Limits...
Metabolomic Screening of Tumor Tissue and Serum in Glioma Patients Reveals Diagnostic and Prognostic Information.

Mörén L, Bergenheim AT, Ghasimi S, Brännström T, Johansson M, Antti H - Metabolites (2015)

Bottom Line: However, in the clinical reality, large variations exist in the survival of patients concerning both glioblastomas and low-grade gliomas.Interestingly, we also found patterns in both tumor and serum with individual metabolite features that were both elevated and decreased in patients that lived long after being diagnosed with glioblastoma compared to those who died shortly after diagnosis (p(tum)(o)(r) = 0.006, p(serum) = 0.004; AUROCC(tumor) = 0.846 (0.647-1.000), AUROCC(serum) = 0.958 (0.870-1.000)).Metabolic patterns could also distinguish long and short survival in patients diagnosed with oligodendroglioma (p(tumor) = 0.01, p(serum) = 0.001; AUROCC(tumor) = 1 (1.000-1.000), AUROCC(serum) = 1 (1.000-1.000)).

View Article: PubMed Central - PubMed

Affiliation: Department of Chemistry, Umeå University, SE 901 87 Umeå, Sweden. lina.moren@umu.se.

ABSTRACT
Glioma grading and classification, today based on histological features, is not always easy to interpret and diagnosis partly relies on the personal experience of the neuropathologists. The most important feature of the classification is the aimed correlation between tumor grade and prognosis. However, in the clinical reality, large variations exist in the survival of patients concerning both glioblastomas and low-grade gliomas. Thus, there is a need for biomarkers for a more reliable classification of glioma tumors as well as for prognosis. We analyzed relative metabolite concentrations in serum samples from 96 fasting glioma patients and 81 corresponding tumor samples with different diagnosis (glioblastoma, oligodendroglioma) and grade (World Health Organization (WHO) grade II, III and IV) using gas chromatography-time of flight mass spectrometry (GC-TOFMS). The acquired data was analyzed and evaluated by pattern recognition based on chemometric bioinformatics tools. We detected feature patterns in the metabolomics data in both tumor and serum that distinguished glioblastomas from oligodendrogliomas (p(tumor) = 2.46 × 10(-8), p(serum) = 1.3 × 10(-5)) and oligodendroglioma grade II from oligodendroglioma grade III (p(tumor) = 0.01, p(serum) = 0.0008). Interestingly, we also found patterns in both tumor and serum with individual metabolite features that were both elevated and decreased in patients that lived long after being diagnosed with glioblastoma compared to those who died shortly after diagnosis (p(tum)(o)(r) = 0.006, p(serum) = 0.004; AUROCC(tumor) = 0.846 (0.647-1.000), AUROCC(serum) = 0.958 (0.870-1.000)). Metabolic patterns could also distinguish long and short survival in patients diagnosed with oligodendroglioma (p(tumor) = 0.01, p(serum) = 0.001; AUROCC(tumor) = 1 (1.000-1.000), AUROCC(serum) = 1 (1.000-1.000)). In summary, we found different metabolic feature patterns in tumor tissue and serum for glioma diagnosis, grade and survival, which indicates that, following further verification, metabolomic profiling of glioma tissue as well as serum may be a valuable tool in the search for latent biomarkers for future characterization of malignant glioma.

No MeSH data available.


Related in: MedlinePlus

Overview of the metabolomic workflow. (Upper panel. Left) Raw gas chromatography-time of flight mass spectrometry (GC-TOFMS) data for the analyzed samples makes up a three dimensional matrix with the Time axis being retention time or index for each metabolite linked to the elution from the chromatographic system, the mass to charge (m/z) axis being the mass over charge ration for the molecular fragments detected by the mass spectrometer and the Samples axis being the analyzed samples. (Middle) To obtain pure chromatographic and spectral profiles for relative quantification and identification of metabolites the raw GC-TOFMS data was processed by hierarchical multivariate curve resolution (HMCR), which is a multivariate deconvolution technique especially developed to resolve complex GC-MS based metabolomics data from multiple samples to make it suitable for multiple sample comparisons by means of e.g. pattern recognition approaches. (Right) The area under each resolved metabolite peak makes up the variables of the resulting data matrix (X) used as input for further pattern recognition and statistical analysis. Each column of X represents one resolved metabolite peak over all samples (rows of X). Chemometric bioinformatics based pattern recognition is applied to, X.; e.g. for investigating the difference between two sample classes (turquoise and grey in X). (Lower panel. Left) The sample variation of X is projected in the model scores allowing interpretation of sample distribution patterns. Each symbol in the scores plot represents one sample described by all variables/metabolites (columns of X). As an example, the pink sample symbol relates to the pink row of X. (Right) The variable/metabolite variation is projected in the model loadings allowing interpretation of sample distribution patterns and explanation of variable contribution to patterns in seen in scores. Each symbol in the loading plot represents on variable/metabolite. As an example, the blue symbol relates to the blue column of X as well as the blue resolved metabolite profile in the upper middle frame.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4588809&req=5

