Automated single-nucleotide polymorphism analysis using fluorescence excitation–emission spectroscopy and one-class classifiers

We have developed a new method of highly automated SNP (single nucleotide polymorphism) analysis for identification of genotypes. The data were generated by the Taqman reaction. A total of 18 half-plates were analysed for different genes, each consisting of 48 wells, including six synthetic DNA samples, three background samples, and 39 human DNA samples. Fluorescence spectra were obtained from each well. The characteristics of the spectra depended on whether the genotype originated from one of three classes—homozygotic wild-type, mutant, or heterozygote. The main problems are: (1) spectral variation from one half-plate to another is sometimes very substantial; (2) the spectra of heterozygotic samples vary substantially; (3) outliers are common; and (4) not all possible alleles are represented on each half-plate so the number of types of spectra can vary, depending on the gene being analysed. We solved these problems by using a signal-standardisation technique (piecewise direct standardisation, PDS) and then built two one-class classifiers based on PCA models (PCA data description) to identify the two types of homozygote. The remaining samples were tested to see whether they could be approximated well by a linear combination of the spectra of two types of homozygote. If they could, they were identified as heterozygotic; if not, they were identified as outliers. The results are characterised by very low false-positive errors and 2 to 6% overall false-negative errors.