J. Biophotonics 7, No. 3–4, 189–199 (2014) / DOI 10.1002/jbio.201300149
Journal of
BIOPHOTONICS FULL ARTICLE
Attenuated Total Reflection Fourier Transform Infrared (ATR-FTIR) spectral discrimination of brain tumour severity from serum samples James R. Hands 1, Konrad M. Dorling 1, Peter Abel 2 , Katherine M. Ashton 3 , Andrew Brodbelt 4 , Charles Davis 3, Timothy Dawson 3 , Michael D. Jenkinson 4 , Robert W. Lea 2 , Carol Walker 4, and Matthew J. Baker*; 1 1 2 3
4
Centre for Materials Science, Division of Chemistry, JB Firth Building, University of Central Lancashire, Preston, PR1 2HE, UK School of Pharmacy and Biomedical Sciences, Maudland Building, University of Central Lancashire, Preston, PR1 2HE, UK Department of Pathology, Lancashire Teaching Hospitals NHS Trust, Royal Preston Hospital, Sharoe Green Lane North, Preston, Lancashire, PR2 9HT, UK The Walton Centre for Neurology and Neurosurgery, The Walton Centre NHS Foundation Trust, Lower Lane, Fazakerley, Liverpool, L9 7LJ, UK
Received 16 September 2013, revised 6 November 2013, accepted 28 November 2013 Published online 7 January 2014
Key words: serum, filtrate, glioma, rapid, spectroscopy, infrared, ATR-FTIR, cancer
Gliomas are the most frequent primary brain tumours in adults with over 9,000 people diagnosed each year in the UK. A rapid, reagent-free and cost-effective diagnostic regime using serum spectroscopy would allow for rapid diagnostic results and for swift treatment planning and monitoring within the clinical environment. We report the use of ATR-FTIR spectral data combined with a RBF-SVM for the diagnosis of gliomas (high-grade and low-grade) from non-cancer with sensitivities and specificities on average of 93.75 and 96.53% respectively. The proposed diagnostic regime has the ability to reduce mortality and morbidity rates.
Centrifugal filter (EMD Millipore, USA).
* Corresponding author: e-mail:
[email protected], Phone: +44 1772 893209
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Journal of
BIOPHOTONICS 190
J. R. Hands et al.: Serum spectroscopy gliomas
1. Introduction Malignant gliomas are among the more lethal of human cancers. In 2010, of the 9,156 newly diagnosed patients, brain cancers accounted for 2,689 deaths in men and 2,208 deaths in woman in the UK [1]. In the last three decades incidence and mortality rates have increased by 23% and 25% for men and women respectively. Brain cancer is the most common cancer for childhood deaths after leukaemia. In 2010, one-third of 217 children newly diagnosed with brain cancer died in England [2]. More than 120 different types of tumour can be found in the brain, of which, gliomas account for 30– 40% of all intracranial neoplasms [3]. The term ‘glioma’ refers to a primary site tumour which originates from glial cells (neuroglia) within the brain and central nervous system (CNS). Glial cells outnumber neuron cells and occupy approximately half of the brain. The two main glial cell types in the brain and CNS are astrocytes and oligodendrocytes. Malignant gliomas are graded according to the classification of the World Health Organisation (WHO). The three main types of malignant glioma are astrocytomas, ependymomas and oligodendrogliomas. A tumour with a mixture of the histological features present in the main three is known as a mixed glioma [4]. Table 1 shows the subtypes of high-grade and low-grade gliomas [5]. Survival rates of people with glial tumours vary depending on tumour type, histopathological grading, age and other patient specific diseases. On average, people diagnosed with glial tumours survive: 6– 8 years with low-grade astrocytomas or oligoendrogliomas; 3 years with anaplastic astrocytomas and 12–18 months for high-grade astrocytomas and Glioblastoma multiforme [6]. Age at diagnosis of a glial tumour affects survival rate after 5-years; children and adolescents (0–19 years) have a relatively good survival rate (58.1%) compared to young adults (20–39 years) and adults (40–59 years) where survival rates decrease steadily (52.8% and 19.1%, respectively). Older adults (60þ) have a poor survival rate of just 4.4% [7]. The majority of new brain cancer cases are malignant and brain cancer has high rates of mortality in adults and children [2, 8]. Table 1 The sub-types of WHO grade I–IV low-grade and high-grade brain tumors [5]. General Tumour Grade
WHO Grade Grade Sub-type
Low Grade
I II II III III IV
High Grade
Pilocytic astrocytoma Oligodendroglioma Astrocytoma Anaplastic astrocytoma Oligodendroglioma Glioblastoma
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Gliomas are primarily detected using a medical imaging technique such as computerised tomography (CT) and magnetic resonance imaging (MRI), amongst others [9]. Evidence of a tumour being present requires confirmatory diagnosis via biopsy or tumour resection [10]. A biopsy specimen is firstly acquired by drilling into the skull before the specimen is histopathologically examined for the presence of any malignancies [11, 12]. Histological grading does not provide clinicians with accurate prognostic and therapeutic details on an individual patient basis [13]. The current diagnostic regime requires histopathological interpretation of tissue specimens by a trained neuropathologist, thus pathological diagnoses are to some extent subjective [14]; Bruner et al. 1997 found that disagreements between original and review diagnoses in 214 of 500 cases, some 42.8% [15]. A study found that the inter-observer agreement was considerably better for specialised neuropathologists as opposed to general