Raman microspectroscopic analysis of the tissue-specific composition of the human osteochondral junction in osteoarthritis: A pilot study

This study investigates the influence of osteoarthritis (OA) disease severity on the bio-composition of the osteochondral junction at the human tibial plateau using Raman microspectroscopy. We specifically aim to analyze the spatial composition of mineralized osteochondral tissues, i.e., calcified cartilage (CC) and subchondral bone plate (SBP) from unfixed, hydrated specimens. We hypothesize that the mineralization of CC and SBP decreases in advanced OA. Twenty-eight cylindrical osteochondral samples ( d = 4 mm) from tibial plateaus of seven cadaveric donors were harvested and sorted into three groups following histopathological grading: healthy ( n = 5), early OA ( n = 8), and advanced OA ( n = 15). Raman spectra were subjected to multivariate cluster analyses to identify different tissues. Finally, the tissue-specific composition was analyzed, and the impact of OA was statistically evaluated with linear mixed models. Cluster analyses of Raman spectra successfully distinguished CC and SBP as well as a tidemark region and uncalcified cartilage. CC was found to be more mineralized and the mineral was more crystalline compared with SBP. Both tissues exhibited similar compositional changes as a function of histopathological OA severity. In early OA, the mineralization tends to increase, and the mineral contains fewer carbonate substitutions. Compared with early OA, mineral crystals are rich in carbonate while the overall mineralization decreases in advanced OA. This Raman spectroscopic study advances the methodology for investigating the complex osteochondral junction from native tissue. The developed methodology can be used to elucidate detailed tissue-specific changes in the chemical composition with advancing OA. Statement of Significance In this study, Raman microspectroscopy was utilized to investigate the influence of osteoarthritic degeneration on the tissue-specific biochemical composition of the human osteochondral junction. Multivariate cluster analyses allowed us to characterize subtle compositional changes in the calcified cartilage and subchondral bone plate as well as in the tidemark region. The compositional differences found between the calcified cartilage and subchondral bone plate in both organic and mineral phases will serve as critical benchmark parameters when designing biomaterials for osteochondral repair. We found tissue-specific ∗ Corresponding author: Shuvashis Das Gupta, M.Sc., Research Unit of Medical Imaging, Physics and Technology, Faculty of Medicine, University of Oulu, POB 50 0 0, FI-90 014

changes in the mineralization and carbonate substitution as a function of histopathological OA severity. Our developed methodology can be used to investigate the metabolic changes in the osteochondral junction associated with osteoarthritis.

