Discovery of Biomarkers for Tasmanian Devil ... - ACS Publications

Subscriber access provided by CORNELL UNIVERSITY LIBRARY

Article

Discovery of biomarkers for Tasmanian devil cancer (DFTD) by metabolic profiling of serum Naama Karu, Richard Wilson, Rodrigo Hamede, Menna Jones, Gregory M Woods, Emily F. Hilder, and Robert A. Shellie J. Proteome Res., Just Accepted Manuscript • DOI: 10.1021/acs.jproteome.6b00629 • Publication Date (Web): 06 Sep 2016 Downloaded from http://pubs.acs.org on September 6, 2016

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Journal of Proteome Research is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

Discovery of biomarkers for Tasmanian devil cancer (DFTD) by metabolic profiling of serum Naama Karu*1˄, Richard Wilson*2, Rodrigo Hamede3, Menna Jones3, Gregory M. Woods4, Emily F. Hilder1†, Robert A. Shellie1‡ 1

Australian Centre for Research on Separation Science (ACROSS), School of Physical Sciences,

University of Tasmania, Private Bag 75, Hobart TAS, 7001 Australia; 2Central Science Laboratory (CSL), University of Tasmania, Private Bag 74, Hobart TAS, 7001 Australia; 3

School of Biological Sciences, University of Tasmania, Hobart, Australia; 4Menzies Institute

for Medical Research, University of Tasmania, Hobart, Australia ^ current address: The Metabolomics Innovation Centre (TMIC), University of Alberta, Department of Biological Sciences, CCIS 5-185 Edmonton, AB, T6G 2E9 Canada. †current address: Future Industries Institute, University of South Australia, Mawson Lakes Campus, GPO Box 2471, Adelaide SA, 5001 Australia. ‡current address: Trajan Scientific and Medical, 7 Argent Place, Ringwood VIC, 3134 Australia. *Corresponding authors: Naama Karu, ACROSS, Department of Chemistry, School of Physical Sciences, University of Tasmania, Sandy Bay 7005, Australia. Email: [email protected]; Richard.Wilson, Central Science Laboratory (CSL), University of Tasmania, Sandy Bay 7005, Australia. Email: [email protected], Phone: +61-362261718. Fax: +61-3-62262494.

1 ACS Paragon Plus Environment


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Abstract Devil facial tumour disease (DFTD) is a transmissible cancer threatening Tasmanian devils (Sarcophilus harrisii) with extinction. There is no pre-clinical test available for DFTD, thus our aim was to find biomarkers for DFTD by metabolic fingerprinting. Seventy serum samples from wild Tasmanian devils (35 controls, 35 with tumours) were analysed by Liquid Chromatography - High Resolution Mass Spectrometry. Features were selected by multivariate models (PLS/DA, random forests) comparing age-matched training set (n = 20 x 2), and further complying with fold-change threshold (≥ 1.4) and Mann-Whitney U-tests with correction for multiple hypotheses (false discovery rate, FDR q < 0.05). An array of overlapping peptide segments of the N-terminal end of fibrinogen were the strongest positive DFTD markers. These peptides recorded fold-change up to 90, FDR-corrected p value below 0.01 and area under ROC curve of at least 0.80, and also correlated with tumour size (Spearman R > 0.45, p < 0.01). Additional potential markers included amino acid and lipid metabolites, while cortisol and urea were most significant health predictors (AUC ≥ 0.90). PLS/DA resulted in AUC = 0.997 for the training set, and overall sensitivity of 91% and specificity of 97%. A support vector machine model utilising only the major peptide marker and seven other metabolites led to overall 94% sensitivity and specificity. The novel findings in this first DFTD metabolomics study shed light on metabolic changes in Tasmanian devils affected by DFTD and provide a valuable step towards development of prognostic biomarkers.

