Rapid intraoperative molecular genetic classification of gliomas using Raman spectroscopy

Abstract Background The molecular genetic classification of gliomas, particularly the identification of isocitrate dehydrogenase (IDH) mutations, is critical for clinical and surgical decision-making. Raman spectroscopy probes the unique molecular vibrations of a sample to accurately characterize its molecular composition. No sample processing is required allowing for rapid analysis of tissue. The aim of this study was to evaluate the ability of Raman spectroscopy to rapidly identify the common molecular genetic subtypes of diffuse glioma in the neurosurgical setting using fresh biopsy tissue. In addition, classification models were built using cryosections, formalin-fixed paraffin-embedded (FFPE) sections and LN-18 (IDH-mutated and wild-type parental cell) glioma cell lines. Methods Fresh tissue, straight from neurosurgical theatres, underwent Raman analysis and classification into astrocytoma, IDH-wild-type; astrocytoma, IDH-mutant; or oligodendroglioma. The genetic subtype was confirmed on a parallel section using immunohistochemistry and targeted genetic sequencing. Results Fresh tissue samples from 62 patients were collected (36 astrocytoma, IDH-wild-type; 21 astrocytoma, IDH-mutated; 5 oligodendroglioma). A principal component analysis fed linear discriminant analysis classification model demonstrated 79%–94% sensitivity and 90%–100% specificity for predicting the 3 glioma genetic subtypes. For the prediction of IDH mutation alone, the model gave 91% sensitivity and 95% specificity. Seventy-nine cryosections, 120 FFPE samples, and LN18 cells were also successfully classified. Meantime for Raman data collection was 9.5 min in the fresh tissue samples, with the process from intraoperative biopsy to genetic classification taking under 15 min. Conclusion These data demonstrate that Raman spectroscopy can be used for the rapid, intraoperative, classification of gliomas into common genetic subtypes.

Diffuse gliomas are the most common and aggressive primary tumors of the central nervous system in adults. The molecular genetic stratification of gliomas is important as more evidence emerges of the predictive and prognostic implications of different genetic subtypes. Surgical resection plays a key role in the management of patients with gliomas, and there is increasing evidence that the greater the extent of surgical resection, the greater the patient's progression-free and overall survival. [1][2][3] The diffuse nature of these tumors means that the surgeon can never achieve complete tumor resection. At present, it has been demonstrated that the clinical benefit of the extent of tumor resection does depend on the specific genetic subtypes. 2 Information on the tumor genetic subtype at the time of surgery, providing real-time predictive and prognostic information, would allow the surgeon to tailor their surgical strategy to each patient's genetic profile.
The 2016 World Health Organization (WHO) classification of gliomas introduced molecular and genetic subtyping integrated with the histological glioma type and grade 4 ( Figure 1). Common and clinically relevant genetic subtypes include gliomas with mutations of the isocitrate dehydrogenase (IDH) genes 1 or 2 with or without chromosome 1p and 19q (1p/19q) codeletion. 5 Within most neuropathology laboratories, 1p/19qcodeletion is established using the cytogenetic technique of fluorescence in-situ hybridization (FISH). The presence of the most common IDH mutation (IDH1 R132H) can be established using immunohistochemistry (IHC) with an antibody highly sensitive and specific for this

Importance of the Study
In this era of personalized medicine and the development of treatments targeting specific glioma genetic subtypes, it will become increasingly important for surgeons to be aware of the genetics of a tumor at the time of operation to inform their surgical strategy and to balance the survival benefit of increasing surgical resection with the risk of causing neurological deficit. In this study, we demonstrate that Raman spectroscopy can accurately and rapidly classify gliomas according to isocitrate dehydrogenase mutations and 1p/19q-codeletion. The protocol can be carried out within a neurosurgical theatre without the need for laboratory input and results are available in under 15 min. shown, only common groups; ** these are provisional categories in the classification: it is now recognized that most of these neoplasms are incipient or undersampled glioblastomas, isocitrate dehydrogenase-wild-type 11 ).

