Impact of Long-Term Storage on Mid-Infrared Spectral Patterns of Serum and Synovial Fluid of Dogs with Osteoarthritis

Simple Summary

Signals generated from passing infrared radiation through biological samples such as serum can create distinct wave patterns, or “fingerprints”, that can be used to distinguish samples with or without diseases such as osteoarthritis. In research, the collection of a large database of samples for studying diseases such as osteoarthritis can take years and requires the storage of samples in freezers. The aging of samples in the frozen state can change both the quality of samples and the results that are obtained from aged samples versus fresh, frozen samples. In this study, serum and joint fluid from dogs that had knee osteoarthritis and those that did not were used to compare the quality of the generated fingerprints from samples after short- and long-term storage in the frozen state. The results showed that even though aging of samples in the frozen state causes statistically significant differences between the fingerprint of samples, these changes only account for 2–3% of the overall variance in the spectroscopic data set (compared, e.g., to the variability stemming from differences between samples having been withdrawn from arthritic versus healthy dogs, which amounts to 77% for serum or even 86% for synovial fluid). The fingerprint of both serum and joint fluid samples after long-term storage of about seven years could still be used to distinguish between arthritic and healthy dogs with a high degree of accuracy.

Abstract

Mid-infrared spectral (MIR) patterns of serum and synovial fluid (SF) are candidate biomarkers of osteoarthritis (OA). The impact of long-term storage on MIR spectral patterns was previously unknown. MIR spectra of canine serum (52 knee-OA, 49 control) and SF (51 knee-OA, 51 control) were obtained after short-term and long-term storage in −80 °C. Multilevel simultaneous component analysis and partial least squares discriminant analysis were used to evaluate the effect of time and compare the performance of predictive models for discriminating OA from controls. The median interval of storage between sample measurements was 5.7 years. Spectra obtained at two time points were significantly different (p < 0.0001); however, sample aging accounted for only 1.61% and 2.98% of the serum and SF profiles’ variability, respectively. Predictive models for discriminating serum of OA from controls for short-term storage showed 87.3 ± 3.7% sensitivity, 88.9 ± 2.4% specificity, and 88.1 ± 2.3% accuracy, while for long-term storage, they were 92.5 ± 2.6%, 97.1 ± 1.7%, and 94.8 ± 1.4%, respectively. Predictive models of short-term stored SF spectra had 97.3 ± 1.6% sensitivity, 89.4 ± 2.6% specificity, and 93.4 ± 1.6% accuracy, while for long-term storage they were 95.7 ± 2.1%, 95.7 ± 0.8%, and 95.8 ± 1.1%, respectively. Long-term storage of serum and SF resulted in significant differences in MIR spectral variables without significantly altering the performance of predictive algorithms for discriminating OA from controls.