metabolites-05-00502-f003: Overview of the metabolomic workflow. (Upper panel. Left) Raw gas chromatography-time of flight mass spectrometry (GC-TOFMS) data for the analyzed samples makes up a three dimensional matrix with the Time axis being retention time or index for each metabolite linked to the elution from the chromatographic system, the mass to charge (m/z) axis being the mass over charge ration for the molecular fragments detected by the mass spectrometer and the Samples axis being the analyzed samples. (Middle) To obtain pure chromatographic and spectral profiles for relative quantification and identification of metabolites the raw GC-TOFMS data was processed by hierarchical multivariate curve resolution (HMCR), which is a multivariate deconvolution technique especially developed to resolve complex GC-MS based metabolomics data from multiple samples to make it suitable for multiple sample comparisons by means of e.g. pattern recognition approaches. (Right) The area under each resolved metabolite peak makes up the variables of the resulting data matrix (X) used as input for further pattern recognition and statistical analysis. Each column of X represents one resolved metabolite peak over all samples (rows of X). Chemometric bioinformatics based pattern recognition is applied to, X.; e.g. for investigating the difference between two sample classes (turquoise and grey in X). (Lower panel. Left) The sample variation of X is projected in the model scores allowing interpretation of sample distribution patterns. Each symbol in the scores plot represents one sample described by all variables/metabolites (columns of X). As an example, the pink sample symbol relates to the pink row of X. (Right) The variable/metabolite variation is projected in the model loadings allowing interpretation of sample distribution patterns and explanation of variable contribution to patterns in seen in scores. Each symbol in the loading plot represents on variable/metabolite. As an example, the blue symbol relates to the blue column of X as well as the blue resolved metabolite profile in the upper middle frame.

