The basic idea behind CoMet is schematically described in Figure 1. The intracellular level of a given metabolite is predicted to be decreased or increased in cancer cells based on the analysis of the relative expression levels of the human genes encoding for all identified enzymes that employ the metabolite as substrate or product (see Material and Methods for details). We applied our approach to the Jurkat cell line, which is derived from an acute T lymphoblastic leukemia patient 21. The comparison of two Jurkat cell samples to three GM15851 lymphoblast cell samples (see Additional File 1) resulted in 104 metabolites predicted to be lowered in Jurkat cancer cells and 78 metabolites predicted to be increased in these cells, out of a total of 982 metabolites considered in the analysis (see Additional File 2). According to our hypothesis, we would expect an enrichment of Jurkat antileukemic agents among the 104 metabolites predicted to be lowered in the cancer cells. By performing an exhaustive search of the literature for experimental evidence, we found that 13 of the 982 analyzed metabolites has previously been demonstrated to exhibit anticancer activity in Jurkat cells. Table 1 shows that 2/13 metabolites are predicted to be lowered in Jurkat cells: thymidine, an antineoplastic agent 22, and prostaglandin D2, which induces apoptosis without inhibiting the viability of normal T lymphocytes 23. Only 1/13 proven anticancer agents in Jurkat cells belongs to the group of 78 metabolites predicted to be increased in these cancer cells: the apoptotic agent 17β-2-methoxyestradiol 24. The remaining 10 known anticancer molecules active in Jurkat cells: testosterone 25, melatonin 26, sphingolipid GD3 27, 2'-deoxyguanosine 28, 2'-deoxyadenosine 29, 2'-deoxyinosine 29, nicotinamide 30, methylglyoxal 31, linoleic acid 32 and cAMP 33 are included in the set of 800 metabolites whose intracellular levels are predicted to be essentially the same in both Jurkat and normal cells. Although the fraction of metabolites with known anticancer activity among the compounds predicted to be lowered in Jurkat cells (2/104 = 0.019) is higher than that corresponding to the rest of the compounds [11 non predicted ones have literature validated anticancer properties; (1+10)/(78+800) = 0.013], the difference is not statistically significant (two-tailed Fisher's exact test at a critical alpha level of 0.05). On the other hand, it has to be noticed that negative results tend to be underreported, making it difficult to obtain unbiased statistics about metabolites that lack anticancer properties.

Based on simple criteria such as low molecular weight, commercial availability and affordability, we selected nine metabolites predicted to be lowered in Jurkat cells in order to test their effect on its proliferation (Table 1). For one of these nine metabolites (seleno-L-methionine) we carried out a validation by quantitative real-time PCR of the microarray data used by CoMet to make its prediction (see Additional File 3). We examined the effect on the growth of Jurkat cells of a 72 h treatment with riboflavin, tryptamine, 3-sulfino-L-alanine, menaquinone, dehydroepiandrosterone (the non-sulfated version of the predicted metabolite dehydroepiandrosterone sulfate), α-hydroxystearic acid (one of the possible compounds compatible with the predicted generic metabolite α-hydroxy fatty acid), hydroxyacetone, seleno-L-methionine and 5,6-dimethylbenzimidazole (the aglycone of the predicted metabolite α-ribazole) at a concentration of 100 μM (see Additional File 1). Figure 2A shows that all the tested metabolites with the exception of sulfino-L-alanine exhibited statistically significant antiproliferative activity on Jurkat cells (as evaluated by two-tailed t-tests at a critical alpha level of 0.05), with growth below 90% of the untreated control in all the cases. Although sulfino-L-alanine alone did not inhibit the growth of Jurkat cells, it significantly potentiated the inhibitory effect of seleno-L-methionine (43.1% to 30.3% growth compared to control). Similarly, a synergistic interaction between 5,6-dimethylbenzimidazole and seleno-L-methionine led to a supra-additive growth inhibitory activity. On the other hand, α-hydroxystearic acid and dehydroepiandrosterone show an additive effect, while α-hydroxystearic acid and seleno-L-methionine exhibited sub-additive or antagonistic inhibitory activity. Menaquinone showed the highest antiproliferative activity (11.3% growth compared to control), whereas the inhibitory activity of riboflavin, tryptamine and hydroxyacetone on Jurkat cells was more moderate (all above 70% growth compared to control). Thus, even with the small set tested, a remarkable richness of antiproliferative responses is seen.