Neuro-Oncology Advances
IDH mutation, 6 which accounts for around 85%-90% of the mutations in gliomas. 7 If there is suspicion of a rare IDH mutation, the sample would undergo targeted IDH-1/ IDH-2 genetic sequencing.
Raman spectroscopy is a vibrational spectroscopy technique that probes the unique molecular vibrations of a sample to characterize its molecular composition. 8 The significant advantage of Raman analysis is that no or minimal sample processing is required, allowing for rapid and cost-effective analysis of tissue without the need for specialist staff, time-consuming laboratory protocols, or expensive consumables.
The main aim of this study was to evaluate the ability of Raman spectroscopy to identify the most common molecular genetic subtypes of diffuse glioma in the neurosurgical theatre setting. An important objective was to develop a simple protocol, which would allow the surgeon or scrub staff to undertake Raman analysis of tissue in the operating theatre without the need for specialized technicians or laboratories. We have also evaluated the application of Raman analysis in the neuropathological diagnostic workflow of glioma tissue, including analysis of cryosections, formalin-fixed paraffin-embedded (FFPE) sections, and glioma cell lines.

Study Participants
Ethical approval for the use of fresh, cryosection, and FFPE tissue was obtained through the Oxford Brain Bank generic Research Ethics Committee approval (15/SC/0639). All tissue used was deemed surplus to diagnostic requirements.

Pathological Diagnosis
All cases were reviewed by a consultant neuropathologist, and diagnoses were established using the European Association for Neuro-Oncology guidelines, 9 which incorporates the updated (2016) WHO classification of tumors of the central nervous system. 4 See Supplementary material for details of histopathological diagnostic protocols. The WHO integrated diagnosis was based on the IHC/genetic results and histological grade (Figure 1), giving following 5 possible diagnostic groups: 1. diffuse or anaplastic astrocytoma, IDH-mutant (grade II or III); 2. glioblastoma, IDH-mutant (grade IV); 3. oligodendroglioma or anaplastic oligodendroglioma, IDH-mutant, and 1p/19q-codeleted (grade II or III); 4. diffuse or anaplastic astrocytoma, IDH-wild-type (grade II or III); 5. glioblastoma, IDH-wild-type (grade IV).

Tissue Processing
For fresh tissue samples, 1 biopsy per patient was taken by the neurosurgeon from an area of tumor bulk defined both visually and using the intraoperative neuronavigation system (Brainlab, Germany). The biopsy was immediately transferred to the neuropathologist, who cut the tissue biopsy into 2 pieces: 1 for intraoperative smear evaluation, and the adjacent piece for Raman analysis. The portion for Raman analysis measured approximately 3 mm 3 . The sample was immediately transferred to a grooved stainless steel slide 10  surface onto which the Raman laser could be focused and compacting the tissue to limit the effect of tissue drying. The adjacent biopsy underwent intraoperative smear preparation with hematoxylin and eosin (H&E) and was reviewed by the consultant neuropathologist to confirm the presence of tumor. Details of cryosection, FFPE, and LN18 cell line preparation can be found in the Supplementary material. Two LN18 IDH-mutant and 2 IDH-wild-type cell lines were cultured and processed separately at 2 different time points. Cell line batches, once harvested, were processed as either cryosections or FFPE sections, with protocols to match the respective tissue sample preparations.

Raman Spectroscopy
Raman spectra were collected using a Renishaw bench-top RA800 series spectrometer (Renishaw plc., UK). For fresh tissue samples, single point spectra were obtained, widely distributed across the smeared sample, aiming for at least 100 spectra per fresh tissue sample. Two accumulations of either 0.5 s or 1 s depending on the amount of background fluorescence and resulting detector saturation for that sample. The total time to acquire data from each area was taken from the spectrometer log. The time from surgical biopsy to completion of spectral collection was also recorded. See Supplementary material for further details.
The second model was a 3-group model, stratifying gliomas into the 3 main genetic subtypes ( Figure 1B) regardless of histological grade: All astrocytomas, IDH-wild-type were glioblastomas (grade IV)-there were no grade II or III astrocytomas, IDH-wild-type tumors; these are often undersampled glioblastomas. 11

