What question did this study set out to answer?

March 14, 2026Open Access

ATR-FTIR Spectroscopy Combined with Machine Learning as a Cost-Effective Screening Tool for SNP Genotyping

Key Points

To evaluate ATR-FTIR spectroscopy combined with machine learning for SNP genotyping.
Utilized ATR-FTIR spectroscopy for genotyping according to UCP1 SNPs.
Applied real-time quantitative PCR for DNA amplification and analysis.
Conducted principal component analysis for spectral data interpretation.
Employed various machine learning models, including deep learning, to classify SNP genotypes.
Achieved approximately 65% AUC and 71% accuracy in differentiating AA and GG genotypes.
PCA indicated substantial overlap among genotypes despite moderate classification performance from supervised models.
Identified spectral regions that effectively reduced reagent interference, aiding in SNP discrimination.

Abstract

Single nucleotide polymorphisms (SNPs) play a central role in genetic susceptibility to multifactorial disorders such as obesity and type 2 diabetes mellitus (T2DM), highlighting the need for scalable and cost-effective genotyping strategies. This study evaluates the feasibility of combining attenuated total reflection Fourier-transform infrared (ATR-FTIR) spectroscopy with machine learning (ML) approaches to discriminate amplified fragments of DNA by polymerase chain reaction (PCR amplicons) according to a single nucleotide polymorphism (SNP) in the UCP1 gene named −3826 A/G. Genomic DNA from 190 individuals was genotyped by real-time quantitative PCR (qPCR SNP genotyping) using TaqMan assays and subsequently amplified by qualitative PCR to generate amplicons for spectral analysis. ATR-FTIR spectra were acquired under standardized conditions, including negative template controls (NTC), and processed using standard normal variate (SNV) normalization. Difference spectra obtained by subtracting the mean NTC spectrum enabled the identification of spectral regions with reduced reagent interference. Exploratory principal component analysis (PCA) was applied to selected spectral ranges, followed by supervised ML and deep learning (DL) models for pairwise genotype discrimination. While PCA revealed substantial overlap among genotypes, supervised models achieved moderate but consistent classification performance. The highest classification accuracy was achieved by a DL multilayer perceptron with a residual-simulated architecture, with an AUC ≈ 65.4% and an accuracy ≈ 71.6% in the AA×GG comparison, while logistic regression yielded the best AUC and accuracy among conventional models when applied to DNA fingerprint regions (900–1100 and 950–1200 cm⁻¹). Although not intended to replace established genotyping methods, this approach demonstrates potential as a rapid, low-cost screening tool to prioritize samples for confirmatory analysis, supporting the development of complementary workflows for large-scale SNP investigations. • ATR-FTIR spectra and machine learning enabled SNP genotype discrimination • Spectral range selection strongly influenced model performance outcomes • Best model achieved ~65% AUC and ~71% accuracy between AA and GG genotypes • SNP variants induced subtle but detectable vibrational changes in DNA amplicons • Findings represent an initial framework with clear potential for refinement

ATR-FTIR Spectroscopy Combined with Machine Learning as a Cost-Effective Screening Tool for SNP Genotyping

Key Points

Abstract

Cite This Study