surgical pathologists when grading fibrillary astrocytomas, likely due to experience. The neuropathologists all agreed or had only one discrepant diagnosis in 86.7% versus the general surgical pathologists who had 43.3% in all thirty reviewed cases [16]. A rapid diagnostic serum test would be beneficial to both patients and clinicians whilst reducing current diagnosis times. A rapid serum screening regime would significantly reduce current diagnosis times and greatly increase the chance of successful treatment. Despite an increase in mortality rates for brain cancer, survival rates are also increasing due to early detection [2]. The ability to rapidly detect diseases, such as glioma, allows for treatment to be offered earlier and survival rates to increase. Previous studies have provided evidence of the benefits of applying spectroscopy to clinical problems [17, 18]. Backhaus et al. (2010) distinguished between breast cancer serum and healthy patient serum samples achieving a sensitivity of 98% and a specificity of 95%. The study also shown the ability to differentiate between breast cancer and other disease states (e.g. Alzheimer’s disease, coronary heart diseases and hepatitis C) using minute quantities of patient serum with a sensitivity of 98% and a specificity of 100% [19]. ATR-FTIR has been used successfully to discriminate between whole and ultrafiltrate serum samples obtained from patients suffering from one of the following; acute myocardial infarction, angina pectoris and other pains in the chest. This study achieved a sensitivity of 88.5% and a specificity of 85.1% in a blind validation test [20]. Disease states can be distinguished between using infrared spectroscopy; Scaglia et al. (2011) achieved a sensitivity of 95.2% and a specificity of 100% when using a leaveone-out cross-validation on spectra collected from normal patients (no hepatic fibrosis) and various degrees of hepatic fibrosis [21]. Our previous study has
www.biophotonics-journal.org
FULL ARTICLE J. Biophotonics 7, No. 3–4 (2014)
shown the diagnostic ability of gliomas from noncancer patients using serum samples with high sensitivities and specificities as high as 87.5% and 100% respectively [22]. Specifically to the spectroscopic grading of cancers a number of studies have been conducted. Steiner et al. (2003) was successfully able to distinguish between control (non-cancer tissue) and astrocytoma and glioblastoma tissue sections using infrared spectroscopy with high classification rates approaching 90% accuracy [23]. Glioma grading has also been shown to be suitable when deciding whether to continue or not continue with tumour resection based on infrared spectroscopic classification; Sobottka et al. (2008) found that FTIR could be used as an intraoperative tool in cerebral glioma surgery to help define tumour margins and detect high grade tumour residues with sensitivities and specificities as high as 100% and 96.9% respectively [24]. Gajjar et al. (2009) has shown the ability to differentiate between patients diagnosed with either ovarian or endometrial cancer from non-cancer controls (including patients with known benign gynaecological conditions) using blood and serum samples applied with ATR-FTIR spectroscopy with classification results as high as 96.7% for ovarian cancer and 81.7% for endometrial cancer [25]. We present the methodological studies implemented for ATR-FTIR diagnostic regime optimisation. These studies include the drying time of serum and spectra obtained, filtration and analysis of whole, 100 kDa (kilodalton), 10 kDa and 3 kDa serum filtrate aliquots from control (non-cancer), lowgrade glioma serum (e.g. astrocytoma) and highgrade glioma serum (Glioblastoma multiforme). Finally, a variance study was carried out to assess the consistency of the spectral data collected. ATRFTIR is a rapid, cost-effective and easy-to-use spectroscopic tool which has great potential as a screening and diagnostic tool for a range of diseases within the clinical environment [26].
2. Methods and materials 2.1 Serum samples Blood samples were collected from 49 patients diagnosed with a Glioblastoma Multiforme (GBM), 23 pa-
191
tients with a diagnosed low-grade glioma (astrocytoma, oligoastrocytoma, oligodendroglioma) and 25 normal (non-cancer) patients with no known health problems. The average age of the entire sample set is 54.62 years. Table 2 provides demographic details by cancer group. Samples were obtained from the Walton Research Tissue Bank and Brain Tumour North West (BTNW) Tissue Bank where all patients had given research consent. The research described in this paper was performed with full ethical approval (Walton Research Tissue Bank Application number: BTNW/WRTB 13_01, Applicant: M. J. Baker/BTNW Application number: 1108, Applicant: P. Abel). We investigated the use of ATR-FTIR spectroscopy and human patient serum to discover a novel process for the rapid diagnosis of gliomas. This research investigated the fundamental differences in ATR-FTIR spectra between serum and serum derived filtrate aliquots from patients diagnosed as having a low-grade glioma, high-grade glioma or non-cancer. All blood samples were taken pre-operatively. The serum tubes were left to clot at room temperature for a minimum of 30 minutes and a maximum of 2 hours from blood draw to centrifugation. Separation of the clot was accomplished by centrifugation at 1,200 g for 10 minutes and 500 ml aliquots of serum dispensed into prelabelled cryovials. Serum samples were snap frozen using liquid nitrogen and stored at 80 C.
