Machine learning-assisted analysis of saliva-containing mixtures using a saliva-specific multi-CpG-microhaplotype marker system.

Saliva is a common and important type of biological evidence in forensic casework; however, detecting saliva-derived alleles in complex mixtures remains challenging. In this study, based on saliva-specific single-CpG sites, highly correlated neighboring CpG sites were further screened and integrated as multiple CpGs (multi-CpGs), which were then combined with adjacent microhaplotype information to construct a multi-CpG-microhaplotype marker system. These markers were detected using massively parallel sequencing (MPS) and combined with the k-nearest neighbors (KNN) algorithm to construct allele-level classification models for identifying saliva-derived alleles in mixtures. During model development, the features derived from multi-CpGs showed better overall allele-level classification performance than the corresponding single-CpG features, supporting their further application in mixture analysis. In the independent test set, the KNN model based on multi-CpGs achieved an allele-level classification accuracy of 94.08%. Based on the model prediction results, saliva detection was supported across mixture ratios ranging from 1:49 to 8:1 (saliva: non-saliva), and preliminary auxiliary support for exploratory contributor inference was obtained in saliva-containing mixtures. In conclusion, the combined analytical framework based on the multi-CpG-microhaplotype marker system and machine learning provides a new complementary strategy for identifying saliva-derived alleles in mixtures, and shows promising potential for forensic saliva-containing mixture analysis.