Metabolic profiling reveals key metabolic features of renal cell carcinoma

Abstract

Recent evidence suggests that metabolic changes play a pivotal role in the biology of cancer and in particular renal cell carcinoma (RCC). Here, a global metabolite profiling approach was applied to characterize the metabolite pool of RCC and normal renal tissue. Advanced decision tree models were applied to characterize the metabolic signature of RCC and to explore features of metastasized tumours. The findings were validated in a second independent dataset. Vitamin E derivates and metabolites of glucose, fatty acid, and inositol phosphate metabolism determined the metabolic profile of RCC. α-tocopherol, hippuric acid, myoinositol, fructose-1-phosphate and glucose-1-phosphate contributed most to the tumour/normal discrimination and all showed pronounced concentration changes in RCC. The identified metabolic profile was characterized by a low recognition error of only 5% for tumour versus normal samples. Data on metastasized tumours suggested a key role for metabolic pathways involving arachidonic acid, free fatty acids, proline, uracil and the tricarboxylic acid cycle. These results illustrate the potential of mass spectroscopy based metabolomics in conjunction with sophisticated data analysis methods to uncover the metabolic phenotype of cancer. Differentially regulated metabolites, such as vitamin E compounds, hippuric acid and myoinositol, provide leads for the characterization of novel pathways in RCC.

Fig 1

Decision Tree Model (ADTree) generated for the two-class problem of discriminating RCC and normal renal tissue samples (A), and localized RCC and metastatic disease (B). Key metabolites are shown with the corresponding normalized relative peak intensity cut-offs. Each metabolite resembles a decision node that is linked to two prediction nodes with the corresponding prediction values. Classification of a hypothetical sample would be based on the sum of final attained prediction node values that are determined by applying the peak intensity cut-offs for all metabolites of the decision tree on the sample-specific data record. Any result> 0 means a class prediction of 0 (A: normal tissue; B: localized tumour), any result < 0 a class prediction of 1 (A: RCC, B: metastatic tumour). The model was trained with the first dataset and used all metabolites irrespective of identified status.

Fig 2

The information gain for the two class discrimination between RCC and normal tissue by key metabolites. Metabolites with the highest gain contribute most to the correct discrimination. The theoretical maximum gain = 1. The black bars indicate metabolites that were not detectable in all samples and were therefore unable to be incorporated into the ADTree model, but all of these metabolites were detected in over 90% of samples, except for 6-phosphogluconic acid (88%).

Fig 3

Descriptive statistics of relative metabolite concentrations in tumour versus normal tissue. Select key metabolites are chosen based on their high informational gain for the tumour/normal discrimination and/or their identification in the decision tree analysis. Boxplots show median, 25^th and 75^th percentiles, range, and extreme values. For better illustration a logarithmic scale was chosen for the relative concentration; absolute concentrations cannot be calculated and therefore no precise scale is given.