1. Introduction

Biobanking biological fluid samples (e.g., serum, urine, joint fluid) is an essential component of prospective clinical studies and a means of maximizing the efficiency of using obtained samples. These biological samples and the contents therein (i.e., cellular, molecular) are at risk of being compromised during the pre-analytical phase of projects. The pre-analytical phase is where sample collection technique, handling, and storage protocols may inadvertently alter the components of interest. Cryopreservation of biological fluids requires standardized protocols, reliable inventory, and specialized facilities to preserve valuable samples. Factors that have been shown to impact the stability of the samples during the storage phase are temperature, storage time, and freeze-thaw cycles. The impact of these factors on various biomarkers and sample components has been investigated in previous studies [1,2,3]. However, recommendations for optimal storage variables vary depending on the stability of the target components (e.g., molecule being measured) and measurement technique. The use of mid-infrared (MIR) spectral patterns of biological fluids to detect variability between disease states as a novel approach to disease pattern recognition has been investigated at an increasing rate in the past few decades [4,5,6,7,8,9,10,11,12]. The spectral pattern of a given sample is the sum of all MIR light absorbance by the infrared-active molecular bonds within the sample, which is displayed as a unique waveform (spectrum) [13]. These unique patterns of spectra of complex samples such as serum can be used as “fingerprints” to distinguish between samples that have significantly different absorption patterns due to their composition. These differences in molecular composition can be due to changes caused by underlying disease processes (e.g., inflammatory diseases such as osteoarthritis). Fourier-transform infrared spectroscopy (FTIR) is one of the techniques utilized in acquiring MIR spectra of biological samples that is adjuvant-free, cost-effective, simple, and requires a small volume of samples [5,6,7,8,9]. Spectral variables (i.e., fingerprints) based on this method have been successfully used in differentiating various types of arthritis in humans [4]. Research in experimental and clinical models of osteoarthritis (OA) in various joints of animal models has also been able to demonstrate the ability to distinguish OA from control samples with high accuracy rates based on the MIR spectra of biological samples (i.e., serum, synovial fluid) [5,6,7,8,9]. Short-term (<one year) longitudinal studies using spectral variables in serum and joint fluid of animals have shown reproducible spectral characteristics of OA that, despite changes over time, remain distinguishable from control samples [7,8,9]. Studies investigating long-term changes in clinical OA in larger animal models (e.g., dogs) and human models require at least several years that warrant biobanking of collected biological samples to allow for batch analysis. These biobanked biological samples (e.g., serum, synovial fluid) are typically stored in −80 degrees Celsius for long-term storage to minimize the impact of time (e.g., degradation of molecules). However, the impact of long-term storage on the spectral pattern of these biological samples has not been previously investigated. The first aim of this study was to evaluate whether long-term storage results in changes in the MIR spectral pattern of serum and synovial samples of dogs with and without knee OA as measured by FTIR spectroscopy. The second aim of this study was to evaluate the impact of any observed changes due to long-term storage on the ability to discriminate between the serum and synovial fluid spectra patterns of dogs with and without knee OA. The hypothesis of the study was that long-term storage would have minimal impact on MIR spectral patterns of serum and synovial fluid of dogs with or without knee OA.

2. Materials and Methods

2.1. Samples

The sample size used in the current study was based on the available number of samples from the original project that had two arms: investigating serum and synovial fluid samples from client-owned dogs with OA secondary to naturally occurring degenerative (non-traumatic) cranial-cruciate ligament (CrCL) tears in one or both knees and controls [8,9]. The presence of CrCL tears and OA changes in the OA group dogs had been confirmed intraoperatively by inspection of the joint via arthrotomy or arthroscopy. The controls were otherwise healthy adult dogs with no orthopedic abnormalities and with both knees free of gross abnormalities upon evaluation immediately after euthanasia for reasons unrelated to the study. Evaluation of the knee joint in control dogs was via an opening of the knee joint and evaluation of the joint, including cartilage surfaces, ligaments, menisci, synovium, and joint capsule. Briefly, venous blood samples were collected, and serum was separated and saved in 0.5–1 mL aliquots in cryovials (Nalgene Cryogenic tubes, VWR International, Batavia, IL, USA) and preserved at −80 °C until batch analysis. Synovial fluid samples had been collected for the original study using an aseptic technique from the knees of dogs with a cranial-cruciate ligament tear in the OA group under general anesthesia immediately prior to surgical intervention to treat knee instability (i.e., tibial plateau leveling osteotomy), and from healthy knees in the control group immediately after euthanasia. The synovial samples were also saved in 0.5 mL aliquots in cryovials and preserved at −80 °C until batch analysis. After the completion of the original project, the unused aliquots of serum and synovial samples were maintained in −80 °C storage for a minimum of five years.

The available sample inventory was reviewed for OA and control samples with adequate volume left for analysis. For the OA group, the serum and synovial fluid samples were those obtained prior to any surgical intervention. If more than one aliquoted sample was available, the clearest was selected (i.e., none to minimal blood contaminated for synovial fluid and none to minimal hemolysis for serum samples). Only one serum sample per dog was selected. In the OA group, even if both knees were affected, only one sample from one knee per dog was included in the study. In the control group, both healthy knees of each dog had been sampled, but preferentially, only one sample from each dog was included.

