Use of Graphical Models in analysis of DNA-mixtures

Specialeforsvar ved Malte Bødkergaard Nielsen

Titel: Use of Graphical Models in analysis of DNA-mixtures


Abstract:  Based on work by Cowell, Graversen, Lauritzen, and Mortera on forensic DNA analysis, I describe how Bayesian networks can serve as a model for analysis of mixed DNA traces. I provide an introduction to the process of producing electropherograms, which constitute the combined signal from the DNA profiles of the contributors to a trace. A gamma model for each peak height in an electropherogram is presented as well as a review of how Bayesian network techniques can be used to incorporate these into an efficient network capable of exact computations over the vast state space of potential genotypes. This efficiency allows for calculations in networks with up to six unknown contributors, using standard hardware. The model's flexibility is outlined by various network extensions to handle artefacts and to create model diagnostics that can be used to assess the model's adequacy. The model is used in a case analysis that combines information from multiple electropherograms both to perform a deconvolution of profiles and to compute the likelihood ratio of a suspect being present in the mixture against a random person being present. Finally, the benefits and challenges are discussed: Most notable is the model's ability to perform likelihood ratio evaluation and mixture deconvolution in a unified framework, which ensures consistency, but it all relies on an assumption on the number of unknown contributors as well as on a set of parameters found through numerical optimization methods, and not least on the estimated allele frequencies for a reference population.




Vejledere: Steffen Lauritzen, Therese Graversen
Censor:      Torben Tvedebrink, Aalborg Universitet