Classification of autism related molecular signatures by simple machine learning models applied to ATR-FTIR spectra of pediatric sera

Background Autism spectrum disorder (ASD) begins early and continues for life with strong personal and societal implications. Complexity of the condition and absence of auxiliary methods that can complement clinical evaluation make an early identification challenging. ASD is reported to be associated with metabolic changes and hence biochemical alterations in the affected children. This study is one of the initial investigations to employ Attenuated Total Reflectance–Fourier Transform Infrared (ATR-FTIR) spectroscopy coupled with machine learning to identify the biochemical alterations in the blood sera of children with ASD as compared to controls.Methods ATR-FTIR spectra from lyophilized blood sera of 13 children with ASD, in the age range of 1.8 to 8 years along with 9 age-matched neurotypical controls were obtained and baseline corrected using rubber-band method. After baseline correction and vector normalization, spectral features for both groups were analyzed using principal component analysis (PCA) for dimensionality reduction, followed by supervised classification with linear discriminant analysis (LDA).Results Results of LDA applied to three selected PCs, including PC3 (4.38% variance, p = 0.0021), PC5 (2.18%, p = 0.095), and PC6 (1.31%, p = 0.083) under leave-one-out cross-validation achieved 86.36% accuracy, 92.31% sensitivity, 77.78% specificity, 88.89% F1-score, and AUC-ROC of 0.803. Analysis of the pre-processed mean spectra identified eight prominent spectral variations including two in the protein and lipid C-H stretching region (3293 and 2962 cm−1) and six in the bio-fingerprint region (1630, 1533, 1456, 1394, 1237, and 1086 cm−1).Conclusions We conclude that ATR-FTIR spectroscopy coupled with machine learning can detect biochemical variations in the sera of children with and without ASD by applying LDA to PCA-reduced pre-processed spectra.