2.2 Drying study Normal human mixed pooled serum (0.2 mm sterile filtered, CS100-100, purchased from TCSBiosciences, UK) was used in volumes of 1 mL to determine the optimal drying time necessary for quality spectral collection. Spectra were collected using a JASCO FTIR-410 spectrometer equipped with a Specac ATR single reflection diamond Golden GateTM at the University of Central Lancashire, in the range of 4000– 400 cm1 , at a resolution of 4 cm1 and over 32 coadded scans. Prior to each spectral collection, a background absorption spectrum was collected for atmospheric correction. One microlitre of pooled serum was pipetted onto the ATR-FTIR crystal and 3 spectra were collected at each 0, 2, 4, 8, 16 and 32 minute interval to ob-
Table 2 Number, age and gender data of patient samples from each tumour grade. Tumour Grade
Number of Subjects
Age range/mean age
Gender
Normal (Non-cancer) Low-Grade High-Grade
25 23 49
26–87/59.1 years 19–60.3/36.9 years 24.7–78.8/60.1 years
15 male, 10 female 11 male, 12 female 29 male, 20 female
www.biophotonics-journal.org
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Journal of
BIOPHOTONICS 192
J. R. Hands et al.: Serum spectroscopy gliomas
serve spectral changes during the drying process. The dried intimate serum film was removed from the crystal using absolute ethanol (purchased from Fisher Scientific, Loughborough, UK). The drying experiment was repeated multiple times to gain spectra representative at specific times during drying.
2.3 Variance study Normal human mixed pooled serum (0.2 mm sterile filtered, CS100-100, purchased from TCSBiosciences, UK) was used in the variance study. A volume of 1 mL was pipetted on to the ATR-FTIR single reflection diamond crystal and dried for 8 minutes, at which time 3 spectra were collected. Three spectra were collected per 1 mL and 50 different 1 mL spots were analysed. The dried serum spot was removed the crystal using absolute ethanol between each variance repeat with absolute ethanol. In total, 150 ATR-FTIR spectra were collected.
2.4 ATR-FTIR spectral diagnostic model All whole serum samples were thawed prior to spectral collection and 100 kDa, 10 kDa and 3 kDa filtration aliquots were prepared using Amicon Ultra0.5 mL centrifugal filters (purchased from Millipore Limited, UK) [Figure 1]. Centrifugal filters filter out components of the serum above the cut-off point of the filters membrane (i.e. 100 kDa), allowing components below the filter membrane cut-off point to pass through. Each whole serum sample (high-grade, low-grade and control) had a filtration aliquot prepared by pipetting 0.5 mL of the whole serum in to the filtration device and centrifuging at 14,000 rpm for; 10 minutes, 15 minutes, and 30 minutes for 100 kDa, 10 kDa and 3 kDa filter devices respectively. Spectra were collected in a random order within the serum sample sets. For each sample, a 1 mL serum spot was dried for 8 minutes on the ATR-FTIR crystal, at which time 3 spectra were collected. This procedure was repeated three times per sample. As a result, for each sample 9 spectra were collected. Prior to spec-
Figure 1 Centrifugal filter (EMD Millipore, USA).
tral collection, a background absorption spectrum was collected (for atmospheric correction) before the 1 mL was pipetted onto the ATR-FTIR crystal, thus a background was collected per serum replicate. The dried serum film was washed off the crystal in between each procedure using Virkon disinfectant (purchased from Antec Int., Suffolk, UK) and absolute ethanol. Spectra were acquired in the range of 4000– 400 cm1 , at a resolution of 4 cm1 and averaged over 32 co-added scans. In total, 3375 ATR-FTIR spectra were collected from all whole and filtration serum samples. Table 3 shows the total number of spectra and patients in each serum grade and filtration category. The number of patients reduces as serum filtrate aliquots are prepared due to serum availability.
2.5 Pre-processing selection data For each whole and filtration serum sample set, an identical approach was used to pre-process the spectral data, and to analyse using multivariate analysis methods. Firstly, to remove any bias from analysis models, the technical replicates from each sample were averaged so that each serum sample set contained three spectra from each patient; one average spectrum from each patient spot. Outliers were then removed from the spectral sets using a quality test. The fingerprint region (1800–900 cm1 ) was selected
Table 3 The number of spectra collected and number of patients (in brackets) for each filtrate composition for the range of cancer serum severities being analysed.
