Comparison of 67 metabolite concentrations from healthy subjects (n:62) and subjects with breast cancer (n:38) revealed significant differences. Application of multivariate statistical data analysis (OPLS-DA) to this dataset resulted in distinction between individuals with breast cancer and those without. Five of the healthy individuals overlapped with the breast cancer category. The model parameters and validation of the PLS-DA (multivariate statistical data analysis) suggested a good model. OPLS-DA class prediction was performed as for the EOC subjects, on a total of 20 subjects, 10 each of breast cancer and healthy. As may be observed, all breast cancer and healthy test subjects were correctly classified [21].

Analysis of urinary metabolite changes revealed that many metabolites decreased in relative concentration with a cancer phenotype when compared with healthy. That the majority of urinary metabolites appeared to decrease in concentration in cancer patients is a similar result to what has been seen in colon cancer tissue metabolomics. Interestingly, some metabolites that were shown to increase in cancer tissue (such as some of the amino acids) were lower in the urine of cancer patients. Concentrations of many amino acids decrease in cancer patients relative to healthy. Decreases in tricarboxylic acid (TCA) cycle intermediates are suggestive of a suppressed TCA cycle. In a study of urinary markers of colorectal cancer, it was observed that several TCA cycle intermediates decrease in those with colorectal cancer as compared with those without [46]. The biological reason behind the metabolite changes is largely speculative at this point but likely involves a shift in energy production, as tumors rely primarily on glycolysis as their main source of energy. This phenomenon is known as the Warburg effect [47], and decreases in TCA cycle intermediates and glucose in the urine could be indicative of this phenomenon. Clearly, lower glucose concentrations were observed in women with ovarian cancer as compared with breast cancer. This could be because of the fact that more of the women with ovarian cancer were in an advanced stage of the disease. Furthermore, the use of amino acids by tumors requires the upregulation of amino acid transporters, [48] pulling these metabolites from the blood. Decreases in circulating glucose and amino acids could subsequently result in an overall decrease in energy metabolism elsewhere in the body, diminishing other metabolic pathways such as the urea cycle, resulting in lower concentrations of urea and creatine, and potentially affecting gut microbial population and/or metabolism.

So, it is suggested that a urine test is faster, easier to administer, less costly, and noninvasive and could be used as a prescreen to other forms of more invasive or uncomfortable screening.

ER and PgR status were best predicted by PLS-DA (Tables 10.2 and 10.3). For ER status, the number of correctly classified blind samples were 44/50 and 42/50 for Kennard-Stone and SPXY sample selection, respectively, while PgR status had a correct blind sample classification of 39/50 for the Kennard-Stone test set and 36/50 for SPXY. Similar results for both Kennard-Stone and SPXY sample selection indicate robust classification by PLS-DA. The sensitivity and specificity of classification were approximately equal; this is in contrast to the results of PNN and BBN where the sensitivity was higher than the specificity. The higher sensitivity may be due to the fact that, especially for ER status, there are more positive than negative samples. This could lead to networks that are more specialized in recognizing positive than negative samples. Since the probability of a sample being positive is much higher than the probability of it being negative, the network achieves a greater number of total correct classified samples by classifying most of the samples as positives. In PNNs, this can be partly overcome by the customized fitness function, allowing the user to insert a penalty whenever a negative sample is classified incorrectly. In this study, the same penalty was used for both the Kennard-Stone and the SPXY training and test sets. Although this improved the classification ability of the networks compared to networks without penalty, the classification error was still higher than that achieved by PLS-DA.