Statistical Analysis and Model Building
All data analysis was carried out using MATLAB R2015b (MathsWorks). For each genetic subtype, normalized (using Extended Multiplicative Scatter Correction 12 ) mean spectra and the negative of the second derivative 13 were plotted against the wave number shift. Statistical differences between peak intensities were evaluated using the Wilcoxon rank-sum test, and the magnitude of the resulting P value was used as an indicator of the magnitude of the difference between the intensities. Tentative molecular assignment of the identified Raman peaks was carried out based on the published literature. [14][15][16] See Supplementary material for further details of spectral preprocessing and analysis.
Principal component analysis (PCA) fed linear discriminant analysis (LDA) was used to build the pathology classification models, which used all the spectra, after preprocessing, from each patient. In this classification method, PCA is used to reduce the data dimensionality, with the analysis of variance used to select only those PCs that describe a statistically significant difference in the data. An LDA model is then built from those significant PCs. Model performance was assessed using leave-one-patient-out cross-validation to generate model sensitivity and specificity (Supplementary material). Multiclass receiver operating characteristic (ROC) curves were generated with the area under ROC curve (AUROC) calculated as a measure of accuracy.

Results
A total of 62 fresh tissue samples were included in the study. No cases were excluded on the grounds of the absence of tumor on the adjacent H&E analysis of the smeared section. Ninety-seven cryosections and 132 FFPE sections were cut for analysis. Of these, 18 cryosections and 12 FFPE sections were excluded after review of the parallel H&E section showed no evidence of tumor. A total of 79 cryosections and 120 FFPE sections were used for analysis. Table 1 shows the number of samples analyzed in each genetic subgroup for each sample preparation method.
A total of 9799 spectra were collected in the fresh tissue group, with a mean of 161 spectra per sample. The mean time for data collection was 6 min and 6 s per sample, and the whole process from intraoperative biopsy to completion of Raman spectroscopy was under 15 min in every case. See Supplementary material for cryosection and FFPE details). Table 2 summarizes the sensitivities and specificities for 3-group and 2-group prediction of glioma genetic subtypes using PCA-LDA classification modeling of the Raman spectra with leave-one-patient-out cross-validation. With fresh tissue samples, the models were able to classify tumors into either IDH-mutant or IDH-wild-type (2-group model) with 0.91-0.95 sensitivity and specificity and 0.98 accuracy. Fresh tissue classification into the 3 genetic subtypes (3-group model) was performed with 0.79-0.94 sensitivity, 0.89-1.00 specificity, and an accuracy of between 0.92 and 0.98 ( Figure 3). Cryosection and FFPE section models classified with slightly less sensitivity and specificity than the fresh tissue models as shown in Table 2.

Neuro-Oncology Advances
The overall accuracy of these models remained high, with cryosection 3-group classification accuracy of 0.

Mean Spectra and Peak Assignment
For the fresh tissue samples, mean spectra and the negative of the second derivative of the mean spectra for each genetic subgroup in the 3-group model are shown in On all second-derivative plots, the wave numbers of prominent Raman peaks are labeled and the statistical significance of the difference between the peak intensities of each genetic subtype in that model is indicated (P < .001, Wilcoxon rank-sum test).
The most dominant Raman peaks in the fresh tissue sample are seen at 1445 cm -1 and 1660 cm -1 . These are both a summation of multiple overlapping peaks and represent the ratio of protein and lipid content within the sample.