Although the fact that the nine tested metabolites predicted to be lowered in Jurkat cells exhibited antiproliferative activity strongly support our hypothesis, the possibility still exists that many endogenous metabolites at the concentration used in our test inhibit the growth of Jurkat cells, independent of the intracellular level status predicted by CoMet. Therefore, we tested metabolites whose intracellular levels in Jurkat cells were predicted to be increased (bilirubin, androsterone, homovanillic acid, vanillylmandelic acid, N-acetyl-L-aspartate and taurocholic acid) or unchanged (pantothenic acid, citric acid, folic acid, β-D-galactose, cholesterol) compared with lymphoblasts. We analyzed the effect on the growth of Jurkat cells of a 72 h treatment with each of the eleven human metabolites at a concentration of 100 μM (see Additional File 1). Figure 2B shows that only two of the six tested metabolites whose concentrations are predicted to be increased in Jurkat cells exhibit significant antiproliferative activity: bilirubin and androsterone (19.6% and 54.5% growth compared to control, respectively). The growth inhibition exerted by each of the remaining tested metabolites was less than 10% and statistically insignificant. Similarly, Figure 2C shows that all the tested metabolites whose intracellular levels in Jurkat cells and lymphoblasts we predict to be comparable, exhibit no significant effect on the growth of cells. Statistical significance was evaluated in all the cases according to two-tailed t-tests at a critical alpha level of 0.05.

To validate the relevance of the identified antiproliferative metabolites as potential anticancer agents, we investigated the selectivity of their antiproliferative activity towards Jurkat cells with respect to human lymphoblast cells. We compared the effect of a 72 h treatment with menaquinone, dehydroepiandrosterone, seleno-L-methionine and 5,6-dimethylbenzimidazole at concentrations of 100 and 50 μM on the growth of Jurkat cells and human GM15851 lymphoblast cells cultured in the same conditions (see Additional File 1). Table 2 shows that menaquinone and dehydroepiandrosterone at concentrations of 100 or 50 μM exhibit significant antiproliferative activity on Jurkat cells but not on lymphoblast cells. Moreover, dehydroepiandrosterone at a concentration of 50 μM shows a statistically significant stimulation of the growth of lymphoblasts (130% growth compared to control), while at the same concentration in Jurkat cells the growth rate is reduced to 83.2%, thereby providing an interesting example of selective inhibition. Seleno-L-methionine inhibits the growth of Jurkat and lymphoblast cells at both tested concentrations; however, the effect on Jurkat cells was considerably more potent. While 5,6-dimethylbenzimidazole at a concentration of 100 μM exhibits significant antiproliferative activity on both tested cell lines, at a concentration of 50 μM it only inhibits Jurkat cells. Statistical significance was evaluated in all the cases according to two-tailed t-tests at a critical alpha level of 0.05.

