Decoding myasthenia gravis: advanced diagnosis with infrared spectroscopy and machine learning

Abstract

Myasthenia Gravis (MG) is a rare neurological disease. Although there are intensive efforts, the underlying mechanism of MG still has not been fully elucidated, and early diagnosis is still a question mark. Diagnostic paraclinical tests are also time-consuming, burden patients financially, and sometimes all test results can be negative. Therefore, rapid, cost-effective novel methods are essential for the early accurate diagnosis of MG. Here, we aimed to determine MG-induced spectral biomarkers from blood serum using infrared spectroscopy. Furthermore, infrared spectroscopy coupled with multivariate analysis methods e.g., principal component analysis (PCA), support vector machine (SVM), discriminant analysis and Neural Network Classifier were used for rapid MG diagnosis. The detailed spectral characterization studies revealed significant increases in lipid peroxidation; saturated lipid, protein, and DNA concentrations; protein phosphorylation; PO₂^-asym + sym /protein and PO₂^-sym/lipid ratios; as well as structural changes in protein with a significant decrease in lipid dynamics. All these spectral parameters can be used as biomarkers for MG diagnosis and also in MG therapy. Furthermore, MG was diagnosed with 100% accuracy, sensitivity and specificity values by infrared spectroscopy coupled with multivariate analysis methods. In conclusion, FTIR spectroscopy coupled with machine learning technology is advancing towards clinical translation as a rapid, low-cost, sensitive novel approach for MG diagnosis.

Figure 1

The average serum spectra of the control and MG groups in different regions. The average serum absorbance and difference spectra (solid red line) of the MG (N = 24) and healthy control (N = 42) groups (A) in the 3800–3020 cm^-1 region. The spectra were normalized with respect to the CH₂ antisymmetric band located at 2929 cm^-1. (B) in the 3020–2800 cm^-1 regions. The spectra were normalized with respect to the amide A band located at 3286 cm^-1. The enlarged olefinic band spectra are given in the sub-panel. (C) in the 1800–800 cm^-1 region. The spectra were normalized with respect to the amide A band located at 3286 cm^-1. Enlarged difference spectra are given in the sub-panel. Difference IR spectra were obtained by subtracting the control spectrum from the MG groups’ spectra. Band assignments are given in Table 1. Control was represented by C, Myasthenia Gravis by MG, and Difference spectra by MG-C dif. This figure was given for a rough visual representation of the MG-induced spectral changes. The data were generated using OPUS 8.v1 software (https://www.bruker.com/en/products-and-solutions/infrared-and-raman/opus-spectroscopy-software/downloads.html).

Figure 2

Visual representation of infrared spectral biomarkers for MG diagnosis: (A) Comparison of area ratios of MG patient serum biomolecules with the control group, (a) Amide A amount, (b) Unsaturation index, (c) Saturated lipid concentration, (d) Protein concentration, (e) Protein Phosphorylation, (f) Protein structural changes, (g) PO₂^- antisym. + sym./Protein, (h) PO₂^- sym./Total lipid, (i) DNA concentration, (j) RNA concentration, and (k) RNA/DNA. (B) Comparison of protein and lipid band position and bandwidth values (a) Amide I band position, (b) Amide I bandwidth, (c) CH₂ antisym. stretching bandwidth. An unpaired t-test was applied to the samples. (*: indicates that the patient group was compared with the control; *p < 0,05, **p < 0,01, ***p < 0,001), **** p < 0,0001) (C: Control group, MG: Myasthenia Gravis patient group). The figure shows, in addition to MG-induced contextual and concentration changes of serum biomolecules, a decrease in the unsaturation index, i.e., an increase in the lipid peroxidation, variations in the amide I band position and bandwidth, i.e., protein structural changes, and a decrease in the bandwidth of CH₂ antisymmetric stretching lipid band, i.e., a decrease in lipid dynamics. The graphs were generated using GraphPad Prism version 7.0 for Windows, GraphPad Software, Boston, Massachusetts USA, www.graphpad.com.

Figure 3

PCA score plots in the PC1-PC2 plane of the Control (C), and MG samples obtained from the second derivative of unit vector normalized spectra (A) for the whole region, (B) for the lipid region, (C) for the protein region, and (D) for the fingerprint region. PCA scores of the Control and MG samples are separated into two clusters. The figures were generated by "The Unscrambler X" Version 10.3, CAMO Software Inc., Oslo Norway(www.camo.com/unscrambler/).

Figure 4

Confusion tables showing the number of observations of the Control and MG samples (First column) and the Receiver Operating Characteristic (ROC) (Second column) for the SVM classifier models obtained for the (A) whole, (B) lipid, (C) protein, and (D) fingerprint regions. The SVM classifier correctly predicts all 42 control samples and 24 MG samples by the classifier designed for the protein region. ROC curves show the plots of sensitivity vs. (1-specificity) for the MG mode classifiers. The models were generated via MATLAB R2023a, Classification Learner App software(The MathWorks Inc. https://www.mathworks.com).