Introduction
Osteoarthritis (OA) is the most common joint disease, causes pain and leads to disabilities [1] . While the exact etiology and pathogenesis of OA remain unclear, the disease is characterized by articular cartilage (AC) degeneration and subchondral bone (SB) sclerosis along with other tissue modifications such as pannus, bone cysts, and bone marrow lesions [2] . Recent studies suggest that cartilage degeneration and SB sclerosis are associated, synchronous phenomena [3 , 4] , and there is substantial evidence that changes in the SB occur in very early stages of OA-in some cases, even before cartilage loss [5][6][7] .
In OA, the thickening of the subchondral plate, i.e., the calcified cartilage (CC) and cortical subchondral bone plate (SBP), is well documented [3 , 4 , 7 , 8] ; this process includes reduced bone turnover and thinning of trabeculae [7] . However, these typically occur in the later stages of OA [7] . On the other hand, observations of surgically-induced animal models of OA suggest that the thickness of the subchondral plate could even decrease in early OA [9 , 10] . Thinning of the subchondral plate is reported to be highly sitespecific and has been suggested to be associated with increased remodeling rate of SB and altered joint-loading conditions [9] . During OA progression, the CC starts to advance into the deep AC in a process similar to endochondral ossification [6 , 11] , developing multiple mineralization fronts-a phenomenon called tidemark duplication [6] . In addition, the cement line advances toward the CC by the chondroclastic resorption process [12] . All these processes, changing the subchondral plate thickness with associated bone remodeling, involve biochemical alterations of mineralized tissues [6 , 13 , 14] . To understand these biomineralization processes in different stages of OA, it is imperative that the biochemical alterations in the different tissues are measured accurately. However, it is a major challenge to distinguish the relatively thin CC from the complex osteochondral junction owing to limitations in the spatial resolution or contrast of current microscopic imaging modalities. Because of this, only few studies have been conducted on the CC compared with other joint tissues. Moreover, the challenge to the researcher is exacerbated when considering that the alterations of the SBP in late OA are quite different from those of the cancellous bone [15] . Investigating the changes in the chemical composition of the CC and SBP as a function of OA severity will help elucidate the metabolic changes associated with the disease development.
To understand the metabolic activities of the CC and SBP in OA, studies on both mineral and organic matrix compositions with high spatial resolution and adequate contrast are essential. In humans, depending on the joint location and age, the average thickness of the CC is 10 0-30 0 μm [16] . It consists of chondrocytes embedded in a calcified type-II collagen-rich matrix [17 , 18] , where the mineral phase is carbonate-substituted hydroxyapatite with crystal size similar to the adjacent bone tissue [18][19][20] . It has been shown that the composition is directly related to the mechanical properties of mineralized tissue [21][22][23][24] . In an earlier study, a higher calcium content was found in the CC than in the underlying bone tissue [25] . Furthermore, the CC in OA was reported to have increased stiffness as it became hypermineralized [14 , 16] . However, in previous studies, the measurements and observations were made with fixed/embedded specimens, which have the potential to introduce artifacts by altering tissue composition or protein conformation [26] . For instance, the commonly used fixative agent, formaldehyde, preserves biological tissue by introducing cross-links to the organic components [27] .
Minimal need for sample preparation and compatibility with hydrated samples make Raman spectroscopy an ideal candidate for studying the composition of fresh tissue specimens. Moreover, the high spatial resolution ( < 1 μm) of Raman microspectroscopy is particularly useful for separating the tidemark and CC from the osteochondral junction. Raman microspectroscopy is based on inelastic scattering of photons and the Raman spectral signatures correspond to the molecular vibrational energies. The biochemical fingerprint region of the Raman spectrum (350-1750 cm −1 ) yields information on both the mineral and organic phases of the tissue [28] . Being a label-free and nondestructive measurement technique, a specimen examined by Raman spectroscopy can be further studied for other purposes, e.g. , investigation of mechanical properties via instrumented indentation techniques. Because of these advantages, Raman microspectroscopy has been applied in numerous studies to investigate different connective tissues, including AC [29][30][31] and bone [21 , 23 , 28 , 32 , 33] . However, to the best of our knowledge, there are no studies on the tissue-specific composition of the human osteochondral junction as a function of histopathological OA severity.
Since advanced OA involves increased bone deposition [7 , 34] and bone remodeling rate [4 , 35] , we hypothesize that such bone tissue is relatively younger, and thus the mineralization in the CC and SBP decrease in advanced OA. To test this hypothesis, in this pilot study, we aim to first identify different tissue types in the osteochondral junction using Raman microspectroscopy, and then quantify the relative biochemical composition of these tissues.

Sample preparation
Twenty-eight human osteochondral samples were collected from the tibial plateau (from the medial and lateral compartments of both legs) of seven cadaveric donors (age 70 ± 6 years: 6 males and 1 female) (approval no 134/2015 The Research Ethics Committee of the Northern Savo Hospital District, Kuopio, Finland). Cylindrical osteochondral plugs (diameter = 4 mm) were prepared with a dental drill from tibial plateaus and stored in 1x Dulbecco's phosphate-buffered saline at −20 °C. The plugs were further prepared for Raman measurement by cutting ~1.7 mm sections with a low-speed diamond saw (Isomet 10 0 0, Buehler, USA). These were polished with grit paper and ultra-sonicated (Wagner Instrusonic, W-181, Finland) for 5 min at room temperature. The samples were kept wet during all preparation stages. The osteochondral sections were placed on a petri dish with dental wax and submerged in distilled water for the Raman measurements. Both the tissue processing and Raman acquisition in aqueous solution can potentially modify the highly labile mineral crystals [36] . Therefore, a separate investigation of the possible demineralization effect of the aqueous medium on the human bone tissue during the Raman measurement period was performed (Supplementary material: section 1). This demonstrated that the acquisition in the aqueous medium did not considerably affect the Raman spectra of the mineralized tissue.
The remaining osteochondral sections were fixed in formalin and decalcified in ethylenediaminetetraacetic acid. After decalcification, the sections were dehydrated and embedded in paraffin. Subsequently, three adjacent thin sections (thickness = 3 μm) were cut using a microtome (Thermo Fisher Scientific, Microm HM 355S, USA) and stained with Safranin-O for the OA histopathology assessment.