High-Grade Serum Low-Grade Serum Normal (Non-cancer) Serum Total
Whole Serum
100 kDa Serum
10 kDa Serum
3 kDa Serum
441 207 225 873
423 207 225 855
423 198 225 846
405 198 198 801
(49) (23) (25) (97)
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
(47) (23) (25) (95)
(47) (22) (25) (94)
(45) (22) (22) (89)
www.biophotonics-journal.org
FULL ARTICLE J. Biophotonics 7, No. 3–4 (2014)
for multivariate analysis. A principal component based noise reduction, using the first 30 principal components of the data, was performed on the spectra to improve the signal-to-noise ratio. Following this, all spectra were vector normalised and mean centred. The spectral data was also analysed using second derivative spectra of the data, but best overall results for PCA and SVMs were achieved using the noise reduction, vector normalisation and mean centring process. Principal component analysis (PCA) was performed on the pre-processed spectra, giving an unsupervised classification from which the loadings could be interpreted. Using LIBSVM code in MATLAB [27], an automatic n-fold cross validation was performed (where n ¼ 3) on the training data to find the best values for the cost and gamma functions. These values were then used to train the SVM in one-versus-rest mode using a randomly selected training set consisting of two thirds of the patient-associated spectral data. The remainder of the data, making up the blind test set, was then projected into the model, and confusion matrices were calculated giving an overall SVM classification accuracy based on the true and predicted data class labels. The spectral data was split independently at a patient level into 1/3 blind and 2/3 training where all spectra from one patient was either in the train set or blind set. Sensitivities and specificities were calculated for each SVM model and for each separate disease group. Three different test and blind data test sets at the patient level, so no patient spectrum was in the test set that was in the blind set., were used to provide a range of sensitivities and specificities for whole serum. Pre-processing and multivariate (MVA) analysis was carried out on the raw spectral data in Matlab (7.11.0 [R2010b] (The MathWorks, Inc. USA) using in-house written software.
3. Results 3.1 Drying study Figure 2 displays the typically observed ATR-FTIR spectral data from 1 mL of whole human serum over a range of 0–32 minutes during drying. The spectra have been offset for ease of visualisation. At room temperature (~18 C) 1 mL of serum has been found to dry after 8 minutes through repeat drying experiments. Effective spectral collection requires intimate contact between the serum sample and the ATRFTIR crystal to allow interaction with the evanescent field; this can be achieved by allowing the liquid serum sample to dry [28]. Drying allows the intensity of the bands to increase exponentially as swelling decreases,
www.biophotonics-journal.org
193
Figure 2 ATR-FTIR spectra showing 1 mL of human serum drying after 8 minutes.
thus reducing the distance between the reflecting interference (water) and the sample molecules [29]. Spectra observed after allowing the 1 mL of serum to dry for 8 minutes creating a dried film on the ATR-FTIR crystal is consistent with serum spectra presented by Backhaus et al. (2010) [19], Petrich et al. (2009) [20], Hands et al. (2013) [22], Gajjar et al. (2013) [25] and Liu et al. (2002) [30].
3.2 Variance study Figure 3 shows ATR-FTIR raw and pre-processed spectra in the region of 3900–900 cm1 (Figure 3a–b) and in the fingerprint region of 1800–900 cm1 (c–d). The four spectra display an average spectrum surrounded by a standard deviation (STD) error margin. The largest variance between the raw (unprocessed) spectral data was at 1637.27 cm1 (STD: 0.4209) in both analysed wavenumber regions. The smallest variance in the raw data was at 3735.44 cm1 (STD: 0.0038) between 3900–900 cm1 and at 1792.51 cm1 (STD: 0.0138) in the fingerprint region. Noise reduction (30 principal components) and vector normalization pre-processing methods were applied to the data to reduce the baseline and to smooth the data. The pre-processing methods significantly reduced the STD and variance of the spectral data. The largest raw data STD at 1637.27 cm1 was reduced from 0.4209 to 0.0043 (pre-processed) a difference of
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Journal of
BIOPHOTONICS 194
J. R. Hands et al.: Serum spectroscopy gliomas
Figure 3 Raw/unprocessed and processed spectral data for whole serum complete spectrums (3900–900 cm1 ) and fingerprint regions (1800–900 cm1 ).
195.9%. The smallest spectral variance STDs were reduced from 0.0038 to 0.00123 at 3735.44 cm1 and from 0.0138 to 0.0004 at 1792.51 cm1 . The average STD across the 3900–900 cm1 raw and pre-processed data was 0.0137 and 0.0015 respectively. The STD values of the raw spectra were low initially but were reduced further by implementing pre-processing methods. The reproducibility of spectral data using ATR-FTIR is high and exhibits minimal variance, especially after pre-processing.
3.3 ATR-FTIR spectral diagnostic model results 3.3.1 PCA and loadings Figure 4(A) shows a three-dimensional PCA scatter plot displaying the separation between normal (noncancer) [blue], low-grade cancer [red] and highgrade cancer [green] patient spectral data. Principal component 3 (PC3) on the x-axis shows good se-
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
paration between low-grade cancer and normal (non-cancer) data. PC2 on the y-axis shows good separation between high-grade cancer from low-grade and normal (non-cancer). Spectral data collected from the three serum grades in this study have been shown to successfully separate into their different spectral groups using PCA. Figure 4(B, C) are the loadings of PC1 and PC3. Table 4 shows the major spectral peaks and proposed biomolecular assignments responsible for PC1 loadings and Table 5 for PC3 loadings [31, 34–36]. The axes of the loading plots have positive and negative directions. The positive and negative peaks in the loading plots of PC1 and PC3 correspond to the spectral peaks that the PCA scatter plot is using to discriminate the groups within the patient spectral dataset. The major peaks of PC1 (Figure 4B) are –ve 1649 cm1 which is assigned as Amide I, +ve 1513 cm1 assigned as Amide II and –ve 1077 cm1 assigned as C––O stretch (DNA/RNA). The positive and negative peaks separate the low-grade cancer (red) and non-cancer (blue) from the high-grade (green) data points. Table 4 shows the proposed biomolecular assignments from the loadings of PC1.
www.biophotonics-journal.org
FULL ARTICLE J. Biophotonics 7, No. 3–4 (2014)
195
Figure 4 3D PCA scatter plot and PC1 and PC3 loadings from whole patient serum.