Keywords Metabolomics, serum, Tasmanian Devil, Sarcophilus harrisii, DFTD, cancer, fibrinostatin


Page 2 of 41

Page 3 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Introduction The Tasmanian devil (Sarcophilus harrisii) is a medium-sized marsupial carnivore now endemic to the island of Tasmania, Australia. The wild population of the Tasmanian devils has declined by more than 80% due to the fatal devil facial tumour disease (DFTD)1, 2, first observed in 1996 in north-eastern Tasmania2. DFTD is a cancer of Schwann cell origin. A second morphologically similar but genetically distinct facial cancer has been recently discovered2, 3. Along with canine transmissible venereal tumour (CTVT) in dogs and recently discovered cancers in freshwater clams4, 5, they are the only known naturally-occurring clonally transmissible cancers6. DFTD is transmitted by direct inoculation of live tumour cells when susceptible and infected individuals bite each other7. After an unknown latent period (estimated 3-12 months based on epidemiological models) lesions develop and rapidly grow, and extensive metastases were observed in very late stages in 65% of euthanized animals in one study2. DFTD cells are generally not recognised by the host immune system, and there is ongoing effort to characterise the DFTD tumours8 and to induce an immune response to DFTD9, 10, with the goal of understanding the devil—tumour interaction and the potential for evolution and for developing a vaccine. The present study is the first non-targeted metabolomics analysis of serum from Tasmanian devils, in search of a unique DFTD metabolic fingerprint with the dual aims of expanding our understanding of the disease pathology and for future development of prediction models for the disease. Only basic biochemical benchmarks were recently established for Tasmanian devils in DFTD and health11, and no available information currently exists on variation between the metabolome of DFTD and healthy devils. Of the applied approaches in systems biology, metabolomics is rarely performed in wild animals.



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The challenges in wildlife metabolomics studies are well documented12,

Page 4 of 41

13

, and include

limited sample availability, behaviour-introduced sampling bias dictated by trapping of animals, sample collection at ambient conditions, varied / unknown nutritional status, and seasonal variation, among others. All of the above factors should be acknowledged as potential limitations in our study. In most cases, metabolomics studies of animal bio-fluids or tissues involve species that are used as animal models for disease, as reviewed by Griffin et al.14). In human cancer research, the majority of metabolomics studies aim to provide prediction models for early diagnosis of various cancer types (see examples15-19 and comprehensive reviews20,

21

). Providing a metabolic background for the devils’ tumour

disease will expand the knowledge-base and direct further research in the omics field and targeted biochemical studies of specific analytes.

Experimental Procedures Chemicals All aqueous solutions were prepared with ultra-pure water having a specific resistance of 18.2 MΩ/cm, purified on a Millipore (Bedford, MA, USA) Milli-Q system. The deuterium-labeled standards (Mass Spectrometry internal standards) kynurenic acid d5 (MW 194.0791), 4-Pregnen-17a-ol-3,20-dione d8 (17α-hydroxyprogesterone d8; MW 338.2778) and octanoyl-L-carnitine d3 (MW 290.236) were purchased from C/D/N Isotopes (Quebec, Canada). All analytical standards for optimisation of electrospray ionisation settings and determination of peak identity were purchased from Sigma (Milwaukee, WI, USA).


Page 5 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Animals Samples were collected from wild Tasmanian devils trapped in a population in northwest Tasmania (West Pencil Pine, 15 km west of Cradle Mountain) during four 10 day trips per year for 3 years (2013-2015), conducted every 3 months in February, May, August and November. This was under the approval of the university of Tasmania animal ethics committee (permit number A0013326). Blood (1.0 mL per sample) was collected from the subcutaneous ear vein by using a sterile Precision Glide needle (0.8mm x 25mm). Blood was immediately transferred into an Eppendorf vial and stored in a portable cooler with ice blocks at approximately 40C. Serum and plasma was separated by centrifugation (1,200 x G for ten minutes) within 6-24 hours after collection. After centrifugation sera was immediately stored at -200C and kept frozen until sample preparation for metabolomics experiments. Sample preparation Serum samples were prepared according to a metabolomics protocol published by Dunn et al.22 with variations. Serum samples were thawed at 10°C for 1 h, then mixed by vortex for 2 s and aliquots of up to 100 µL (depending on availability) transferred into a 0.5 mL microcentrifuge tube. Ice-cold methanol (analytical grade, Merck) was added 3:1 (v/v) to the tube and mixed by vortex for 15 s, then stored at 10°C for up to 20 min until centrifugation. The samples were then centrifuged at 16600x g for 15 min at 10°C and the supernatant carefully transferred into a 1.5 mL microcentrifuge tube. The samples were then stored at -20°C for up to 3 days until drying by reduced-pressure centrifugal evaporator without heating. The dried samples were stored at -20°C for up to a week, then batches of samples were thawed for 1hr and reconstituted in solution at the same volume of original



