Non-parametric simulation of DNA mixture profiles from one-person trace profiles

Evidential test data of genotyping softwares must clearly carry the characteristics of real-life evidential trace profiles. Ensuring that simulated data, which is used in research and validation, is similar and behaves similarly to real-life DNA mixture profiles, is important for making sure that the software is reliable and trustworthy when assessing strength of evidence. In present day, mixtures for such purposes are typically generated in vitro (laboratory-created mixtures) or in silico (peak heights are sampled from a model). Both of these simulation methods have their drawbacks, as simulating in vitro can be expensive and is heavily dependent on the availability of physical contributors, and simulating in silico relies on a specific pre-decided statistical or machine learning model which approximates the real-life mixture generating process. In this paper, we examine a mixture simulation approach, which is a hybrid of these two strategies. To simulate mixture profiles, we use one-person trace profiles, which can be modified to be originating from any person with a known genotype, and paste them together. This approach makes it very easy to create mixture profiles from any set of contributors, including related donors, while maintaining the variability of the used profiles. To illustrate, we replicate 2-4-person trace profiles from PROVEDIt and compare them to the mixtures they were replicated for. The results show that the simulated mixtures are very similar and behave very similarly to the mixtures from PROVEDIt when compared by likelihood ratio distributions and various estimated parameters obtained from DNAStatistX, such as expected peak height, coefficient of variation, etc. To investigate the difference with a fully in silico approach, we also compare the non-parametrically replicated mixtures to mixtures simulated with EuroForMix and highlight the differences between the two approaches.