At a concentration of 100 μmol/L, all nine metabolites predicted to be lowered in Jurkat cells selected for cell proliferation assays exhibited a statistically significant inhibition of Jurkat cell growth below 90% of the untreated control, whether alone, in most of the cases, or in combination, in the case of 3-sulfino-L-alanine (Figure 2A). On the other hand, 2/6 tested metabolites predicted to be increased in Jurkat cells unexpectedly exhibited antiproliferative activity (Figure 2B), while none of the five tested metabolites that we predicted to be unchanged inhibited Jurkat cell growth (Figure 2C). Thus, 18/20 assayed metabolites behave according to our working hypothesis. If we jointly consider the novel antiproliferative compounds presented in this work and those metabolites whose anticancer activity in Jurkat cells was previously known (discussed above), the fraction of anticancer metabolites among the compounds predicted to be lowered in Jurkat cells [(9+2)/104 = 0.106] is considerably higher than that corresponding to the rest of the compounds [(1+2+10+0)/(78+800) = 0.015]. A Fisher's exact test indicates that the positive association between lowered metabolite levels in Jurkat cells as predicted by CoMet and antiproliferative activity of the metabolite in that cell line is highly significant (two-tailed p-value = 8.7 × 10^-6). These findings clearly support our hypothesis regarding the active role of endogenous metabolites in cancer, especially the concept that the metabolism of cancer cells may be regulated so that the intracellular concentration of certain antiproliferative metabolites is reduced, resulting in uninhibited cellular growth.

The growth inhibitory effects of some of the predicted compounds may seem relatively low and the tested concentrations of 50 and 100 μmol/L too high compared with most anticancer drugs of synthetic or natural origin. However, this range of concentrations is not unreasonable for metabolic compounds, since many metabolites can be found at similar levels in the cytosol and/or extracellular fluids 34. Furthermore, when the effect of selected metabolites on growth inhibition was tested in Jurkat and lymphoblast cells cultured in identical conditions, a pattern of selectivity of the antiproliferative effect towards the cancer cell line became evident (Table 2). In an extreme case, dehydroepiandrosterone at a concentration of 50 μM inhibited the growth of Jurkat cells but stimulated the proliferation of lymphoblasts. Interestingly, several of our newly found antiproliferative metabolites exhibited synergistic interactions consistent with the Systems Biology approach of our method: the prediction was performed on the entire metabolome and not on individual metabolites or pathways. This raises the intriguing question of what the result would be if concentrations close to those observed in the normal cells could be achieved in the cancer cell for most metabolites, i.e. a reversion to a normal-like metabolic profile, at least for metabolites that inhibit cancer cell growth. If the antiproliferative activities we observed in a cancer cell line have therapeutic value, different combined strategies can be devised where sets of predicted metabolites are concurrently selected according to their association with the same or different metabolic pathways. For example, we can employ a strategy where multiple metabolites target a single pathway, or on the contrary, where each metabolite acts specifically on a different pathway. In addition, some active metabolites might serve as completely novel lead compounds for further drug design and development, with the advantage of reduced initial toxicity. While we have only performed cell proliferation assays, it is reasonable to speculate that some metabolites may also exhibit anti-metastatic (anti-invasive), anti-angiogenic, immunostimulant or other anticancer properties that would not be evident in an in vitro study of cell growth inhibition.

We did not investigate here the mode of action of the antiproliferative metabolites found by CoMet, and it is possible that some may exert their effect based on completely novel mechanisms, whose elucidation goes well beyond the immediate scope of this study. However, for most, we can suggest a possible mode of action based on their effect on other cancer cells or on the known properties of closely related molecules. For example, 5,6-dichlorobenzimidazole, a bioisosteric derivative of the active metabolite 5,6-dimethylbenzimidazole, induces differentiation of malignant erythroblasts by inhibiting RNA polymerase II 35. The tested metabolite tryptamine is an effective inhibitor of HeLa cell growth via the competitive inhibition of tryptophanyl-tRNA synthetase, and consequent inhibition of protein biosynthesis 36. 9-hydroxystearic acid, an isomer of the active metabolite α-hydroxystearic acid, arrests HT29 colon cancer cells in G0/G1 phase of the cell cycle via overexpression of p21 37, induces differentiation of HT29 cells 38 by inhibition of histone deacetylase 1 39 and interrupts the transduction of the mitogenic signal 40. Menaquinone (vitamin K2), the most efficient compound among the tested metabolites, has been previously reported to induce G0/G1 arrest, differentiation and apoptosis in acute myelomonocytic leukemia HL-60 cells 41.

