DNA methylation-based assay and Random Forest classification model for identification of biological materials

Determination of the biological origin of trace material is an important forensic tool for supporting crime investigations. Differentially methylated DNA positions (DMPs) can be used to discriminate among forensically relevant biological materials. In this study we developed and validated a multiplex assay that targets 13 DMPs using combined bisulfite conversion and SNaPshot (single-base extension) analysis for the identification of semen, blood, menstrual blood, vaginal secretion, nasal secretion, saliva, and skin swabs. The assay also incorporates a DNA methylation site for sex determination and two sites for control of bisulfite conversion. We evaluated the specificity of DMPs for their intended biological material and uncovered previously unknown overlapping DNA methylation profiles by additionally testing buccal mucosa and hair root samples. To translate DNA methylation levels into biological source predictions, two Random Forest (RF) models were built: RF1 for identifying single‑component samples (trained on 284 experimental single-component samples) and RF2 for predicting both single‑component samples and mixtures (trained on 284 experimental single-component samples and 2870 in-silico-generated mixtures). In validation experiments RF1 achieved high accuracy for identification of single-component samples (92.7%, tested on 124 samples). The prediction performance is limited to bisulfite-conversion input of 1 ng down to 0.25 ng DNA depending on the biological material. Analyzing two-component mixtures, the strong linear correlation between the DNA methylation levels and mixture component proportions was confirmed. In-silico mixture approach of RF2 showed a promising performance to identify even unbalanced experimental mixed samples (n = 120), proving the potential for practical applications.