The major peaks of PC3 (Figure 4(C)) are þve 1655 cm1 and ve 1644 cm1 both which are assigned Amide I and þve 1589 cm1 assigned as Amide II. The major þve and ve peaks separate the low-grade from the non-cancer PCA data points and ve 1644 cm1 . Table 4 and 5 show the spectral assignments for PC1 and PC3 loadings respectively (Figure 4(B, C)). Table 5 shows the proposed biomolecular assignments from the loadings of PC3.
3.4 SVM results ATR-FTIR spectra from all whole serum and serum filtrate aliquot samples were analysed to investigate sensitivities and specificities possible on patient and spectral levels using SVM-RBF analysis. Three different test and blind spectral datasets were used to
provide a range of sensitivities and specificities for whole serum (Table 6). One test and blind spectral dataset was used to achieve sensitivities and specificities for the filtrate aliquots. The whole serum SVM diagnostic model achieved 93.75% sensitivity and 96.53% specificity (overall average) on a patient level. The whole serum dataset also achieved 92.61% sensitivity and 96.12% specificity on a spectral level. The best whole serum RBF-SVM diagnostic model misclassified three patients in the blind dataset; one low-grade patient and two high-grade patients. The blind dataset consisted of one average spectrum per 1 mL of patient serum analysed resulting in three spectra per patient in the dataset. If two or more spectra from the three averaged for each patient were classified as either high-grade, low-grade or non-cancer, then this is regarded as the diagnosis. In comparison to the whole serum SVM-RBF results, the 100 kDa (Table 7), 10 kDa (Table 8) and 3 kDa
Table 4 Major peaks and proposed biomolecular assignmens of PC1 loadings [Figure 3(B)]. Spectral assignments taken from [28, 32–34]. Direction Loading on PC 1 (B)
Wavenumber (cm1 )
Proposed Biomolecular Assignment
þve
1698
þve þve þve ve
1513 1445 1386 1649
ve
1619
ve ve
1553 1077
Amide I of antiparallel b-sheet/aggregated strand protein structures (C––O stretch (76%), C––N stretch (14%), CCN deformation (10%)) Amide II of proteins (b-pleated sheet structures) d N––H (60%), n C––N (40%) CH2 deformation of methylene group, lipids CH3 deformation, lipids Amide I of a-helical protein structures (C––O stretch (76%), C––N stretch (14%), CCN deformation (10%)) Amide I of proteins (b-pleated sheet structures) n C¼O (80%), n C––N (10%), d N––H (10%) Amide II of proteins (b-pleated sheet structures) d N––H (60%), n C––N (40%) C––O stretch, deoxyribose/ribose, DNA, RNA
www.biophotonics-journal.org
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Journal of
BIOPHOTONICS 196
J. R. Hands et al.: Serum spectroscopy gliomas
Table 5 Major peaks and proposed biomolecular assignments of PC3 loadings [Figure 3(C)]. Spectral assignments taken from [28, 32–34]. Direction Loading on PC 3 (C)
Wavenumber (cm1 )
Proposed Biomolecular Assignment
þve
1655
þve
1589
þve þve þve þve þve þve ve
1462 1416 1394 1361 1119 1040 1644
ve ve ve ve ve
1513 1505 1220 1006 935
Amide I of antiparallel b-sheet/aggregated strand protein structures (C––O stretch (76%), C––N stretch (14%), CCN deformation (10%)) Amide II (NH bend (43%), C––N stretch (29%), CO bend (15%), C––C stretch (9%), N––C stretch (8%)) CH2 stretch deformation of methylene group (lipids) CH3 asymmetric stretch (lipids) CH3 deformation, lipids CH3 stretch (symmetric) C––O stretch (antisymmetric), COH bend, lipids n C––O, deoxyribose/ribose DNA, RNA Amide I of a-helical protein structures (C––O stretch (76%), C––N stretch (14%), CCN deformation (10%)) C¼C stretch C¼C stretch C––C stretch, C––H bend Phenylalanine (ring breathing) C––C residue a-helix
(Table 9) serum filtrate aliquots did not achieve sensitivities or specificities as high. The 100 kDa diagnostic model achieved 69.05% sensitivity and 85.87% specificity on a patient level and the 10 kDa diagnostic model achieved a sensitivity and specificity on a patient level with 79.68% and 88.75% respectively. The 3 kDa dataset achieved the lowest sensitivities and specificities and on a patient level achieved 65.28% and 80.81% respectively. Table 10 shows the optimal Cost and Gamma values for the whole serum and serum filtrate SVM analysis where training accuracy is the correctly predicted spectra from the training set and total accuracy is the correctly predicted spectra from the blind set. Whole serum contains a range of biochemical components such as cytokines and chemokines which are redundant proteins that are secreted with