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

serum sample taken. The reconstitution solution was MS-grade acetonitrile : water (5:95 v/v) with 0.1% formic acid (identical to HPLC separation start conditions; all from Merck), also containing 100 ng/L of three deuterated standards: kynurenic acid d5, 4-Pregnen-17aol-3,20-dione d8 and octanoyl-L-carnitine d3. The reconstitution solution was prepared in bulk, aliquoted and frozen, then thawed as necessary for sample batch preparation. Reconstitution was followed by a quick mix by vortex, then precipitation of non-solubilised particles by centrifugation at 16600x g for 15 min at 15°C. The supernatant was transferred into an HPLC glass insert within an HPLC vial, sealed and stored at -20°C until analysis. Samples were thawed at room temperature for 30 min, followed by a brief mix and air bubble purge. A QC sample was prepared by taking 20 µL each from 40 serum samples (representing different health status and trips out of the 140 available samples), and then proceeding with sample preparation as depicted above. The QC bulk sample was aliquoted and stored at -20°C until analysis per batch. LC-MS analysis The LC-MS system was a Waters Alliance 2690 HPLC coupled to a Thermo Scientific LTQ-Orbitrap XL mass spectrometer. The HPLC separation column was an Agilent ZORBAX SB-C18 (3.0 x 150 mm; 3.5 µm particles) equipped with a SB-C18 4.6 µm guard cartridge and maintained at 30°C. The Waters Autosampler was set to 10°C, and the injection volume was 10 µL. The LC solvents were 0.1% (v/v) formic acid in Milli-Q water (Solvent A) and 0.1% (v/v) formic acid in HPLC-grade acetonitrile (Solvent B) at a flow rate of 0.4 mL/min. The mobile phase flow was split post-column using a minimum-void T-piece (Upchurch), to provide a stream of approximately 0.15 mL/min to the mass spectrometer. A multi-step gradient was applied over a 25 min run: Isocratic 5% B for 2 min, a linear ramp to 30% B


Page 6 of 41

Page 7 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


over 10 min, a linear ramp to 98% B over 5 min, hold time of 3 min before returning to start conditions within 1 min and re-equilibrating for 4 min.

The mass spectrometer was

operated in electrospray ionisation (positive ion) mode with a capillary temperature of 275°C, sheath gas flow set to 26au and without auxiliary gas. The capillary voltage and tube lens voltage were set to -20 V and -70 V, respectively, and an electrospray voltage of 4.2 kV used. Centroid data were acquired using the FT analyser over the m/z scan range of 90-800 u at a target resolution of 30,000. Approximately 800 scans were collected over each 25 min run. Data analysis Thermo Orbitrap raw files were converted to mzxml files using MSconvert freeware (ProteoWizard 3.023) in accordance with the requirements of the metabolomics platform XCMS 24. XCMS online was utilised for feature detection, retention time correction and alignment. To focus on health status rather than within-class biological variability and noise, the XCMS analysis was set to include peaks appearing above a threshold in at least half of the samples in one of the experimental classes. Sample-wise normalisation by median-fold was conducted as part of the data pre-analysis by XCMS, and was assessed for its correction for batch effect and instrumental drift according to normalised box plots including the QC samples. All peaks were manually examined on XCMS and the Xcalibur Qual Browser platform (Thermo Xcalibur 2.1, Thermo Fisher Scientific, US) to exclude noise and other malintegrated peaks, and also to validate adducts and isotopes. Prepared data files were further processed using Microsoft Excel™ prior to statistical analysis.