Mentions: Pattern recognition utilizes multivariate projection methods to extract and verify co-varying patterns or signatures of variables that are significant for explaining systematic variation in experimental data. In metabolomics, pattern recognition works to compress the variable space, i.e., the detected and relatively quantified metabolic features, into a few latent variables, e.g., principal components, explaining the majority of the systematic variation in the data. In this way interpretation of changes in metabolic signatures as well as detection of robust and relevant sample patterns caused by those signature changes are largely facilitated. In this work, processed metabolomics data from tissue and serum samples were analyzed separately using different pattern recognition approaches. In a first step, principal component analysis (PCA) [44] was applied to get an unsupervised overview of the variation in the data and to detect deviating samples, so-called outliers. For further multivariate sample comparisons with the aim to look for differences between pre-defined sample classes, orthogonal partial least squares-discriminant analysis (OPLS-DA) [45] was used. OPLS is a supervised multivariate regression method allowing a separation of the variation into predictive variation (related to the response(s) of interest) and orthogonal variation (variation unrelated to the response(s) of interest). This has been shown to facilitate the interpretation of complex multivariate data and the interactions therein. The combined data processing and pattern recognition procedure can be overviewed in Figure 3. Initially, the diagnostic potential of the metabolic profiles in terms of discriminating between GBMs and oligodendrogliomas was evaluated. The low grade astrocytomas (grade II and III) were excluded from further evaluation due to the low number of samples making statistical analysis unfeasible. Then differences in tumor grade were investigated comparing samples from oligodendroglioma patients with grade II and grade III tumors, respectively. Finally, metabolic patterns associated with survival time were investigated in glioblastomas and oligodendrogliomas separately. From the calculated OPLS-DA models, model weight values (w*), i.e., variable contribution values for the pre-defined sample class separations, were extracted and only variables related to the class separation were included in the final OPLS-DA models (w*average ± 2 SD). Furthermore, a Mann–Whitney U-test was used to calculate a probability value (p-value) for each included metabolite in relation to the class separations of interest. All models were validated using cross-validation and p-values for the cross-validated model were calculated using CV-ANOVA [46]. A seven-fold cross-validation procedure was applied using 1/7 of the data as the test set, while the remaining 6/7 of the data were modeled and then repeating this seven times. Furthermore, for all OPLS-DA models the number of latent variables (OPLS components) (A), the variation described in the metabolite data (R2X), the between class variation described (R2Y) and the between class variation predicted based on cross-validation (Q2) were reported. When comparing time of survival in GBM, patients that died shortly after diagnosis (≤4 months) were compared to patients that lived long after diagnosis (≥3 years). For oligodendrogliomas, patients that died within 2 years after diagnosis (short survival) were compared to patients that lived for more than 3 years (long survival). Survival groups were selected based on retrospective data from our institution with an expected 4 months survival in GBM of 65% and 3 year survival of only 8%. In oligodendroglioma grade II, 2- and 3-year survival is expected to be 78% and 75%, respectively. For patients with oligodendroglioma grade III expected survival of 2 and 3 years is 50% and 50%, respectively. However, in our consecutively collected material, we did not have enough patients within any clinically relevant time limits. Instead, we had two almost equally sized groups of patients; one group that died within two years of diagnosis and the other group that lived longer than 3 years, which is why those limits were the most inherent to get reliable statistics. All pattern recognition analysis, including cross-validation and CV-ANOVA, was performed in SIMCA (version SIMCA-P + 13.0; Umetrics, Umeå, Sweden). Model plots were created using SIMCA or GraphPad Prism (5.04; GraphPad Software Inc., La Jolla, CA, USA) in combination with Adobe Illustrator CS5 (15.0.0; Adobe Systems Inc., San Jose, CA, USA). To summarize the results, Receiver Operating Characteristic (ROC) curves were calculated for the detected significant metabolic patterns associated with survival. This way of utilizing ROC curves for metabolic patterns as compared to the conventional way using single markers is novel and makes it possible to evaluate the diagnostic and prognostic potential of metabolic patterns or signatures in a way that is familiar to the clinical community. The ROC calculations were performed in ROCCET: ROC Curve Explorer & Tester (www.roccet.ca) [47].


Metabolomic Screening of Tumor Tissue and Serum in Glioma Patients Reveals Diagnostic and Prognostic Information.

Mörén L, Bergenheim AT, Ghasimi S, Brännström T, Johansson M, Antti H - Metabolites (2015)

Overview of the metabolomic workflow. (Upper panel. Left) Raw gas chromatography-time of flight mass spectrometry (GC-TOFMS) data for the analyzed samples makes up a three dimensional matrix with the Time axis being retention time or index for each metabolite linked to the elution from the chromatographic system, the mass to charge (m/z) axis being the mass over charge ration for the molecular fragments detected by the mass spectrometer and the Samples axis being the analyzed samples. (Middle) To obtain pure chromatographic and spectral profiles for relative quantification and identification of metabolites the raw GC-TOFMS data was processed by hierarchical multivariate curve resolution (HMCR), which is a multivariate deconvolution technique especially developed to resolve complex GC-MS based metabolomics data from multiple samples to make it suitable for multiple sample comparisons by means of e.g. pattern recognition approaches. (Right) The area under each resolved metabolite peak makes up the variables of the resulting data matrix (X) used as input for further pattern recognition and statistical analysis. Each column of X represents one resolved metabolite peak over all samples (rows of X). Chemometric bioinformatics based pattern recognition is applied to, X.; e.g. for investigating the difference between two sample classes (turquoise and grey in X). (Lower panel. Left) The sample variation of X is projected in the model scores allowing interpretation of sample distribution patterns. Each symbol in the scores plot represents one sample described by all variables/metabolites (columns of X). As an example, the pink sample symbol relates to the pink row of X. (Right) The variable/metabolite variation is projected in the model loadings allowing interpretation of sample distribution patterns and explanation of variable contribution to patterns in seen in scores. Each symbol in the loading plot represents on variable/metabolite. As an example, the blue symbol relates to the blue column of X as well as the blue resolved metabolite profile in the upper middle frame.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4588809&req=5