Neuro-Oncology Advances
differences between oligodendroglioma and IDH-wildtype astrocytomas were more pronounced than between oligodendroglioma and IDH-mutant astrocytomas.
There is consistency between the dominant Raman peaks across the different tissue preparations. Depending on the model, between 12 and 17 peaks of the 43 peaks identified are present in all 3 tissue preparations. Of these, 11 peaks are also present in both LN18 preparations (cryosections and FFPE sections). In general, cryosection and FFPE section spectra have more prominent DNA and amino acid peaks and less contribution from lipid and protein peaks. The cryosection spectra have a significant background in the 1200-1500 cm -1 region, which is thought to be due to water condensation on the objective lens as the cryosection dries out. There is also a strong contribution from a prominent peak at 1336 cm -1 (assigned to protein:lipid ratios) in this wave number region, which is not seen in the fresh tissue spectra. In terms of differentiating between genetic subtypes, peaks that consistently give the most significant differences between subtypes across the 3 tissue preparations are those assigned to tyrosine (635 and 649 cm -1 ), phenylalanine (1003 cm -1 and 1036 cm -1 ), lipid (825 cm -1 and 853 cm -1 ), protein (1551 cm -1 ), and lipid:protein ratios (1445 cm -1 :1551 cm -1 ). In the 3-group models including oligodendroglioma, peaks for DNA (725 cm -1 , 963 cm -1 , and 1206 cm -1 ) are also consistent across all sample preparations.

Rare IDH
A total of 13 rare IDH mutations were present overall, in 3 fresh tissue samples, 6 cryosections, and 4 FFPE sections ( Table 3). All cases were correctly classified into the IDHmutated group (using the 2-group model), with the highest probability of correct classification occurring in the fresh tissue samples, with a mean probability of 91% for predicting IDH mutant.

LN18 Cells
All LN18 IDH-mutant cell lines stained strongly for R132H IDH mutation on IHC. Mean spectra and the negative of the second derivative are shown in Supplementary Figure 3. The mean spectra for the LN18 cells and key Raman peaks, which differentiate genetic subtypes correlate with the tissue preparation spectra. Table 4 gives the sensitivity,

Discussion
We have demonstrated that Raman spectroscopy can determine the most common genetic subtypes of adult diffuse gliomas using fresh tissue straight from the operating theatre, cryosections, and FFPE sections. Most significantly, our fresh tissue workflow requires no sample processing, and results are available in under 15 min, making this technique ideal for rapid intraoperative diagnosis.
Surgery plays an important role in the treatment of gliomas. There is mounting evidence that the greater the extent of surgical resection the longer the progression-free and overall survival. [1][2][3] The aim of surgery is to remove all the contrast-enhancing tumor in high-grade tumors or all the T2-FLAIR signal in low-grade gliomas, but, due to their diffuse nature, it is known that tumor cells will have infiltrated beyond these boundaries. Removal of the brain surrounding the tumor may result in neurological disability for the patient. The crucial balance for the surgeon is to maximize resection of the tumor whereas minimizing the risk of disability. Surgeons vary widely as to how aggressively they pursue tumor resection and the techniques they use (such as awake surgery and white matter tractography) to define this tumor:functional boundary. This is reflected in a wide range of 15%-70% complete resection rates reported in the literature. 2,17 In this era of personalized medicine, the ability to perform accurate genetic subtyping of gliomas at the time of surgery will allow the surgeon to adapt their surgical strategy with some knowledge of the likely prognostic and treatment potential for the lesions. In this study, we have shown that Raman spectroscopy can accurately identify the 3 most common and clinically relevant glioma groups. The presence of an IDH mutation is a strong predictor of longer progressionfree survival [18][19][20] and overall survival 18-25 when compared with IDH-wild-type tumors. 1p/19q-codeletion in the context of an IDH mutation is the genetic hallmark of oligodendrogliomas 26 and is a strong favorable prognostic factor for overall survival 20,23,27 and a strong predictor of response to radiochemotherapy. [28][29][30][31] More relevant to intraoperative decision-making, Wijnenga et al. 2 examined the impact on overall survival of the extent of surgical resection, measured as volume of residual tumor seen on postoperative contrast magnetic resonance imaging, in molecularly stratified low grade gliomas (WHO grade II astrocytomas and oligodendrogliomas). They reported