growth, differentiation, and the activation function that is involved in regulating and determining the nature of an immune response, such as with the development of cancer [32]. Cytokines have previously been shown to differ between serum from glioma and non-cancer patients [22]. The whole serum SVM-RBF diagnostic model achieved the greatest patient and spectral sensitivities and specificities as a result of the infrared spectral dataset exhibiting bands characteristic of all serum components such as cytokines and chemokines which have varied molecular weights [33]. Filtrate serum aliquots produced from whole serum samples achieved lower sensitivities and specificities as a result of filtering and removing serum biomolecular components above the filter cut-off range. Petrich et al. (2009) reported the significant differences in infrared spectra of dried
Table 6 Dataset 1 including range from dataset 2 and 3: Sensitivities and specificities for whole serum on patient and spectral levels for normal (non-cancer), low-grade and high-grade cancer with an overall average and overall range. Normal (%) Patient 100 sensitivity Patient 95.83 specificity Spectra 95.83 sensitivity Spectra 97.06 specificity
Normal Range (%) 75.00–100.00 95.45–100.00 78.26–95.83 95.45–100.00
Low (%) 87.5 100 86.36 100
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Low Range (%)
High (%)
High Range (%)
Overall Average (%)
Overall Range (%)
87.50–87.50
93.75
92.86–93.75
93.75
75.00–100.00
95.45–100.00
93.75
87.50–93.75
96.53
87.50–100.00
85.00–91.67
95.65
92.86–95.65
92.61
78.26–95.83
95.45–100.00
91.30
86.36–91.30
96.12
86.36–100.00
www.biophotonics-journal.org
FULL ARTICLE J. Biophotonics 7, No. 3–4 (2014)
197
Table 7 Sensitivities and specificities for 100 kDa serum filtrate on patient and spectral levels for normal (non-cancer), low-grade and high-grade cancer with an overall average.
Patient sensitivity Patient specificity Spectra sensitivity Spectra specificity
Normal (%)
Low (%)
High (%)
Overall Average (%)
50 95.45 54.17 94.12
57.14 95.45 61.90 94.12
100 66.7 93.75 67.44
69.05 85.87 69.94 85.39
Table 8 Sensitivities and specificities for 10 kDa serum filtrate on patient and spectral levels for normal (non-cancer), low-grade and high-grade cancer with an overall average.
Patient sensitivity Patient specificity Spectra sensitivity Spectra specificity
Normal (%)
Low (%)
High (%)
Overall Average (%)
85.71 85 75 80.33
70 100 66.67 98.25
83.33 81.25 78.38 78.72
79.68 88.75 73.35 85.77
Table 9 Sensitivities and specificities for 3 kDa serum filtrate on patient and spectral levels for normal (non-cancer), lowgrade and high-grade cancer with an overall average.
Patient sensitivity Patient specificity Spectra sensitivity Spectra specificity
Normal (%)
Low (%)
High (%)
Overall Average (%)
62.5 73.68 70.83 76.36
66.67 100 69.23 98.04
66.67 68.75 68.57 74.47
65.28 80.81 69.54 82.96
serum before and after filtration. A Wilcoxon test revealed that whole serum has more spectral regions which differentiate between serum samples from patients with acute myocardial infarction (AMI) from serum samples from non-AMI patients compared to 100 kDa and 10 kDa serum filtrates. The molecules responsible for the significant differences have a molecular mass below 100 kDa and 10 kDa, thus whole serum spectra has more areas of significance when differentiating disease states. The amide-II region exhibited statistically significant differences between the diseased and non-diseased groups [20]. The results achieved show that the combination of multivariative support vector machine analysis and spectroscopic profiles of the molecular vibra-
tions within 1 mL dried patient serum sample films on a ATR-FTIR crystal can be used to diagnose and differentiate between high-grade, low-grade and non-cancer with high sensitivities and specificities. Once a serum sample has been collected from a patient a result can be obtained within 10 minutes of serum deposition onto the ATR-FTIR crystal.
4. Conclusion The results show that the combination of support vector machine analysis and ATR-FTIR spectroscopic profiles of the molecular vibrations within
Table 10 The optimal Cost and Gamma values for the whole serum and serum filtrate.