Statistical analysis and

plotting were performed using XLStat software (version 2015.6.01, Addinsoft, France), Metaboanalyst (The University of Alberta, Canada17) and INVEX (The University of British Columbia, Canada25). Prior to analysis, peaks below intensity threshold were replaced with 7 ACS Paragon Plus Environment


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

values corresponding to 20% of lowest peak in matrix, and data were log2-transformed to promote normal distribution. Data centring and Pareto-scaling was conducted before multivariate analyses, apart from modelling for feature selection. Unsupervised methods such as PCA and hierarchical analysis were used to overview the data, inspect the replicates of a pooled quality control (QC) sample and identify potential outliers. For a subset of 40 age-matched samples, partial least-squares-discriminant analysis (PLS/DA) and Random Forests (RF) were utilised to rank the importance of features for discriminating between the experimental classes. SVM (support vector machine learning) model was utilised to construct a minimal model for class prediction based on selected features. Each of the models was evaluated internally by permutation tests and also cross-validations, to assist in minimising the risk of false discoveries due to “over-fitting” the model, especially of importance when working with a small sample. Mann–Whitney U test was used to compare between the classes, and was adjusted for multiple comparisons by the false discovery rate (FDR). The univariate performance of selected features was evaluated for log-transformed data by the non-parametric receiver-operator characteristic (ROC) analysis, calculating area under ROC curve (AUC) with 95% confidence interval based on cross validation performance (Monte Carlo method). Spearman rank correlation test was utilised to reveal significant relationships between metabolites, age and also to support adduct/fragment annotations. 2-way ANOVA was used to search for interactions between metabolite levels and the factors sex, age group and season. Chemical identity elucidation Putative identification of features was conducted by searching m/z values in three databases: Metlin24, HMDB26 and Lipid Maps27. The physicochemical properties that affect


Page 8 of 41

Page 9 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


retention mechanism were also taken into account, as well as the biological likelihood to be found in serum of wild carnivores. Then, the chemical formula was compared to the elemental composition calculated in Thermo Xcalibur Qual Browser, allowing restricted ppm deviation (± 1.2 ppm). Additionally, the mass spectrum was evaluated for expected adducts, fragments (if known) and isotope abundance (13C etc.). For more common endogenous compounds, peak identity was verified against authentic standards. MS2 experiments were used to provide fragmentation patterns for comparison with spectra libraries, chemical standards, or for protein/peptide identification as follows. Acquired peptide precursor masses and MS/MS spectra were searched against the Sarcophilus harrisii protein database (UniProt txid9305; 25,346 entries) using PEAKS® Studio software (version 7.5; Bioinformatics Solutions, Inc.,Waterloo, Canada) according to the following parameters: enzyme (none), instrument (Orbitrap), fragmentation (CID) parent ion and fragmention ion tolerances (10 ppm and 0.02Da, respectively) and variable modifications (oxidation). Database matches were considered signification at a -log10 p value > 20. Putative identification of peptides that did not produce sufficiently high quality MS/MS spectra was based on matching the observed and calculated peptide masses with a mass tolerance of +/- 2 ppm.

Results and Discussion Sample properties The 70 serum samples (from 53 Tasmanian devils) were divided into two classes: 1) 35 devils that were considered DFTD asymptomatic, with no signs of tumours for at least 6 months following the examined sample; 2) 35 devils with facial tumours. Of the DFTD group, 33 had a visually-confirmed DFTD (DFTD score of 5 ,definite)2, one possibly with DFTD (score of 4,high probability DFTD, but not counted as DFTD here), and one with tumour



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 41

which is less likely to be a DFTD (score 2) . The sample characteristics are summarised in Table 1, including the progression of the disease according to the state of the visible DFTD tumours (metastases are expected in advanced cases but their status is unknown). The sample selection in a study involving wild population of animals is limited by natural availability, and in Tasmanian devils this is exacerbated by the different age distribution between healthy and diseased, with only a few samples available from healthy adult devils (≥ 2 years of age), while in healthy wild population life expectancy is usually about 5-6 years. Most devils with DFTD trapped in the total examined population were 2-3 years old, and the youngest was 16 months old. Since age in mammals affects many aspects of metabolism (amino acids, nucleosides, hormones etc.), and the study aim is to find metabolites which are markers for DFTD with minimal confounding effect of age, initial metabolite selection was conducted in a subset of age-matched devils. To match the ages between healthy and DFTD animals as best as possible, we chose only animals between 16 months (dispersed juvenile) to 25 months old (first sample taken as an adult). The age-matched training set for feature selection comprised 20 devils with DFTD and 20 healthy controls, with the age median slightly higher in the DFTD class (Table 1). Student’s t-test rejected the significance of differences in the mean age between the two classes (p = 0.12) and Fisher’s F-test confirmed that the variances were similar (p = 0.67). Of the age-matched DFTD group, 16/20 samples were from animals confirmed to be within the first six months of visible DFTD, although some have already presented with mid-late progression of the DFTD.