Raman microspectroscopy
A DXR TM 2xi confocal Raman imaging microscope (Thermo Fisher Scientific, USA) equipped with a 60x/1.00 immersion objective and wide range grating (50-3250 cm −1 with a spectral resolution of 5 cm −1 ) was used for collecting Raman maps. To excite the Raman signal, a depolarized 785 nm laser (30 mW) and 50 μm confocal pinhole aperture were used. Firstly, a mosaic image of the whole sample was acquired and based on the visual inspection, a rectangular region containing the deep AC, tidemark, CC, and SBP was selected for Raman mapping. Owing to the remaining sample surface irregularities, measurements were divided into small rectangular sections (green boxes in Fig. 1 A) and section-specific focus was set ~10 μm below the surface of the sample. The section containing the tidemark was imaged with a step size of 1 μm (region size: 60 μm × 60 μm) and remaining sections with 3 μm (region size: ~60 μm × 270 μm). Every Raman spectrum was collected for 0.5 s and averaged 15 times. Overall, this resulted in acquisition time of approximately 12 h for each sample.

Spectral preprocessing
The raw spectrum from each pixel ( Supplementary Fig. 3) was truncated to the fingerprint region (350-1750 cm −1 ) and subjected to cosmic spike removal (sensitivity: 0.4, spike width: 7) in the commercial MATLAB Cytospec toolbox 2.00.05, built 353 (Berlin, Germany). A principal component analysis based noise filter was used to reduce the spectral noise. Ten principal components, which explained over 99% of the variance ( Supplementary Fig. 4) within the spectral data, were used. Supplementary Figs. 5 and 6 compare the raw and filtered spectra for the AC and CC, respectively. The Table 1 The assignments of major mineral and organic Raman peaks observed in the Raman spectra from the deep articular cartilage, calcified cartilage and subchondral bone plate with their respective Raman shift. Amide I [49] spectral baseline caused by tissue autofluorescence was subtracted by fitting a third-order polynomial. Subsequently, the Raman spectra were vector-normalized to remove the differences in total Raman intensity due to the differences in laser focusing and optical properties of tissues. Apart from the cosmic spike removal step, all preprocessing steps of Raman spectra were performed using inhouse scripts written in MATLAB (MathWorks Inc., MA, USA). Examples of preprocessed mean spectra from the deep AC, CC, and SBP are shown in Fig. 1 B. Detailed Raman peak assignments are given in Table 1 and examples of raw spectral data for each sample are given in the Supplementary material: section 8, respectively.

Identification of tissue types
Classification of the different tissue types was performed with unsupervised cluster analysis, where spectra were classified based on their (dis)similarity or "distance." K -means clustering (KMC) [37] was used to segment the CC and SBP from the Raman maps. In particular, to identify the tidemark, hierarchical cluster analysis (HCA) [38] was used as it produces a dendrogram, which is useful in interpreting the relationships between clusters.
We presumed that CC and SBP can be identified from the Raman data exclusively based on their organic and/or mineral signatures as they have different organic matrices, which might have an effect on tissue mineralization [14 , 16 , 25] . To test this, the Raman spectra were split into two ranges and cluster analyses were performed separately on them. One range contained only organic vibrations (1144-1750 cm −1 ) while the other contained mostly mineral vibrations (350-1143 cm −1 ). To improve the sensitivity of cluster analysis, the dominant V 1 PO 4 3band (900-987 cm −1 ) along with some organic peaks (proline peaks between 800 cm −1 and 1020 cm −1 ) were excluded from the mineral vibrational range.
All Raman maps collected from the entire sample set were combined into one data matrix for KMC analysis. By increasing the number of clusters iteratively, the optimal number of clusters to distinguish different tissue types was found to be five and six for organic and mineral vibrational ranges, respectively. The generated false-colored cluster images were annotated using the acquired dark-field mosaic images as references. The mean spectra of each cluster were calculated to investigate spectral differences between the CC and SBP tissues in organic and mineral vibrations.
As described earlier, OA involves the advancement and duplication of the tidemark. We used HCA to identify these mineralization fronts from the Raman maps. All Raman maps containing the tidemark (60 μm × 60 μm regions collected with a step size of 1 μm) were combined into a single data matrix, and the mineral vibrational range (as described in KMC) were fed into the HCA algorithm. The dendrogram was created using Ward's algorithm [39] and the Euclidean distance metric. Four clusters were found optimal by analyzing the extended dendrogram structure (Supplementary Fig. 9).

