect size of about 0.6. Further study of the interaction between glucocorticoid treatment and mouse genotype is warranted since this effect size, as well as an effect size of zero, is included in the 95% confidence interval for the magnitude of the ultimate torque interaction effect [0,1.22].

The regression models for prediction presented in Fig. 3 and Table 1 suggest that Raman signatures from bone, if obtained transcutaneously, may be capable of screening for fracture risk more reliably than the conventional measurement of BMD via DXA or computed tomography. DXA is the current standard for diagnosing osteoporosis since it is the most extensively validated technology for clinical prediction of fracture risk.³¹ Computed tomography can also be used to measure BMD and provides additional quantitative information about bone structure. However, computed tomography requires a higher radiation dose and remains more expensive and less accessible than DXA. The scatter plots in Fig. 3, along with Table 1, provide estimates of the fracture risk prediction accuracy of DXA [Fig. 3], computed tomography [Fig. 3], and Raman spectroscopy [Fig. 3]. Among these technologies, Raman spectroscopy produced the smallest prediction error. Raman spectroscopy could also provide a potential cost and space benefit over clinical DXA and computed tomography scanners, and it is arguably safer as it does not require the use of ionizing radiation. While transcutaneous acquisition of bone Raman signatures will inevitably reduce the signal strength, the ability to acquire such spectra in mice,^32- 33 rats,³⁴ chickens,³⁴ canines,³⁵ and humans³⁶ has recently been demonstrated.

Previous Raman spectroscopy studies of osteoporotic bone have primarily focused on changes to individual spectral features associated with specific biochemical properties of the bone.¹¹ Recent studies have also revealed correlations between various Raman spectral features and a number of parameters related to the bone's biomechanical strength.^37- 39 Specifically, correlations between the ultimate stress (maximum applied stress prior to fracture) and the Raman-derived MTMR and crystallinity (which is related to the inverse of the

This study is the first to evaluate bone changes using a variety of minimally invasive imaging and spectroscopic techniques and to establish predictive correlations with the biomechanical properties of cortical bone in a mouse model of rheumatoid arthritis. The data presented here examined mice at only a single GC-exposure time point (6 weeks). Interesting future directions include examining mice at earlier time points to evaluate whether Raman spectroscopy can predict early, subtle chemical changes to the bone before later changes in BMD and deterioration of biomechanical properties are manifest. Toward clinical translation of this diagnostic technology, longitudinal in vivo studies in mice utilizing transcutaneous detection will also be pursued.

Raman spectroscopy has been used to study the chemical perturbations to cortical bone associated with both glucocorticoid-induced osteoporosis and rheumatoid arthritis. Each disease was associated with statistically significant changes in the degree of phosphate mineralization and overall mineral composition in the tibial cortical bone of mice. Raman spectroscopy and microcomputed tomography were also separately used to predict tibial fracture risk as quantified by a biomechanical measurement of ultimate torque (maximum applied torque prior to fracture). Predictions based upon the measured Raman spectrum were as accurate as those produced by microcomputed tomography derived parameters, and more accurate than the clinically-used parameter of bone mineral density.

This work was supported in part by a University of Rochester Provost's Multidisciplinary Award and NIH Grant No. NIH 1R21AR061285-01 . We would like to thank Dr. Edward Schwarz for valuable advice and for providing the breeder mice for the hTNF-Tg mouse colony used in this study, Dr. Subhash Juneja for his help with maintaining and genotyping the hTNF-Tg mouse colony, and Mr. Michael Thullen for his excellent technical assistance with the microcomputed tomography analysis.