In terms of

sex matching, the existing samples posed a limitation, and the groups could not be well balanced. Nevertheless, due to sample size, stratification by sex was not practical, and differentiating between female and males devils in the context of DFTD is beyond the scope of the presented work. 10 ACS Paragon Plus Environment

Page 11 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Analytical performance A total of 745 peaks were found under the instrumental conditions used and following XCMS pre-analysis. The maximal retention time deviation was under 20 s, and the mass accuracy deviation approximately 1 ppm. Median-fold normalisation was applied and improved the intensity distribution (Figure S-1), and one technical (chromatography) outlier was removed prior to sample selection. The deuterated internal standards showed acceptable variation in levels across the samples, with RSD of < 20% in the 19 replicates of the QC sample, and below 30% in the 70 samples. Peaks with high coefficient of variance (RSD > 30%) in the QC replicates were removed (20% of 745 peaks), and the peak list was further reduced to 232 after thorough examination of chromatograms and spectra, with stringent removal of chemical and instrumental noise, isotopes and low-abundance adducts. Based on these features, an unsupervised principle component analysis (PCA) obtained for the 70 samples and 19 QC replicates showed good clustering and centring of the QC samples, with no indication of any obvious outlier (Figure S-2). Multivariate analysis in training set To specify the metabolic variation associated with DFTD, a combination of statistical approaches was applied, utilising the age-matched samples (n = 20 x 2). The multivariate classification methods were based on variable importance ranking, known to reduce the risk of bias due to multicollinearity. While PLS/DA is very common in metabolomics studies, it is considered susceptible to noise and over-fitting. Therefore random forests was utilised as an additional method, and being non-parametric and unaffected by the feature scale, it is somewhat complementary to PLS/DA. Score plots obtained in multivariate analysis such as PCA or PLS/DA are often used to convince of the discriminating power of the model,



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

however their non-linearity prevents further interrogation and comparison with other studies16. To provide complementary evaluation of the predictive power, receiver operating characteristic (ROC) curve analysis was applied for multivariate as well as univariate analyses. A PCA model obtained for the age-matched samples utilising 232 features revealed only a tendency for separation between the two classes, while a PLS/DA model clearly separated between the DFTD and control samples (Figure S-3). The predictive accuracy using two latent variables was 0.847 (cross-validated probabilities of 5/40 misclassified: 4 DFTD samples and 1 control), the model fitness capability R2Y (cum) was 0.79, and the predictive capability Q2 (cum) was 0.49 after 100 cross validations. The model significance was validated via a permutation test with 1000 iterations resulting in p = 0.009. ROC analysis resulted in a very high AUC of 0.911 (95% confidence interval of 0.76 - 1). Most variables of high importance (variable importance in the projection, VIP > 1.5) and low coefficient of variance were features which dramatically increased in DFTD, and perhaps demonstrate the limitations of this model. To complement these results, a random forests (RF) model was generated, growing 1000 trees with 12 predictors tested per node (Figure S4). The RF model resulted in out-of-bag (OOB) error of 0.125 (4 DFTD samples and 1 control misclassified), and AUC of 0.927 (95% CI 0.776-1). Selection of differentiating features First consideration for inclusion in the list of potential biomarkers was high rank in the multivariate models. The PLS/DA model included 55 features with VIP ≥ 1, which mostly overlapped with highly ranked features in the RF model (according to mean decrease in accuracy of classification). Next, univariate analysis was conducted to further reduce the list of potential biomarkers to include features that individually changed between the classes