The age of samples at the time of the first spectral analysis was calculated based on the date the sample was obtained from the dog and the time when the sample was thawed and the first spectral analysis was performed. Samples from this initial storage period (i.e., analyzed within a year after being acquired) are referred to as “short-term storage” samples. The age of samples at the time of the second spectral analysis after being in storage for a minimum of 5 years was calculated based on the date of sample acquisition from the dog and the time when the sample was run for the second time. Samples from the longer storage period (>1 year) are referred to as “long-term storage samples”. The interval between measurements for the short-term and long-term storage was calculated based on the dates the MIR spectra of each sample were obtained.

2.2. FTIR Spectroscopy

At the time of the first spectral acquisition (i.e., after short-term storage), serum and synovial fluid samples were thawed at 22 °C, and dried films were prepared as described previously [8,9,14]. For each sample, an aliquot was drawn and diluted in a potassium thiocyanate (KSCN) (SigmaUltra, Sigma-Aldrich Inc., St Louis, MO, USA) solution (4 g/L) at a 2:1 serum/SF–to–KSCN ratio (40:20 μL). KSCN was used as an internal control. The absorption peak for KSCN at 2060 cm⁻¹ served as a reference point for normalization purposes. For each sample, replicate (8 μL per replicate) dry films were made on a silicon microplate [5]. After drying at room temperature (20–22 °C), the microplate was mounted on a multi-sampler (HTS-XT, Autosampler, Bruker Optics, Milton, ON, Canada) interfaced to the infrared spectrometer (Tensor 37, Bruker Optics). Mid-infrared absorbance spectra in the range of 400 to 4000 cm⁻¹ were recorded with proprietary software (OPUS software, version 6.5, Bruker Optics, GmbH, Ettlingen, Germany). For each acquisition, 512 interferograms (scans) were accumulated and Fourier transformed to generate a spectrum with a nominal resolution of 4 cm⁻¹ [5,6,15]. Six (short-term storage) or three (long-term storage) replicates (due to the limited volume of available samples) were prepared and analyzed for each sample, and the corresponding spectra were averaged prior to the successive data processing. At each measurement time point, all the spectra were acquired within a short time span (~10 days).

2.3. Data Analysis

Analyses of non-spectral data were performed using SPSS software (IBM SPSS Statistics, v. 25). Variables without a normal distribution were described by their median and interquartile range (IQR). All acquired serum and synovial-fluid spectra files from both time points were imported in MATLAB^® (R2015b (8.6.0.267246); The Mathworks, Natick, MA, USA) for the successive data processing which was carried out utilizing in-house written scripts. For each sample, at each time point, the average of the replicate spectra was used for analysis. Prior to any modeling, the data were preprocessed by the first derivative (using the Savitzky-Golay algorithm [16] with 19 points window and second-order polynomial) followed by mean centering.

Modeling was conducted in two different stages. At first, multilevel simultaneous component analysis (MSCA) was used to verify whether there could be any significant difference between the spectra of the same samples measured after short- and long-term storage. Indeed, MSCA can be considered as a multivariate generalization of repeated measurements ANOVA [17,18]. In particular, the overall (preprocessed) spectral matrix $\mathit{X}$ is partitioned into the individual contributions of between-sample (${\mathit{X}}_b$) and within-sample (${\mathit{X}}_w$) systematic sources of variability plus the residuals (${\mathit{X}}_{res}$, i.e., the variation not accounted for by the model), according to

$$\mathit{X}={\mathit{X}}_b+{\mathit{X}}_w+{\mathit{X}}_{res}$$

In particular, the effect of sample aging on the spectroscopic signal is related to the within-sample variation described by ${\mathit{X}}_w$ and can be quantified as the sum of squares of the elements in that matrix. The significance of the contribution is evaluated by comparing such effect with its distribution under the null hypothesis, which is estimated by means of a permutation test [19]. On the other hand, if the effect is found to be significant, its impact on the spectroscopic signal can be interpreted by PCA of the associated matrix ${\mathit{X}}_w$. In the present study, MSCA analysis was conducted independently on the serum and synovial fluid samples.