Neuro-Oncology Advances
that the effect of less tumor residual volume on overall survival was much greater in IDH-mutant astrocytomas compared with oligodendrogliomas. In the IDH-mutant astrocytoma group, even a small amount of residual tumor of 0.1 cm 3 affected negatively on survival compared with 0 cm 3 residual. This was not seen in the oligodendroglioma group, with a small residual having no significant effect on overall survival. Similar finding was reported by Kawaguchi et al. 32 who found a significant survival advantage of gross total resection over subtotal resection in IDH-mutant astrocytoma but no significant survival advantage for oligodendroglioma or IDH-wild-type astrocytoma. Consequently, pushing the boundaries of surgical resection when there is a possible increased risk of disability might be in the patient's best interest if they have an IDHmutant astrocytoma but would not be in their interest if they have an oligodendroglioma: intraoperative information of the glioma genetic subtyping is therefore essential. As we learn more about glioma subtypes and develop novel therapies tailored for their treatment, including the insertion of therapy agents at the time of surgery, the need for accurate genetic subtype information at the time of surgery will increase. We, and others, have shown that preoperative non-invasive magnetic resonance spectroscopy at 3-or 7-Tesla may provide the neurosurgeon with information on IDH status of diffuse gliomas 33 ; however, to the best of our knowledge, this currently does not allow distinction between IDH-mutant astrocytomas and oligodendrogliomas.
One aim of our study was to build up the training dataset for model development and therefore true intraoperative classification was not possible until enough data had been acquired. Once a classification model was developed, samples could be analyzed and classified immediately intraoperatively. This was piloted in the last 5 cases collected and yielding similar classification performances to the results presented. The major advantage of Raman spectroscopy compared with conventional techniques is the lack of sample preparation required. We developed a protocol that could easily be replicated by anyone in the operating theatre. The fresh tissue samples simply require the biopsy to be smeared into a grooved slide and then placed on the Raman microscope stage (Figure 2). In our study protocol, the sample was split into 2 sections by the neuropathologist before Raman analysis, as adjacent intraoperative or permanent tissue staining was essential to establish the  quality of the biopsy and tumor content, and spectral collection was supervised by a researcher. These steps would not be required as part of a Raman diagnostic pathway, particularly with the straightforward automation of the process, which, at the press of a button, would perform spectrometer quality control, laser autofocusing, spectral acquisition, preprocessing, and classification. This sets Raman spectroscopy apart from other techniques that have been used to acquire intraoperative genetic subtyping, such as polymerase chain reaction-based rapid genotyping assays. 34,35 More traditional techniques can give genetic stratification within the timeframe of an operation and are more accurate than Raman spectroscopy, but they all require complex laboratory-based sample processing and analysis by qualified laboratory technicians. They have significant consumables costs per test, compared with the minimal cost of the grooved stainless steel slide used for Raman analysis, which after the procedure could be sterilized, reused, and become part of the standard neurosurgical instrument set. Others have reported the use of Raman and attenuated total reflectance-Fourier transform infrared spectroscopy spectroscopy to distinguish IDH-mutant versus IDH-wild-type but have used snap frozen sections before spectroscopy, which still requires laboratory-based sample preparation. 36,37 These vibrational spectroscopy studies have not attempted to distinguish oligodendrogliomas as a separate group, which is essential for clinical decision-making. Raman spectroscopy has also been used intraoperatively in the form of a probe device to attempt to differentiate tumor boundaries. 38 A probe device is essential for boundary detection but adds no benefit to the aim of this study: as the tumor is being removed by the surgeon anyway, in-or near-operating theatre analysis is sufficient and allows for more sensitive Raman spectrometry than would be achievable via a fiber optic probe setup. A barrier to the translation of Raman spectroscopy into the clinical setting is that clinicians are skeptical about the reliance on a change in spectral patterns to define different pathological entities. The spectral characteristics of glioma found in this study add to a body of published evidence, which consistently defines similar key spectral features of glioma tissue. 