Page 12 of 41

Page 13 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


above a 30% threshold. A volcano plot (Figure S-5) found 86 features with Mann-Whitney Utests p < 0.05 and at least 1.3 fold change between the classes (n = 20 x 2), that include 57 features with FDR q < 0.05. The results from the two multivariate models and the univariate analysis were combined, yielding 40 features, summarised in Table 2. Section (A) of the table includes 30 features which changed at least 1.4-fold between DFTD and healthy devils, with FDR-adjusted q < 0.05, and PLS/DA VIP ≥ 1 (26 features) or PLS/DA VIP ≥ 0.75 and top50 rank in the RF model (4 features). Ten additional features only partly met these strict criteria (Table 2 section B), and were recorded mainly owing to biological interest and in consideration of the small sample size in this analysis. The chemical identity of the features was elucidated using MS spectra and chemical standards, as described in the methods section and also detailed in Table S-1. The panel of differential features comprised of common endogenous compounds, many of which were lower in DFTD, and putative peptides, which were all elevated in DFTD. The individual discriminating performance of the features was further supported by univariate ROC analysis using cross validations, and the AUC and 95% confidence interval were added to Table 2. Most of the features in Table 2(A) had AUC ≥ 0.80, and Figure 1 presents box plots and ROC curves for the top four differentiating features in Table 2, namely urea, cortisol (both lower in DFTD), and two putative peptides (elevated in DFTD), all with AUC ≥ 0.89. Performance of selected features Figure 2 presents a heat map of the 40 features in Table 2 across the full cohort (n = 70), demonstrates the differences between DFTD and healthy controls, and depicts the sample metadata for sex, age and season of sample collection. To assess the performance of the selected features in class prediction, PLS/DA analysis was conducted using the age-



1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

matched samples (n = 20 x 2) as training set, and the remaining 30 unmatched samples (n = 15 x 2) as test set. The scores plot in Figure 3 shows a PLS/DA scores plot for the model (R2Y(cum)= 0.73; Q2(cum) = 0.58 ). DFTD and control samples were mostly separated, with close to perfect classification of the training set (AUC = 0.998, 1 early stage DFTD classified as control, red filled triangle in the scores plot). Class predictions of the 30 samples in the test set resulted in one healthy sample misclassified as DFTD, and one sample with score 2 tumour predicted not to have DFTD, along with two early-DFTD samples. Overall, the sensitivity was 91% (31/34) and specificity 97% (34/35). The discriminating performance of each feature was further examined via univariate ROC analysis for the full dataset (n = 70). Error rates of classification for all samples were added, as the AUC value can be misleading due to its independence of the decision boundaries. When tested in the non age-balanced samples, features which either highly improved their AUC or did not retain their discriminative power would be suspected to depend on age, not only DFTD. This issue is further addressed in supporting information (Supp1; page S15), along with examination of the potentially confounding factor of sex. As discussed in the next section, more than ten highly differentiating features elevated in DFTD were annotated as peptides, and could have originated in fragmentation of the same larger peptide, either in vivo or ex vivo, during sample processing and analysis. Regardless of the individual contribution of each of these features to the former prediction models, we developed a model which included only a single peptide in addition to non-peptide features. To reduce the similarity with previous models which dictated the feature selection, a linear SVM model was developed based on the 232 features. The resulting model combined eight features which also appear in Table 2: the triply-charged peptide with m/z 526.28, urea, Ncarbamoyl sarcosine, creatine, serine, kynurenic acid, hexanoyl carnitine, and the putative 14 ACS Paragon Plus Environment

Page 14 of 41

Page 15 of 41

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Tyr-Ser (m/z 251.10). Figure 4A shows the excellent ROC curve for both the training set (AUC 0.96, 95% CI of 0.86-1) and the test set (AUC 0.96). The average prediction accuracy based on 100 cross validations was 0.88 for the training set and 0.93 for the test set. The probabilities of prediction are presented in Figure 4B, with total of 4 misclassified samples, leading to overall 94% sensitivity and specificity. The model was highly significant with p < 0.001 after 1000 permutations. Association of metabolites with DFTD progression The vast majority of devils with DFTD in this study (90%) showed ulceration of the tumour and 70% had secondary infection, however no significant associations could be established between these parameters and any of the differentiating metabolites, either by correlation or univariate logistic regression. Tasmanian devils can develop multiple facial tumours, and so the total volume of the tumours (confirmed by visual inspection) was calculated for devils with DFTD (n=33). The correlation between the total tumour volume and the 232 features was assessed by Spearman rank test, yielding 20 features with significant association (R≥0.37 or R≤-0.39, all p

Discovery of Biomarkers for Tasmanian Devil ... - ACS Publications

Recommend Documents