metabolites-05-00502-f003: Overview of the metabolomic workflow. (Upper panel. Left) Raw gas chromatography-time of flight mass spectrometry (GC-TOFMS) data for the analyzed samples makes up a three dimensional matrix with the Time axis being retention time or index for each metabolite linked to the elution from the chromatographic system, the mass to charge (m/z) axis being the mass over charge ration for the molecular fragments detected by the mass spectrometer and the Samples axis being the analyzed samples. (Middle) To obtain pure chromatographic and spectral profiles for relative quantification and identification of metabolites the raw GC-TOFMS data was processed by hierarchical multivariate curve resolution (HMCR), which is a multivariate deconvolution technique especially developed to resolve complex GC-MS based metabolomics data from multiple samples to make it suitable for multiple sample comparisons by means of e.g. pattern recognition approaches. (Right) The area under each resolved metabolite peak makes up the variables of the resulting data matrix (X) used as input for further pattern recognition and statistical analysis. Each column of X represents one resolved metabolite peak over all samples (rows of X). Chemometric bioinformatics based pattern recognition is applied to, X.; e.g. for investigating the difference between two sample classes (turquoise and grey in X). (Lower panel. Left) The sample variation of X is projected in the model scores allowing interpretation of sample distribution patterns. Each symbol in the scores plot represents one sample described by all variables/metabolites (columns of X). As an example, the pink sample symbol relates to the pink row of X. (Right) The variable/metabolite variation is projected in the model loadings allowing interpretation of sample distribution patterns and explanation of variable contribution to patterns in seen in scores. Each symbol in the loading plot represents on variable/metabolite. As an example, the blue symbol relates to the blue column of X as well as the blue resolved metabolite profile in the upper middle frame.
Mentions: Pattern recognition utilizes multivariate projection methods to extract and verify co-varying patterns or signatures of variables that are significant for explaining systematic variation in experimental data. In metabolomics, pattern recognition works to compress the variable space, i.e., the detected and relatively quantified metabolic features, into a few latent variables, e.g., principal components, explaining the majority of the systematic variation in the data. In this way interpretation of changes in metabolic signatures as well as detection of robust and relevant sample patterns caused by those signature changes are largely facilitated. In this work, processed metabolomics data from tissue and serum samples were analyzed separately using different pattern recognition approaches. In a first step, principal component analysis (PCA) [44] was applied to get an unsupervised overview of the variation in the data and to detect deviating samples, so-called outliers. For further multivariate sample comparisons with the aim to look for differences between pre-defined sample classes, orthogonal partial least squares-discriminant analysis (OPLS-DA) [45] was used. OPLS is a supervised multivariate regression method allowing a separation of the variation into predictive variation (related to the response(s) of interest) and orthogonal variation (variation unrelated to the response(s) of interest). This has been shown to facilitate the interpretation of complex multivariate data and the interactions therein. The combined data processing and pattern recognition procedure can be overviewed in Figure 3. Initially, the diagnostic potential of the metabolic profiles in terms of discriminating between GBMs and oligodendrogliomas was evaluated. The low grade astrocytomas (grade II and III) were excluded from further evaluation due to the low number of samples making statistical analysis unfeasible. Then differences in tumor grade were investigated comparing samples from oligodendroglioma patients with grade II and grade III tumors, respectively. Finally, metabolic patterns associated with survival time were investigated in glioblastomas and oligodendrogliomas separately. From the calculated OPLS-DA models, model weight values (w*), i.e., variable contribution values for the pre-defined sample class separations, were extracted and only variables related to the class separation were included in the final OPLS-DA models (w*average ± 2 SD). Furthermore, a Mann–Whitney U-test was used to calculate a probability value (p-value) for each included metabolite in relation to the class separations of interest. All models were validated using cross-validation and p-values for the cross-validated model were calculated using CV-ANOVA [46]. A seven-fold cross-validation procedure was applied using 1/7 of the data as the test set, while the remaining 6/7 of the data were modeled and then repeating this seven times. Furthermore, for all OPLS-DA models the number of latent variables (OPLS components) (A), the variation described in the metabolite data (R2X), the between class variation described (R2Y) and the between class variation predicted based on cross-validation (Q2) were reported. When comparing time of survival in GBM, patients that died shortly after diagnosis (≤4 months) were compared to patients that lived long after diagnosis (≥3 years). For oligodendrogliomas, patients that died within 2 years after diagnosis (short survival) were compared to patients that lived for more than 3 years (long survival). Survival groups were selected based on retrospective data from our institution with an expected 4 months survival in GBM of 65% and 3 year survival of only 8%. In oligodendroglioma grade II, 2- and 3-year survival is expected to be 78% and 75%, respectively. For patients with oligodendroglioma grade III expected survival of 2 and 3 years is 50% and 50%, respectively. However, in our consecutively collected material, we did not have enough patients within any clinically relevant time limits. Instead, we had two almost equally sized groups of patients; one group that died within two years of diagnosis and the other group that lived longer than 3 years, which is why those limits were the most inherent to get reliable statistics. All pattern recognition analysis, including cross-validation and CV-ANOVA, was performed in SIMCA (version SIMCA-P + 13.0; Umetrics, Umeå, Sweden). Model plots were created using SIMCA or GraphPad Prism (5.04; GraphPad Software Inc., La Jolla, CA, USA) in combination with Adobe Illustrator CS5 (15.0.0; Adobe Systems Inc., San Jose, CA, USA). To summarize the results, Receiver Operating Characteristic (ROC) curves were calculated for the detected significant metabolic patterns associated with survival. This way of utilizing ROC curves for metabolic patterns as compared to the conventional way using single markers is novel and makes it possible to evaluate the diagnostic and prognostic potential of metabolic patterns or signatures in a way that is familiar to the clinical community. The ROC calculations were performed in ROCCET: ROC Curve Explorer & Tester (www.roccet.ca) [47].

