Harnessing Self-Organizing Maps Algorithm: Classification of Nuclear Magnetic Resonance Spectra for Untargeted Metabolomics of Breast Cancer

Metabolomics has emerged as a promising tool for identifying disease biomarkers, and Nuclear Magnetic Resonance (NMR) spectroscopy enables the simultaneous detection of a wide range of metabolites [1]. However, due to the complex interactions in metabolic networks, metabolites often exhibit high correlation and collinearity. Principal Component Analysis (PCA) and Partial Least Squares-Discriminant Analysis (PLS-DA) have been commonly used in metabolomic studies [2]. Nevertheless, to extract meaningful information and ensure accurate biological interpretation, raw NMR spectral data requires preprocessing, which may introduce its own set of challenges. Self-Organizing Maps (SOMs), a learning method that utilizes artificial neural networks (ANNs) to visualize different patterns in data, offering a powerful alternative to PCA, especially for nonlinear data [3]. Kohonen Maps are self-organizing systems applied to unsupervised problems. In addition, counter-propagation artificial neural network (CP-ANN) are very similar to the Kohonen Maps and are essentially based on the Kohonen approach, but combine characteristics from both supervised and unsupervised learning. In this article, the Kohonen map and CP-ANN algorithms were employed to visualize the relationships between metabolites in control and breast cancer (BC) patient samples to identify potential biomarkers. Metabolites were extracted from blood serum samples of the control group (n=42) and BC patients (n=18) using methanol and chloroform as solvents. The acquired 1HNMR spectra were analyzed using Mestrenova software and subsequently underwent phase and baseline corrections, and regions related to the solvent and those without informative data were excluded. By analyzing the obtained topological map, several important variables, including lactic acid, cysteine, glucose, serine, and tagatose were identified as influential factors in differentiating between healthy and BC patient classes. Also, the use of CP-ANN algorithm was able to successfully differentiate the two classes with accuracy and sensitivity of 91% and 90% for the control group and patients, respectively. These findings shed light on potential biomarkers for BC diagnosis, and the application of the SOMs algorithm proves to be a valuable approach to exploring complex metabolomic data.