Besides the general problems that affect any microarray study 4243, there are some factors not accounted in our method that can influence the actual intracellular levels of a metabolite. First, our qualitative treatment of metabolic flux is highly simplified; however, more quantitative approaches such as flux balance analysis require the knowledge of the regulatory effects of covalent modifications and the kinetic rate constants of the associated enzymes, information that is both incomplete and not accurate enough to generate large-scale models 44. Second, the information available about both the subcellular location where the metabolic conversions take place and the transport of metabolites between different intra- or extracellular compartments is very limited. To partially address this issue, we are currently incorporating information about transporter genes into CoMet for qualitative metabolic flux predictions. Finally, a factor that could confound the hypothetical correlation between lowered metabolites in cancer and their potential as therapeutic agents is the existence of moonlighting activities, such as transcriptional regulation and apoptosis, exhibited by several metabolic enzymes 45. Since the growth control functions of moonlighting enzymes do not involve catalysis, their regulated levels in cancer cells may not be correlated with the intracellular concentration of their metabolic substrates or products. Nevertheless, in spite of these in principle limitations, in practice, CoMet greatly succeeded in identifying antiproliferative metabolites. On the other hand, its ability to predict changes in intracellular metabolite concentrations still needs to be proven by experimentally determining the relative intracellular levels of a representative set of metabolites in normal and cancer cells, a task that we are planning to carry out in the near future.

The main source of biological information is the Kyoto Encyclopedia of Genes and Genomes (KEGG) from July 5, 2007 46. We obtained the enzyme function annotation for human genes from the KEGG GENES database, the chemical information about human metabolites from the KEGG LIGAND database, and the metabolic pathway data from the KEGG PATHWAY database. Enzyme function annotations from KEGG were complemented by high confidence predictions made by EFICAz 47, our highly precise approach for enzyme function inference that significantly increases annotation coverage 48. For the mapping between microarray probe identifiers and Entrez GeneID identifiers, the Affymetrix HG-U133 Plus 2.0 NetAffx Annotation file of May 31, 2007 was used 49.

The first step of the CoMet approach consists of the classification of each enzyme-coding human gene into four possible groups: G1) upregulated in cancer cells, G2) downregulated in cancer cells, G3) expressed in both, normal and cancer cells, at levels that are statistically indistinguishable, and G4) not expressed in both, normal and cancer cells. We used two types of data for the classification: the log base 2 signal intensities and the presence calls of the corresponding probe sets, as reported by the Affymetrix Microarray Suite Software 5.0 (MAS 5.0). First, an "off" status is assigned to each gene in each of the two studied conditions (normal and cancer) if the mean fraction of presence calls labeled as "marginal" or "absent" in the corresponding probe sets is at least 80%; otherwise, an "on" status is assigned. Then, each gene is classified into the G1, G2, G3 or G4 group, according to its on/off status in normal and cancer conditions and the following criterion for differential expression: the signal intensities in normal and cancer samples exhibit a statistically significant difference in at least 40% of the corresponding probe sets, as evaluated by an ANOVA two-tailed test with P < 0.005.

The second step is the in silico estimation of the effect that the differentially expressed enzyme-encoding genes may exert on the intracellular levels of metabolites. First, all human metabolic pathways from the KEGG PATHWAY database, a compilation of maps representing the molecular interactions and reaction networks for different types of biological processes, were retrieved. For biological process labeled as Metabolism, there are eleven groups of pathways: 1) Carbohydrate Metabolism, 2) Energy Metabolism, 3) Lipid Metabolism, 4) Nucleotide Metabolism, 5) Amino Acid Metabolism, 6) Metabolism of Other Amino Acids, 7) Glycan Biosynthesis and Metabolism, 8) Biosynthesis of Polyketides and Nonribosomal Peptides, 9) Metabolism of Cofactors and Vitamins, 10) Biosynthesis of Secondary Metabolites, and 11) Xenobiotics Biodegradation and Metabolism. The pathway maps are available as graphical images and also as KEGG Markup Language (KGML) files that facilitate the parsing of relevant biological data. We extracted all the biochemical reactions from the KGML human metabolic pathway maps, including information about substrates, products, direction/reversibility and associated enzyme-coding genes. We then combined this information with gene expression data from normal and cancer cells to construct a genetic-metabolic matrix that links each of 1,477 metabolites with the specific human genes encoding for enzymes that consume and/or produce each metabolite, storing for each gene the differential expression status given by the four-group classification described in the previous section.