Compositional analysis
The mineral-to-matrix ratio, type-B carbonate substitution, carbonate-to-matrix ratio, and mineral crystallinity were calculated to quantitatively study the mineral quality of the mineralized tissues ( Table 2 ). To understand the organic phases, the relative proteoglycan (PG) contents (normalized to organic matrix) and relative saturated and unsaturated lipid contents (normalized to organic matrix) were investigated. Furthermore, changes in these compositional Raman parameters of the tidemark region, CC, and SBP were examined as a function of histopathological OA severity.
Chemical maps were constructed by calculating the ratios of the area under the desired Raman bands for each pixel. The area under each band was calculated with respect to the local baseline within a given wavenumber range. The cluster images were used as a mask to identify pixels of different tissue types. Finally, for statistical analysis, the pixel-by-pixel calculated compositional parameters were averaged for the total pixels of each tissue type.

Histological assessment
The cartilage histopathology assessment system developed by the Osteoarthritis Research Society International (OARSI) [40] was used to assess OA severity in the osteochondral samples. Three experts assessed the OARSI grade (with subgrade) of each sample from Safranin-O stained sections. In the event of disagreement, consensus grading was conducted, and the consensus grade was used as the final grade of the sample.

Statistical analysis
The sample set was divided into three groups to observe how the composition changes from healthy to early stages of OA and to advanced stages. Samples in group 1 (healthy) had OARSI grades from 0 to 1.5 ( n = 5); group 2 (Early OA) had grades from 2.0 to 3.5 ( n = 8); and group 3 (Advanced OA) had grades from 4.0 to 4.5 ( n = 15). To account for inherent dependencies between samples from the same subject, comparisons between the grade groups were made with a linear mixed model [41] . For calculating statistical differences between the groups, the OARSI grade and compartment location ( e.g. , either medial or lateral) were set as fixed variables and the subject was set as a random variable. Furthermore, for calculating the differences between the tissue-specific compositional parameters, the tissue type ( e.g. , either CC or SBP) was set as a fixed variable and the subject was set as a random variable. Restricted maximum likelihood was used as the estimation method with the linear mixed model. All statistical analyses were performed using SPSS 25 software (IBM SPSS Statistics, SPSS Inc., Chicago, IL, USA).

K-means cluster (KMC) analysis results
The KMC analyses with the organic and mineral vibrational ranges are shown in Fig. 2 and Supplementary Fig. 10. The clusters were annotated using the optical microscopic dark-field images ( Fig. 2 A), where all the relevant features (tidemark, CC, SBP, and bone marrow space) can be identified. The comparisons between the cluster images and dark-field images indicate that the deep AC, CC, and bone tissues are more accurately identified using the organic vibrational range. Moreover, with the organic vibrational range, there is a clear transition between the CC and SBP. However, different mineralized layers between the deep AC and CC could be only detected using the mineral vibrational range.
The mean Raman spectra for each cluster are shown in Fig. 3 . In the organic vibrational range, the mean AC spectrum mainly differs from the CC spectrum in terms of the amide III, glycosaminoglycans (GAGs), and lipid peaks ( Supplementary Fig. 11). Compared with the mean SBP spectrum ( Fig. 3 D), the CC has higher intensity in GAGs, lipids, phenylalanine, and tyrosine peaks; and lower intensity in v (C -O-C), amide III, lipids, and amide I peaks. In the mineral range ( Fig. 3 C), major differences between six clusters were observed in the phosphate and carbonate bands, where intensities of V 2 PO 4 3 − , V 4 PO 4 3 − , and type-B carbonate bands are higher in the CC than in SBP.