15,16,36,39,40 Moreover, the most significant spectral differences seen between genetic subgroups are broadly consistent with our understanding of the biology of these tumors. IDH-mutant and IDH-wild-type gliomas are known to have significant differences in lipid metabolism, 41-44 collagen maturation, 45 and DNA methylation, 7,46 which correlates with the most differentiating Raman peaks assigned to lipids, phenylalanine (involved in collagen synthesis), and DNA. The 881 cm -1 peak is markedly different in the fresh tissue oligodendroglioma spectrum. We currently do not have an experimentally validated explanation for this; however, although superficially similar (regarding high levels of 2HG as a result of an IDH mutation), IDH-mutant astrocytomas and oligodendrogliomas are otherwise clinically and molecularly distinct. It is also known that certain metabolic features are associated with the 1p/19q codeletion rather than IDH mutation (eg, cystathionine 47 ). As altered phospholipid metabolism is known to occur in subsets of glioma, 48 we cannot rule out that the increased peak at 881 cm -1 reflects these metabolic changes rather than an increased amount of DNA. It is important to note that this study did not aim to investigate proteomic or metabolomic profiles of different genetic subtypes-there are more suitable techniques to do this than Raman spectroscopy.
There is considerable consistency of Raman spectral peaks across all of the different sample preparation methods and cell lines used in this study. As might be expected, the performance of the classification model decreases as the tissue becomes more manipulated by the sample preparation, with fresh tissue models performing better than cryosection, which in turn classifies better than FFPE sections. Spectra of cryosections and FFPE sections have fewer contributions from proteins and lipid, differentiating genetic subtypes more on DNA and amino acid peaks. This is consistent with the known effect of protein denaturing due to freezing, cross-linking of proteins as part of formalin fixation, and the loss of lipid due to paraffin embedding. 49 The LN18 cell line classification was very accurate within individual cell line preparations. In the cryosection preparations, the performance dropped off when the model includes spectra from both cell lines. This may be the result of large intragroup variability, as there are only 2 groups of each mutation and wild-type cells in the model. The performance would be likely to increase when the number of cell line batches is much greater, causing a decrease in intragroup variability and an increase the intergroup variability.
This study has also shown that Raman spectroscopy has application in the neuropathology and research laboratory settings, offering rapid genetic subtyping of gliomas using cryosections, FFPE sections, and glioma cell lines. The minimal sample preparation and rapid analysis advantages of Raman spectroscopy are less critical within a laboratory setting. However, routine IHC testing for IDH mutation is only sensitive to the common IDH1 mutation, 7 whereas Raman spectroscopy measures the downstream biological consequences of the mutation, most likely associated with the intratumor accumulation of 2-hydroxyglutarate, which is common to all glioma-associated IDH1 and IDH2 mutations. 50 In our study, the IHC test for the common IDH1-R132H mutation was, as expected, negative in those cases harboring rare IDH1 or IDH2 mutations (Table 3). This amounted to a significant proportion of our cohort (5%) and all cases were correctly classified as IDH-mutant in the Raman analysis. High-throughput assays based on next-generation sequencing or microarray methylation chip techniques are likely to replace single marker assessment by IHC, Sanger sequencing, or FISH, but these are not currently widely available and have a 5-to 14-day turnaround. Raman analysis may provide a faster and lowercost alternative, certainly for initial tissue screening in environments where access to molecular diagnostics may not be available.
In summary, this study has demonstrated the potential for Raman spectroscopy to be used in a straightforward way within the neurosurgical and neuropathological glioma diagnostic pathway to provide genetic subtyping of gliomas within 15 min of the biopsy. The resulting data give the surgeon the opportunity to tailor the surgical strategy to the patient's specific tumor genetics. Despite this being the largest study of its kind published to date, the small sample Neuro-Oncology Advances size is a limitation. Further multicenter and larger studies need to be undertaken to expand the training data set and to evaluate the effect of inter-user and inter-machine variability on the predictive model. Finally, the predictive model needs to be tested with a prospective independent cohort of patients to robustly validate the technique.

Supplementary Material
Supplementary data are available at Neuro-Oncology Advances online. Keywords brain tumor | intraoperative | isocitrate hydrogenase | Raman spectroscopy | personalized medicine.

Funding
The The views expressed are those of the authors and not necessarily those of the National Health Service (NHS), the NIHR or the Department of Health. This work uses data provided by patients and collected by the NHS as part of their care and support and would not have been possible without access to this data. The NIHR recognizes and values the role of patient data, securely accessed and stored, both in underpinning and leading to improvements in research and care. Spectroscopic instrumentation was loaned by Renishaw plc. (UK).