In a second stage, to verify whether MIR analysis of synovial fluid or serum samples could provide the basis for a reliable discrimination between OA and control even after long-term storage and if the same spectroscopic markers could be found as when analysis was carried out after short-term storage, supervised pattern recognition (classification) models were built and validated [20]. Due to its ability to deal with highly collinear predictors (e.g., spectral variables), partial least squares discriminant analysis (PLS-DA) was selected [21,22]. The PLS-DA method is based on projecting the data onto a low-dimensional sub-space of latent variables, which are relevant to highlight differences between the groups. Accordingly, model building is required to identify the optimal dimensionality of such subspace. On the other hand, the quality of the predictive models, which can be summarized by different figures of merit, such as sensitivity, specificity, overall classification accuracy, and the area under the ROC curve (AUC), needs to be evaluated on samples not used for model development, to avoid over-optimistic results. Accordingly, to build the models and validate the prediction results and the identified spectroscopic markers, a repeated double cross-validation (rDCV) procedure coupled with permutation tests was adopted [23,24]. Double cross-validation (DCV) is a resampling procedure where the available samples are split according to two nested loops of cross-validation: the outer loop samples are treated as external validation samples, which do not take part either in model building or in model selection; indeed, optimization of the model parameter is carried out on the inner loop samples. The term repeated derives from the fact that the whole DCV procedure is repeated a sufficient number of times (here, 50), changing the distribution of the samples within the cancelation groups so that, on one hand, the results are not dependent on the particular sample splitting among the groups and, on the other, it is possible to obtain confidence intervals for all the classification figures of merit and model parameters. Furthermore, to rule out any possibility of obtaining good results just due to chance correlations, permutation tests were used to non-parametrically evaluate the null distributions of the classification figures of merit, providing p values to estimate the significance of the observed discrimination [23]. Statistical significance was set at p < 0.05. The analyses were conducted independently to report model performance based on whether serum or synovial fluid samples were used.

3. Results

There were 52 serum and 51 synovial fluid samples available from the 70 dogs originally in the OA group, while 49 serum and 52 synovial fluid samples were available from the 50 dogs in the control group. In the control group, one dog did not have adequate serum samples available. Two of the control dogs did not have adequate synovial fluid samples; therefore, the additional three samples were selected from the contralateral knees of three of the control dogs that had adequate samples from both knees.

In the OA group, all 70 dogs had radiographic evidence of OA changes (i.e., joint effusion, intra-articular and peri-articular osteophytosis). The knee OA changes were present unilaterally in 63 (30 right, 33 left) and bilaterally in seven dogs. In cases with bilateral OA secondary to CrCL tears, only the more clinically affected stifle that was operated on was sampled.

The median (IQR) of time in a frozen state before the first run (short-term storage) and the second run (long-term storage) were 1.1 (1.1) and 6.8 (1.1) years, respectively. The time interval between the first and second measurement was 5.7 (0.03) years. At first, MSCA was used to evaluate whether there could be any significant difference in the spectral profiles of the samples after short- or long-term storage, and serum samples were considered. When the overall variability in the preprocessed spectral profiles was partitioned according to the ANOVA scheme, it was found that only 1.61% of the total variability among the serum spectra was ascribable to the effect of aging (between-samples differences accounted for 77.23% of the total variance, while the remaining part is residual variation, associated with random error). The variability observed between the two time points was not impacted by the type of sample (OA versus control). Although the relatively low amount of spectral variance was associated with the differences between short-term and long-term storage, permutation testing indicated that these differences were statistically significant (p < 0.0001). Principal component analysis (PCA) of the effect matrix resulting from the ANOVA decomposition indicated that long-term stored serum samples were characterized, on average, by a lower intensity of most of the peaks. The same analysis was then conducted on the synovial fluid samples. Also, in this case, when the overall variability in the preprocessed spectral profiles was partitioned according to the ANOVA scheme, it was found that the largest part of the spectral variance was due to between-samples differences (85.54%), with only 2.98% of the total variability among the spectra corresponding to the effect of aging. Permutation testing indicated that the effect of sample ageing on the spectral signature, though relatively small, could still be deemed statistically significant (p < 0.0001). Interpretation of the observed difference in synovial-fluid spectra by PCA of the corresponding effect matrix confirmed that long-term storage results, on average, in less intense absorption peaks.