From the genetic-metabolic matrix, we excluded: i) 209 non-physiological metabolites, here defined as those that only participate in reactions that belong to the "Biosynthesis of Secondary Metabolites" and the "Xenobiotics Biodegradation and Metabolism" groups of metabolic pathways, e.g. ecgonine or parathion, ii) 90 metabolites that are considered ubiquitous and often carry out generic roles in many reactions 50, here defined as those that are involved as substrate or product in 20 or more reactions, e.g. H₂O, ATP, NAD(⁺)(P) or O₂, and iii) 289 metabolites that participate in reactions that are mainly catalyzed by orphan human enzymes. We define the number of reactions where a metabolite m acts as substrate or product in human metabolic pathways as Nr_{m, human}, and in reference (non organism specific) metabolic pathways as Nr_{m, ref}. If Nr_{m, human}/Nr_{m, ref} < 0.5, then the metabolite m belongs to the third exclusion category. The reactions absent in human pathways may be due to orphan enzymes, reactions that only occur in other organisms or reactions that may occur in humans but have not yet been detected. For example, the metabolite 1-alkyl-sn-glycero-3-phosphate is excluded because out of four enzymes that use it as substrate or product, two, EC 2.3.1.105 and EC 1.1.1.101, are orphans in human, and one, EC 2.7.1.93, has only been found in rabbit 51. The total number of metabolites remaining in the genetic-metabolic matrix after the three types of exclusion is 982 (see Additional File 2).

In the last step, we apply a simple set of rules to scan the genetic-metabolic matrix for metabolites whose intracellular levels in cancer cells are likely to differ from those in normal cells. The rules are based on the supposition that lower levels of enzymes catalyzing the production of a metabolite (or higher levels of enzymes catalyzing the consumption of it) imply decreased level of such metabolite, and vice versa (see Figure 1). In our methodology, a given metabolite is predicted to have decreased levels in cancer cells when: 1) both of the following apply: 1.1) there is no gene encoding for an enzyme able to catalyze the production of the metabolite whose differential expression status is G1 (upregulated in cancer cells) or G3 (significantly expressed at similar levels in normal and cancer cells) and 1.2) there is no gene encoding for an enzyme able to catalyze the consumption of the metabolite whose differential expression status is G2 (downregulated in cancer cells), and 2) either or both of the following applies: 2.1) there is at least one gene encoding for an enzyme able to catalyze the production of the metabolite whose differential expression status is G2 (downregulated in cancer cells) and 2.2) there is at least one gene encoding for an enzyme able to catalyze the consumption of the metabolite whose differential expression status is G1 (upregulated in cancer cells). Similarly, a metabolite is predicted to have increased levels in cancer cells when: 1) both of the following applies: 1.1) there is no gene encoding for an enzyme able to catalyze the consumption of the metabolite whose differential expression status is G1 or G3 and 1.2) there is no gene encoding for an enzyme able to catalyze the production of the metabolite whose differential expression status is G2, and 2) either or both of the following applies: 2.1) there is at least one gene encoding for an enzyme able to catalyze the consumption of the metabolite whose differential expression status is G2 and 2.2) there is at least one gene encoding for an enzyme able to catalyze the production of the metabolite whose differential expression status is G1. Thus, the methodology attempts to consider, as much as is practical, the entire proteome complement of enzymes that produce and consume the metabolite. That is, our approach is Systems Biology based.