A PLS-DA model of the whole dataset with three latent variables (LVs) explains 43.8 % of the X-variance and 42.7 % of the Y-variance. The score values for ER + and ER- samples are significantly different for all three LVs (t-test, p < 0.001), and it is possible to discriminate between ER + and ER- samples in a score plot of LV1, LV2, and LV3. ER + and ER- samples are mainly separated on the first LV that represents 70 % of the Y-variance explained by the model, and ER- samples have higher score for LV1 than ER + samples. The loading profile for LV1 reveals that samples with higher score for LV1 have more of the metabolites glycine (Gly), glycerophosphocholine (GPC), choline (Cho), and alanine (Ala) and less ascorbate (Asc), creatine (Cr), taurine (Tau), and phosphocholine (PC) than samples with lower LV1 scores. The regression vector of the PLS-DA model gives an indication of the overall influence of the variables based on all three LVs. The regression vector of ER- samples appears similar to LV1 and shows the same metabolic patterns. In addition, lactate (Lac) appears to be more expressed in ER- samples.

Axillary lymph node status was best predicted by BBN with 34 of 50 blind samples correctly classified. However, this was only true for the samples chosen by SPXY sample selection, and the same number of correctly classified samples was not achieved using Kennard-Stone sample selection. PLS-DA and BBN gave similar results, and overall, all three methods gave unacceptably high classification errors. However, the number of correctly classified samples was better than expected by chance for all methods. This indicates that there is a difference between the MR spectra of lymph node-positive and lymph node-negative patients and that the metabolic profile is altered in patients with lymphatic spread compared to patients without spread.

In conclusion, ER and PgR status were successfully predicted by MR metabolomics. There is also a relationship between metabolic profile and lymph node status, although prediction of lymph node status based on MR spectra did not reach a reliable level of correctly classified samples. By combining MR spectroscopy with multivariate modeling, the biological differences between different metabolic profiles could be revealed. Here hormone receptor-negative patients appear to have more of the metabolites glycine (Gly), glycerophosphocholine (GPC), and choline (Cho) than receptor-positive patients. The data also indicate different metabolic profiles between ER status and PgR status. Thus, this study has shown that MR profiles contain prognostic information that may be of benefit in treatment planning and patient follow-up, and MR metabolomics may become an important tool for clinical decision-making in breast cancer patients.

The appeal of metabolomics is concurrent assessment of tumor and host. Indeed, survival of a specific tumor in a specific host relies on a dynamic interaction, with evasion of normal host immunity and favorable stromal environment for metastatic deposits as key factors. In this recent study, metastatic subjects were characterized by higher values of phenylalanine, glucose, proline, lysine, and N-acetyl cysteine and lower values of lipids, when compared to the spectra of both post- and preoperative patients [60].

A strength of metabolomics, as compared with current prognostic tools, may be confirmation rather than assumption of micrometastatic disease. Results reveal differential metabolomic fingerprints for most early and metastatic breast cancer patients. Among the normal noise of the metabolomic fingerprint, most patients were distinguished based on metabolomic analysis of one serum sample [60].

Metabolomic analysis assigns more patients to low risk than are assigned by Adjuvantionline. Similarly, when compared with conventional clinical and pathological factors, prognostic gene expression signatures generally identify more patients of low risk. The 21-gene Oncotype Dx shows direct concordance of 36 % in relapse risk stratification compared with an adjusted Adjuvantionline [61]. The 70-gene MammaPrint, when compared with Adjuvantionline, had stronger predictive power and provided lower-risk estimates for more patients [62]. These low-risk patients may be spared or receive less intensive adjuvant treatment.

In conclusion, the benefit of metabolomics is the incorporation of both a specific tumor profile with metastatic features and a specific host profile conducive to tumor growth. A preliminary exploration in a limited number of patients of a potential role for the evolving field of metabolomics in assessment of micrometastatic disease in early breast cancer has been presented. Clearly, this approach requires refinement and validation, but the distinction identified between early and late disease and the prognostic role of the metabolomic fingerprint provide an exciting platform for further work.

The development of a metabolomic-based profile for the early detection of breast cancer recurrence is presented in a recent study [23]. The investigation makes use of a combination of analytical techniques, NMR and MS, and advanced statistics to identify a group of metabolites that are sensitive to the recurrence of breast cancer.