Optimal Cost (C) Optimal Gamma (g) Training Accuracy SVM Total Accuracy
www.biophotonics-journal.org
Whole Serum
Whole Serum (2)
Whole Serum (3)
100 kDa Data
10 kDa Data
3 kDa Data
22.63 4 85.86 % 96.875 %
32 5.66 87.96% 86.46%
22.63 8 86.46% 91.58%
2048 0.85 72.58 % 79.57 %
2048 16 90.80 % 79.76 %
2048 16 78.53% 75.64 %
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Journal of
BIOPHOTONICS 198
1 mL dried patient serum sample films on an ATRFTIR crystal can be used to diagnose and differentiate between high-grade, low-grade and non-cancer with high sensitivities and specificities. Once a serum sample has been collected from a patient a result can be obtained within 10 minutes of serum deposition onto the ATR-FTIR crystal. This work has shown that a 1 mL spot of serum on an ATR-FTIR crystal can dry to a state where excellent spectra can be collected, consistent with serum spectra presented in other research studies [20–22, 28–29]. The raw spectra collected from 1 mL serum spots after 8 minutes of drying have a low standard deviation which is further reduced with noise reduction and vector normalisation pre-processing methods. The variance between the spectral data after pre-processing is as little as 0.0015 for spectra collected from 3900–900 cm1 . The reproducibility of spectral data using ATR-FTIR is high and exhibits minimal variance especially after pre-processing. We have demonstrated that the combination of mid-infrared spectral data collected from whole serum samples from high-grade, low-grade and normal (non-cancer) patients and multivariative support vector machine (SVM) statistical analysis has the ability to be implemented within the clinical environment as a rapid, reagent-free and cost-effective diagnostic regime. Serum filtrate aliquots in this study (100, 10 and 3 kDa) have been shown to not achieve as high sensitivities and specificities as whole serum; filtration removes key biomolecules which exhibit spectral bands that are involved in the MVA diagnostic process. The loadings from PC1 and PC3 (Figure 4(A, B)) show the peaks which are responsible for the separation between high-grade, low-grade and non-cancer spectral datasets. The major peaks from PC1 shows that Amide I, Amide II and C––O stretch (DNA/RNA) bands are responsible for the separation between non-cancer and low-grade cancer from high-grade cancer PCA data points. The major peaks from PC2 shows that Amide I and Amide II are responsible for the separation of lowgrade cancer from non-cancer PCA data points. We have presented a method which has the ability to diagnose gliomas from whole patient serum ATR-FTIR spectral data to sensitivities and specificities as high as 100.00% and 95.83% respectively. The described spectroscopic methodology has the potential to be used in monitoring the therapeutic response of the cancer treatments given to brain tumour patients. A blood sample collected from a patient being treated for brain tumour could be collected and analysed using ATR-FTIR. The spectral data can then be used to determine whether the patient is responding well to the treatment or whether the patient would benefit from an alternative treatment plan. We plan to collect ATR-FTIR spectral data from sera obtained from patients diagnosed
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
J. R. Hands et al.: Serum spectroscopy gliomas
with metastatic tumours of the brain from various primary sites and meningioma. Gajjar et al. (2009) has shown the ability to differentiate between patients diagnosed with either ovarian or endometrial cancer from non-cancer controls (including patients with known benign gynaecological conditions) using blood and serum samples applied with ATR-FTIR spectroscopy. The ability to detect a disease state using patient sera and ATR-FTIR has many advantages such as early diagnosis, thus applying this methodology to a larger sample set including metastatic sera may enable the site of origin to be distinguished [25]. The ability to not only diagnose a metastatic brain tumour, but also to be able to determine the primary site from a serum sample using ATR-FTIR spectroscopy would have many benefits for both patients and clinicians. The development of an ATR-FTIR spectroscopic prognostic index would assist clinicians when using the proposed regime in the clinical environment where a prognosis is required. The diagnostic regime proposed over the course of this paper describes a rapid diagnostic methodology which would greatly reduce current diagnosis times and mortality rates, thus revolutionising the clinical environment. Acknowledgements The authors would like to acknowledge the support of Brain Tumour North West, The Sydney Driscoll Neuroscience Foundation, The Centre for Materials Science at the University of Central Lancashire and Dr Alex Henderson, from the University of Manchester, for providing the necessary statistical codes required to analyse the ATR-FTIR spectral data. Author biographies online.
Please see Supporting Information