HCA results
The chemical maps of the matrix-to-mineral ratio along with the HCA results of Raman maps (step size: 1 μm) are shown in Fig. 4 . Fig. 4 D shows all four mean Raman spectra of the respective clusters. The AC spectrum (purple) has the lowest intensity while the CC spectrum (yellow) has the highest intensity in the phosphate and type-B carbonate peaks among the clusters. The orange and blue spectra have intensities between those of the AC and CC spectra, indicating that they represent the transitional region between the deep AC and CC. This corresponds to a region containing

Integral
Commonly used V 1 PO 4 3and amide I bands are strongly sensitive to laser polarization in the confocal Raman microscopy system [28 , 32] . Moreover, V 2 PO 4 3band is mineral crystallite orientation-dependent [24] . As asymmetric V 4 PO 4 3 − band is less vulnerable to orientation effects [32] , it represents the mineral content. Owing to limited amount of validation studies, both amide III band (less susceptible to tissue orientation effects [32] ) and hydroxyproline + proline bands (collagen specific [28] ) represent the organic matrix. Integral Type-B carbonate symmetric stretch band normalized by the mineral band [23] . The carbonate is an imperfection in the apatite crystals and the carbonate-to-phosphate ratio reflects the mineral crystallinity [21 , 54] .

Table 3
Estimated marginal mean value with the standard error of the investigated compositional Raman parameters of the calcified cartilage and subchondral bone plate, along with the differences between the means with 95% confidence interval and p -values from the linear mixed models.

Raman Parameters
Calcified the tidemark as viewed earlier by optical microscopy. The identification of the tidemark region is also supported by the dendrogram structure with four leaves in Fig. 4 B, where the root node divides into two leaves indicating that the Raman dataset was first divided into two groups based on mineralization. One leaf represents nonmineralized tissue and the other represents mineralized tissue. The mineralized leaf again divides into three subleaves, and the chemical maps ( Fig. 4 A) clearly show that the yellow cluster (in Fig. 4 C) represents CC tissue. Thus, the remaining two clusters of two different mineralized layers represent the tidemark region.

Compositional analysis
KMC images obtained using the organic vibrations were used as a mask to identify pixels of the CC and SBP from the chemical maps (Supplementary Fig. 12). Various compositional differences were found between them ( Table 3 ). Both mineral-to-matrix ratios are greater ( p < 0.001) in the CC than SBP. The carbonate-tophosphate ratio is higher ( p < 0.001) in the SBP, while the mineral crystallinity is higher ( p < 0.001) in the CC tissues. Moreover, the organic matrix of CC has a relatively higher ( p = 0.005) PG contents compared with SBP.

Compositional alterations of the tidemark region, calcified cartilage, and subchondral bone plate as a function of histopathological OA severity
In the tidemark region, the V 4 PO 4 3 − /amide III ratio decreased ( p = 0.025), while the V 1 CO 3 2 − /V 4 PO 4 3 − ratio increased ( p = 0.023) in advanced OA from healthy group ( Table 4 ). The V 1 CO 3 2 − /amide III ratio also increased ( p = 0.004) while the relative saturated lipids decreased ( p = 0.023) in advanced OA compared with the healthy group.
In the CC, both mineral-to-matrix ratios tend to increase ( p = 0.095 and p = 0.083, respectively) in early OA stages, and then decrease ( p = 0.001 and p = 0.003, respectively) in advanced OA stages compared with early OA. Meanwhile, the carbonate-tomineral ratio tends to decrease ( p = 0.074) in the early OA stages, and increases ( p = 0.016) in advanced OA stages (with respect to early OA).
In the SBP, the V 4 PO 4 3 − /(hydroxyproline + proline) ratio tends to increase ( p = 0.08) in early OA stages and decreases ( p = 0.004) in advanced OA with respect to early OA, while the V 1 CO 3 2 − /V 4 PO 4 3ratio decreases in advanced OA from the healthy group.