Based on the outcomes of this exploratory analysis, which indicated that sample ageing could result in small but significant differences in the spectral signature, in a second stage of the study we wanted to verify if the analysis of long-term stored samples could still provide reliable predictions and, if so, whether models built on the two set of samples (short- and long-term stored) resulted in the same set of putative markers. First, the spectra collected on the serum samples were considered. The quality of the predictive models was evaluated by considering the figures of merit calculated on the outer loop of the rDCV procedures, which correspond to the classification performances on samples not used for model building or optimization. Moreover, since rDCV uses resampling, this allows for the estimation, for each figure of merit, of not only a single value but also a confidence interval (in particular, in the remainder of the text, 95% confidence level will always be considered). The predictive models for discriminating serum OA samples (n = 50) from controls (n = 49) built on the spectra from short-term storage showed sensitivity, specificity, and accuracy of 87.3 ± 3.7%, 88.9 ± 2.4%, and 88.1 ± 2.3%, respectively. The predictive models based on the same serum samples after long-term storage showed sensitivity, specificity, and accuracy of 92.5 ± 2.6%, 97.1 ± 1.7%, and 94.8 ± 1.4%, respectively. In both cases, the very good discriminant ability of the models are graphically visualized in Figure 1, where the mean scores (and the corresponding confidence intervals) of the samples along the single canonical variate (discriminant direction) of the models are displayed. In both cases, it can be observed how almost all of the OA samples have positive scores along the direction, whereas many of the controls have negative coordinates.

Having verified that both models provided accurate discrimination between OA and control samples, inspection of the spectral variables mostly contributing to the observed differentiation was conducted by analyzing the values of the variable importance in projection (VIP) indices. The VIP indices summarize the contribution of each experimental variable to the PLS-DA model and are normalized so that a greater-than-one criterion can be used to assess their significance. Accordingly, the wavenumbers mostly responsible for the observed differences between OA and control serum samples at each time point are presented in Figure 2.

The analysis was then repeated for the synovial-fluid samples based on the OA (n = 51) and control (n = 52) groups. The classification model based on short-term storage synovial fluid spectra had sensitivity, specificity, and accuracy of 97.3 ± 1.6%, 89.4 ± 2.6%, and 93.4 ± 1.6%, respectively, when evaluated on the outer loop of the rDCV procedure. The predictive models based on the same synovial-fluid samples after long-term storage showed sensitivity, specificity, and accuracy of 95.7 ± 2.1%, 95.7 ± 0.8%, and 95.8 ± 1.1%, respectively. Also, in this case, the very good discriminant ability of the models is graphically visualized in Figure 3, where the mean scores (and the corresponding confidence intervals) of the samples along the single canonical variate (discriminant direction) of the models are displayed. On the other hand, the wavenumbers responsible for the observed differences between OA and control SF samples at each time point, identified by inspection of the VIP indices calculated from the corresponding PLS-DA models, are presented in Figure 4.

4. Discussion

To the authors’ knowledge, this is the first study evaluating the impact of long-term sample storage on the quality of MIR spectra of serum and synovial-fluid samples used for discriminating OA from control samples. Both serum and synovial fluid samples spectra had statistically significant differences in spectral variables due to the aging of the samples; therefore, our hypothesis was rejected. However, the contributions of these differences to the overall spectral variability for both serum and synovial fluid were relatively small (i.e., <3%). It is important to note that despite the statistically significant differences between the spectra of samples from short-term versus long-term storage, these differences were not a significant component of the observed differences between OA and control samples. This is reflected in the continuously high performance of the predictive models (>87% for sensitivity, specificity, and accuracy) for both serum and synovial fluid samples despite the time of measurements. Visual assessment of spectra from different time points for serum and synovial fluid samples (Figure 2 and Figure 4) also shows how there is a high consistency in the variables identified as putative spectroscopic markers of OA between the two models, thus confirming that even if prolonged storage time can result in small but significant spectral differences, these differences do not affect the reliability and the interpretation of classification models built on the data. This finding supports the use of biobanking of samples for ~5 years if MIR spectral fingerprints are used for discriminating between OA and control samples using the described methodology.

