Abstract
Motivation Analysis of mixed DNA profiles is one of the common tasks for the forensic practitioners. Due to the complexity of the data the analysis is often performed with Bayesian probabilistic genotyping algorithms. These trade off the precision of the results against the execution time. When the default settings are used, as large as a 10-fold change in likelihood ratios (LR) might be observed when the software is run twice on the same case. In the past this variance has been attributed to the stochasticity of Markov chain Monte Carlo (MCMC) algorithm. As likelihood ratios translate directly to the strength of evidence in a trial, the users of the algorithms expect that the resulting likelihood ratios will be near to identical between runs.
Results We present an implementation of Hamiltonian Monte Carlo that achieves this goal. Our method reduces the standard deviation of the log10 likelihood ratio around 10 times in case of a previously analysed mixture. We show that choice of the used convergence metric influences the precision. The method classifies correctly contributors and non-contributors in case of selected MIX05, MIX13, and ProvedIt mixtures closely reproducing previously published results. Our software implementation, however, is the first one to benefit from GPU acceleration. The inference process is quick and on a single-replicate mixtures we considered it takes less than 7 minutes for 3 contributor mixtures, less than 35 minutes for 4 contributor mixtures and less than an hour for 5 contributor mixtures with one known contributor.
Competing Interest Statement
The authors have declared no competing interest.