Discussion
In this study, we applied Raman microspectroscopy to investigate human osteochondral samples with high spatial resolution from unfixed and hydrated tissue specimens. To enable tissue characterization, we implemented cluster analysis techniques to segment different tissue types from measured Raman microspectroscopic maps of the osteochondral junction. Without any a pri-ori information, the unsupervised identification of different tissues based on their composition allowed us to study chemical composition across this biomechanically crucial junction, and detect tissue-specific chemical changes at different histopathological OA stages. We found that the CC is always more mineralized than the SBP, and the mineralization in both tissue types tends to increase in early OA but decreases in advanced OA stages. Using only organic vibrations, the KMC successfully distinguished the deep AC, CC, and SBP from the Raman maps ( Fig. 2 ). The transitions between different tissue types were accurately identified with the organic vibrational range instead of the mineral vibrational range. The tidemark region ( Fig. 4 ) was identified in the Raman maps with 1 μm step size. This region contained the tidemark as identified by the dark-field microscopic images ( Fig. 2 A), and the mineralization of this region was different from the underlying CC ( Fig. 4 ). Thus, the use of Raman microspectroscopy in conjunction with cluster analysis techniques enables a finer definition of the boundary between mineralized and nonmineralized tissues compared with optical microscopy; hence, it provides the opportunity for a better understanding of the complex mass transport processes across this boundary.
The CC was found to be more mineralized than the SBP. Previous studies [14 , 16] have speculated that this mineralization difference may be due to the differences in the organic matrix, as the CC contains type-II collagen while the SBP contains type-I collagen. Unlike SBP, the organic matrix of CC contains more water; hence, it allows more extra-fibrillar space during the mineralization process, which might contribute to the enhanced mineralization. On the other hand, we also found that the carbonate-to-phosphate ratio is higher in the SBP than CC. This indicates that the mineral crystals have greater stoichiometric perfection in the CC than those in the SBP. This is further supported by the smaller full width at half maximum (FWHM) of the V 1 PO 4 3 − band, indicating higher crystallinity ( i.e., a measure of mineral crystal size and/or perfection [22] ) observed in the CC ( Table 3 ).
The organic matrix of CC was found to contain more PG than SBP. Besides attracting charged ions, and consequently water into the tissue, PGs play a pivotal role in biomineralization [22] . The cartilage mechanical properties like compressive and shear strength are positively correlated with the relative amount of PGs and collagen, respectively, as well as with the structural orientations and physical interactions between them [42] . Consequently, the difference found in relative PG contents (normalized to amide III band) between the CC and SBP could result in a difference in compressive strength. Given the positive association between the tissue stiffness and degree of mineralization [14 , 16 , 22 , 24] , it is reasonable to assume that the CC and SBP differ in their stiffness. However, it has also been reported that the SB is stiffer with a lower degree of mineralization compared with CC, and the correlation between the local stiffness and mineral content is significantly different between them [16] . In another study, the mean indentation modulus was reported to be the same for the mature human CC and SB despite changes in mineralization [14] .
The tissue-specific biochemical characterization of the osteochondral junction, especially the compositional differences found between the CC and SBP, could serve as critical benchmark parameters when designing biomaterials for osteochondral repair. In biphasic ( i.e., cartilage-bone) scaffold design for tissue engineering, integrating a physiologically-relevant calcified cartilage layer is often underemphasized [43] . The scaffolds should be designed to recapitulate the key composition of the CC interface, which will be instrumental in restoring integrated musculoskeletal tissue systems as well as physiological and biomechanical function.
Alterations in the compositional Raman parameters owing to histopathological OA severity were examined in the tidemark region, CC, and SBP ( Table 4 ). The results suggest that only saturated lipids changed in the tidemark region in early OA. Moreover, in conjunction with decreased mineralization, increased type-B carbonate substitution was observed in the advanced OA group (compared with the healthy group), indicating more imperfections in the apatite crystals in the tidemark region. The decrement in mineralization is typical of less mature tissue, suggesting recent mineral deposition during tidemark advancement [6] .
