Metabolomics by NMR Combined with Machine Learning to Predict Neoadjuvant Chemotherapy Response for Breast Cancer

Abstract

Neoadjuvant chemotherapy (NACT) is offered to patients with operable or inoperable breast cancer (BC) to downstage the disease. Clinical responses to NACT may vary depending on a few known clinical and biological features, but the diversity of responses to NACT is not fully understood. In this study, 80 women had their metabolite profiles of pre-treatment sera analyzed for potential NACT response biomarker candidates in combination with immunohistochemical parameters using Nuclear Magnetic Resonance (NMR). Sixty-four percent of the patients were resistant to chemotherapy. NMR, hormonal receptors (HR), human epidermal growth factor receptor 2 (HER2), and the nuclear protein Ki67 were combined through machine learning (ML) to predict the response to NACT. Metabolites such as leucine, formate, valine, and proline, along with hormone receptor status, were discriminants of response to NACT. The glyoxylate and dicarboxylate metabolism was found to be involved in the resistance to NACT. We obtained an accuracy in excess of 80% for the prediction of response to NACT combining metabolomic and tumor profile data. Our results suggest that NMR data can substantially enhance the prediction of response to NACT when used in combination with already known response prediction factors.

Keywords: 1H-NMR; breast neoplasms; drug resistance; magnetic resonance spectroscopy; metabolism; untargeted metabolomics.

Figure 1

Representative 1H-NMR spectra from serum obtained from women before they initiated the NACT. Resistant and sensitive refer to response to neoadjuvant chemotherapy.

Figure 2

Volcano plot showing the metabolite variation by NACT outcomes (log2(resistant/sensitive)) in function of their statistical significance (log2(p-value)). Non-significant metabolites are plotted below the horizontal dashed line, and the metabolites above this line presented significant variation (p-value < 0.05). In addition, we show a boxplot for metabolites with significant variation (blue represents resistant patients and pink the sensitive ones); p-values were calculated using the Mann–Whitney–Wilcoxon test or t-test as a function of whether the data came from a normal distribution, proved with the Shapiro test.

Figure 3

Classification models obtained with the combination of metabolites and clinical information. (A) The plot of the ROC curves for each model, containing the area under the curve (AUC). (B) The coefficients of the predictors and (C) performance of models II and V for training (cross-validation) and validation sets.