Bottom Line: However, in the clinical reality, large variations exist in the survival of patients concerning both glioblastomas and low-grade gliomas.Interestingly, we also found patterns in both tumor and serum with individual metabolite features that were both elevated and decreased in patients that lived long after being diagnosed with glioblastoma compared to those who died shortly after diagnosis (p(tum)(o)(r) = 0.006, p(serum) = 0.004; AUROCC(tumor) = 0.846 (0.647-1.000), AUROCC(serum) = 0.958 (0.870-1.000)).Metabolic patterns could also distinguish long and short survival in patients diagnosed with oligodendroglioma (p(tumor) = 0.01, p(serum) = 0.001; AUROCC(tumor) = 1 (1.000-1.000), AUROCC(serum) = 1 (1.000-1.000)).

View Article: PubMed Central - PubMed

Affiliation: Department of Chemistry, Umeå University, SE 901 87 Umeå, Sweden. lina.moren@umu.se.

ABSTRACT
Glioma grading and classification, today based on histological features, is not always easy to interpret and diagnosis partly relies on the personal experience of the neuropathologists. The most important feature of the classification is the aimed correlation between tumor grade and prognosis. However, in the clinical reality, large variations exist in the survival of patients concerning both glioblastomas and low-grade gliomas. Thus, there is a need for biomarkers for a more reliable classification of glioma tumors as well as for prognosis. We analyzed relative metabolite concentrations in serum samples from 96 fasting glioma patients and 81 corresponding tumor samples with different diagnosis (glioblastoma, oligodendroglioma) and grade (World Health Organization (WHO) grade II, III and IV) using gas chromatography-time of flight mass spectrometry (GC-TOFMS). The acquired data was analyzed and evaluated by pattern recognition based on chemometric bioinformatics tools. We detected feature patterns in the metabolomics data in both tumor and serum that distinguished glioblastomas from oligodendrogliomas (p(tumor) = 2.46 × 10(-8), p(serum) = 1.3 × 10(-5)) and oligodendroglioma grade II from oligodendroglioma grade III (p(tumor) = 0.01, p(serum) = 0.0008). Interestingly, we also found patterns in both tumor and serum with individual metabolite features that were both elevated and decreased in patients that lived long after being diagnosed with glioblastoma compared to those who died shortly after diagnosis (p(tum)(o)(r) = 0.006, p(serum) = 0.004; AUROCC(tumor) = 0.846 (0.647-1.000), AUROCC(serum) = 0.958 (0.870-1.000)). Metabolic patterns could also distinguish long and short survival in patients diagnosed with oligodendroglioma (p(tumor) = 0.01, p(serum) = 0.001; AUROCC(tumor) = 1 (1.000-1.000), AUROCC(serum) = 1 (1.000-1.000)). In summary, we found different metabolic feature patterns in tumor tissue and serum for glioma diagnosis, grade and survival, which indicates that, following further verification, metabolomic profiling of glioma tissue as well as serum may be a valuable tool in the search for latent biomarkers for future characterization of malignant glioma.

No MeSH data available.


Related in: MedlinePlus