The present study only partially confirmed our hypothesis. We did not find any decrease in the mineralization of CC or SBP in advanced OA compared with the healthy group; however, we observed that the degree of mineralization in both tissues had a tendency to increase in the early stages of OA, and then decreased in the advanced stages. It is possible that if we included endstage tissue samples (OARSI 5.0-6.5), the level of mineralization would have been lower compared with healthy samples. If indeed, the degree of mineralization and tissue stiffness are interrelated [14 , 16 , 22 , 24] , then we propose that the stiffness of the CC and SBP increases in the early stages of OA owing to alterations in tissue quality, and later decreases because of increased volume fraction [3] . Therefore, an imaging biomarker accurately reflecting the degree of mineralization would provide information about mechanical changes and, therefore, early OA. The degree of mineralization is also related to bone formation by osteoblasts as well as bone matrix remodeling activity [13 , 44] . Thus, the mineralization changes observed in the present study comparing early to late OA could provide a chronological snapshot of the dynamics of remodeling in SBP, and perhaps even in CC [45] . For instance, it has been suggested that the increased mineralization in the CC in OA is a process similar to endochondral ossification [11] . In typical endochondral ossification for long bone formation, chondrocyte apoptosis in the early stages is considered to play a prominent role in matrix calcification [46] . Therefore, it can be proposed that increased mineralization in the CC in early OA is analogous to the early stages of such endochondral ossification process, where thickening of the CC and advance of tidemark occur in conjunction with chondrocyte death in the deep zones of the AC. Subsequently, in endochondral ossification for long bone formation, the bone front advances, and similar to the OA process, the cement line advances and diminishes the CC in the process. This sequence of events was also suggested in an earlier study [47] , which showed that the early stages of degeneration resulted in thickening of the CC, followed by thinning, and an advance of the cement line.
In this study, no changes were observed for the carbonate-tomatrix ratio in either the CC or SBP at any stage of OA ( Table 4 ). On the other hand, the carbonate-to-phosphate ratio changed in the different OA stages. This suggests that, in OA, the changes in Table 4 Descriptive statistics describing the differences between the estimated means with 95% confidence interval of the OARSI grade groups for every investigated compositional Raman parameters in the tidemark region, calcified cartilage, and subchondral bone plate. Results are displayed as the estimated mean value with standard error from the linear mixed models.    phosphate content ( i.e., the degree of mineralization) could have a greater impact on the changes in carbonate-to-phosphate ratio than carbonate content. The fact that there were no changes in carbonate content ( i.e., stoichiometric perfection of the mineral crystals) in OA is also supported by the lack of variations in the FWHM of the V 1 PO 4 3 − band. Finally, we did not observe any changes in the relative PG contents with OA severity.
There is one major limitation of this study. Owing to the challenges in obtaining intact cadaver (donor) specimens, we were limited to a small subject size ( N = 7 donors). A total of 28 osteochondral samples with no visible AC denudation ( i.e., OARSI grades between 0 and 4.5) were extracted from donors and used in this pilot study. Despite that, after histopathological assessment, only five samples were classified as healthy, which reduced the statistical power of the group-wise comparison and prevented us from drawing firm conclusions. For instance, the effect of sex on composition could not be examined because there was only one female donor. Additionally, the compartment-specific ( i.e. , medial or lateral) analysis could not be conducted as all the healthy samples were from the lateral compartment. In future studies, the OA-related compositional alterations of the osteochondral junction need to be investigated with a larger subject population. Another important limitation that might influence the results is the freeze-thaw cycle before sample preparation. This could cause degradation of tissues, which again may interfere with our Raman measurement results [48] . However, other sample preparations and the data measure-ment steps were kept constant for all samples. These shortcomings did not prevent us from interpreting our results to characterize the tissue-specific composition of the CC and SBP in different histopathological OA grades.
To conclude, we characterized the biochemical composition of different tissue types at the osteochondral junction as a function of histopathological degeneration, as well as the transition from the CC to the AC. Raman microspectroscopy was carried out on unfixed, hydrated human specimens. We observed that the CC was more mineralized, and the minerals had higher crystallinity than the SBP. The degree of mineralization in both tissues started to change from early OA; in advanced OA, the mineral crystals were rich in carbonate while the overall mineralization decreased.

Declaration of Competing Interest
SS reports grants from the European Research Council (Grant Agreement number 336267 ) and grants from the Academy of Finland (Grant numbers 268378 & 303786 ), during the conduct of the study. LR reports grant from the Academy of Finland (project number 310466 ) during the conduct of the study. Other authors have no disclosures in relation to this manuscript.