This difference between the levels of variability due to aging in the samples (i.e., 1.61% in serum versus 2.98% in synovial fluid) may be partly due to the types of molecules in the synovial fluid that may be more vulnerable to freezing and degradation over time. But more importantly, the sample storage protocols may have a role in the observed differences. Serum is cell-free, whereas the synovial fluid preparation in this study involved freezing the sample without any further processing (i.e., no centrifugation or separation of cellular components pre-freezing). Therefore, the presence of cellular components and their unique degradation patterns may have resulted in the greater variability between the synovial fluid spectra based on time in a frozen state. The influence of different sample collection, processing, and freezing methods on results of metabolomics studies using synovial fluids have developed optimized protocols to minimize loss of data during measurements [25,26]. Therefore, if the dried-film FTIR spectroscopy technique used in this study is to be compared with metabolomics and proteomic techniques, the sample processing and storage techniques must be uniform to allow for direct comparisons.

This study shows a small but acknowledgeable impact of long-term storage that needs to be considered when utilizing this methodology for prospective clinical trials where samples are collected over years, particularly for chronic and insidious disease processes such as OA. In such long prospective studies or when biobanked samples are utilized, batch analysis of the samples is preferred to reduce other variations in testing conditions (e.g., operator variation, environmental variables, etc.). Therefore, ensuring that confounding variables such as sample aging are not significantly impacting test results is paramount. The results of this study cannot be extrapolated to spectral patterns of other disease processes, as degradation of sample components unique to other disease processes may not follow the same pattern as in OA or controls. Additionally, the results of this study cannot be used to directly identify the molecular changes responsible for the observed difference in the spectral fingerprint. However, the degradation of both serum and SF samples appears to be relatively consistent between the time points in this study, resulting in less intense absorbance peaks. The absorbance peaks of the spectra are directly related to the concentration of the sum of all molecular bonds that are MIR active at each wavenumber (i.e., Beer’s law) [13]. Therefore, a decrease in peak intensity in the face of unchanged testing conditions (e.g., sample dilution, spectrometer settings) is reflective of the degradation of these MIR active molecular bonds. Quantitative evaluation of differences in molecules and compounds in these biological samples from different time points requires prospective studies that perform quantitative assays (e.g., proteomics, metabolomics). The identified molecules with differences over time can then be matched with the MIR peaks at wavenumbers that correspond to their molecular bonds. The main intent of the FTIR spectroscopy of dried films technique used in this study is as a cost-effective, sample-sparing, adjuvant-free screening tool to distinguish between samples with or without OA rather than identify the underlying causes of the observed differences [27].

Another limitation of this study is the lack of additional, shorter intervals between measurements and longer storage periods to assess possible limits for sample storage time and trends as to when the sample variability due to age-related degradation overwhelms the differences based on disease state. Future studies can prospectively store samples and sequentially measure MIR spectra over shorter intervals (e.g., six-month intervals) to document these changes. Additionally, some of the serum and synovial fluid samples in this study had undergone up to two freeze-thaw cycles due to limited sample volumes available, and the impact of these additional freeze-thaw cycles on the overall observed changes using the FTIR spectroscopy methodology in this study cannot be separated out. The impact of the number of freeze-thaw cycles on serum and synovial samples have been shown to have a variable impact on the results based on measured variables and methods of measurement [28,29]. Future studies can specifically evaluate the impact of the number of freeze-thaw cycles on the quality of the spectra using the methodology in this study. However, the fact that despite these confounding factors, the overall changes observed do not significantly impact the predictive modeling for discriminating OA from controls shows merit in this technique despite the limitations. Future studies could evaluate the impact of the number of freeze-thaw cycles on the quality of MIR spectral patterns obtained using the methodology in this study.

5. Conclusions

In conclusion, storing serum and synovial fluid samples of dogs with knee OA and controls in −80 °C results in changes in the spectral patterns after ~5 years. However, these changes have minimal impact on the ability to use the spectral fingerprints of these samples after long-term storage for discriminating between OA and control samples. Future studies could evaluate measured MIR spectra of samples in storage prospectively at shorter intervals to establish trends of sample degradation over longer follow-up times and set maximal limits on sample viability for biobanking purposes.