References [1] Cancer Research UK, Brain, other CNS and intracranial tumours Key Facts. 2013. [Accessed on 11/08, 2013]; Available from: http://www.cancerresearchuk. org/cancer-info/cancerstats/keyfacts/brain-cancer/ brain-cancer. [2] Office for National Statistics, Brain cancer rates have risen by a quarter over the past three decades. 2013. [Accessed on 11/08, 2013]; Available from: http:// www.ons.gov.uk/ons/rel/vsob1/cancer-statistics-registrations–england–series-mb1-/no–41–2010/sty-braincancer-awareness.html. [3] K. Swanson, M. Simon, A. Deutsch, C. Schaller, and H. Hatzikirou. Math Models Methods Appl Sci 15(11), 1779–1794 (2005). [4] Cancer Research UK. Types of primary brain tumours. [Accessed on 11/08, 2013]; Available from:
www.biophotonics-journal.org
FULL ARTICLE J. Biophotonics 7, No. 3–4 (2014)
[5] [6]
[7]
[8] [9]
[10] [11]
[12] [13]
[14] [15] [16]
[17] [18]
[19]
[20]
http://www.cancerresearchuk.org/cancer-help/type/ brain-tumour/about/types-of-primary-brain-tumours. W. Chen, J Nucl Med 48(9)1468–1481 (2007). J. Y. Ljubimova, M. Fugita, N. M. Khazenzon, A. Das, B. B. Pikul, D. Newman, K. Sekiguchi, L. M. Sorokin, T. Sasaki, and K. L. Black, Cancer 101(3), 604–612. E. Crocetti, A. Trama, C. Stiller, A. Caldarella, R. Soffietti, J. Jaal, D. C. Weber, U. Ricardi, J. Slowinski, and A. Brandes, Eur J Cancer 48(10), 1532– 1542 (2012). S. Kesari and C. D. Stiles, Neuron 51(2), 151–153 (2008). M. Tate, C. M. S. Berger, M. C. Tate, and M. S. Berger, Primary Brain Tumors and Cerebral Metastases in: Cancer Metastasis (2011), Cambridge University Press. R. Stupp, J. C. Tonn, M. Brada, and G. Pentheroudakis, Ann Oncol 21(5), 190–193 (2010). C. Chen, C. P. W. Hsu, T. W. Erich Wu, S. T. Lee, C. N. Chang, K. C. Wei, C. C. Chuang, C. T. Wu, T. N. Lui, Y. H. Hsu, T. K. Lin, S. C. Lee, and Y. C. Huang, Clin Neurol Neurosurg 111(10), 835–839 (2010). J. Vaquero, R. Martinez, and M. Manrique, Surg Neurol 53(5), 432–437 (2000). R. V. Iyer, A. Golash, R. W. Lea, C. H. G. Davis, and P. Roberts, Diagnosis and treatment of malignant glioma in: The Molecular and Cellular Pathology of Cancer Progression and Prognosis, 379–408 (2006). C. Krafft, S. B. Sobottka, G. Schackert, and R. Salzer, Analyst, 129(10), 921–925 (2004). J. M. Bruner, L. Inouye, G. N. Fuller, and L. A. Langford. Cancer 79(4), 796–803 (1997). R. A. Prayson, D. P. Agamanolis, M. L. Cohen, M. L. Estes, B. K. Kleinschmidt-DeMasters, F. Abdul-Karim, S. P. McClure, B. A. Sebek, and R. Vinay, J Neurol Sci 175(1), 33–39 (2000). K. M. Dorling and M. J. Baker, Trends Biotechnol 31(8), 437–438 (2013). M. J. Baker, E. Gazi, M. D. Brown, J. H. Shanks, N. W. Clarke, and P. Gardner. J Biophotonics 2(1–2), 104–113 (2009). J. R. Backhaus, Mueller, N. Formanski, N. Szlama, H.G. Meerpohl, M. Eidt, and P. Bugert, Vib Spectrosc 52(2), 173–177 (2010). W. Petrich, K. B. Lewandrowski, J. B. Muhlestein, M. E. Hammond, J. L. Januzzi, E. L. Lewandrowski, R. R. Pearson, B. Dolenko, J. Fruh, M. Haass, M. M.
www.biophotonics-journal.org
199
[21]
[22]
[23]
[24]
[25]
[26] [27]
[28] [29] [30] [31]
[32] [33] [34] [35]
[36]
Hirschl, W. Kohler, R. Mischler, J. Mocks, J. OrdonezLlanos, O. Quarder, R. Somorjai, A. Staib, C. Sylven, G. Werner, and R. Zerback, Analyst 134(6), 1092– 1098 (2009). E. Scaglia, G. D. Sockalingum, J. Schmitt, C. Gobinet, N. Schneider, M. Manfait, and G. Thiefin, Anal Bioanal Chem 401(9), 2919–2925 (2011). J. R. Hands, P. Abel, K. Ashton, T. Dawson, C. Davis, R. W. Lea, A. J. McIntosh, and M. J. Baker, Anal Bioanal Chem 405(23), 7347–7355 (2013). G. Steiner, A. Shaw, L. P. Choo-Smith, M. H. Abuid, G. Schackert, S. B. Sobattka, W. Steller, R. Salzer, and H. H. Mantsch. Biopolymers 72(6), 464–471 (2002). S. B. Sobottka, K. D. Geiger, R. Salzer, G. Schackert, and C. Krafft. Anal Bioanal Chem 393(1), 187–195 (2009). K. Gajjarr, J. Trevisan, G. Owens, P. J. Keating, N. J. Wood, H. F. Stringfellow, P. L. Martin-Hirsch, and F. L. Martin, Analyst 138(14), 3917–3926 (2013). K. M. Dorling and M. J. Baker, Trends Biotechnol 31(6), 327–328 (2013). C. L. C. Chang, LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology. 2011. [Accessed on 05/07, 2013]; Available from: www.csie.ntu.edu.tw/~cjlin/ libsvm. S. G. Kazarian and K. L. Chan, Analyst 138(7), 1940– 1951 (2013). E. Goormaghtigh, V. Raussens, and J. M. Ruysschaert, Biochim Biophys Acta 1422(2), 105–185 (1999). K. Z. Liu, R. A. Shaw, A. Man, T. C. Dembinski, and H. H. Mantsch, Clin Chem 48(3), 499–506 (2002). M. J. Baker, C. Clarke, D. Demoulin, J. M. Nicholson, F. M. Lyng, H. J. Byrne, C. A. Hart, M. D. Brown, N. W. Clarke, and P. Gardner, Analyst 135(5), 887– 894 (2010). L. C. Borish and J. W. Steinke, J Allergy Clin Immunol 111(2), 460–475 (2003). L. R. Grinwald, Chemokine Research Trends, 2007, New York: Nova Science Pub Inc. A. D. Meade, F. M. Lyng, and P. Knief, Anal Bioanal Chem 387(5), 1717–1728 (2007). D. E. Maziak, M. T. Do, F. M. Shamji, S. R. Sundaresan, G. Perkins, and P. T. T. Wong, Cancer Detect Prev 31(3), 244–253 (2007). M. Meurens, J. Wallon, J. Tong, H. Noel, and J. Haot, Vib Spectrosc 10(2), 341–346 (1996).
# 2014 by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim