Automated estimation of the number of contributors in autosomal short tandem repeat profiles using a machine learning approach. (November 2019)

• Record Type: Journal Article
• Authors: Benschop, Corina C.G.; van der Linden, Jennifer; Hoogenboom, Jerry; Ypma, Rolf; Haned, Hinda
• Is Part Of: Forensic science international. Volume 43(2019)
• Publication Date: 2019-11
• Subjects: DNA profile interpretation -- DNA mixtures -- Number of contributors -- Machine learning; Forensic genetics -- Periodicals; Génétique légale -- Périodiques; Forensic genetics; Electronic journals; Periodicals; 614.1
• DOI: 10.1016/j.fsigen.2019.102150
• Languages: English
• ISSNs: 1872-4973

Abstract

Highlights: A machine learning model to assign the number of contributors was developed. Information on allele counts, frequencies and peak heights was included. The selected model was a random forest classifier with 19 features. This NOC machine learning model was trained, tested and validated. The NOC machine learning model outperformed MAC and nC-tool. Abstract: The number of contributors (NOC) to (complex) autosomal STR profiles cannot be determined with absolute certainty due to complicating factors such as allele sharing and allelic drop-out. The precision of NOC estimations can be improved by increasing the number of (highly polymorphic) markers, the use of massively parallel sequencing instead of capillary electrophoresis, and/or using more profile information than only the allele counts. In this study, we focussed on machine learning approaches in order to make maximum use of the profile information. To this end, a set of 590 PowerPlex® Fusion 6C profiles with one up to five contributors were generated from a total of 1174 different donors. This set varied for the template amount of DNA, mixture proportion, levels of allele sharing, allelic drop-out and degradation. The dataset contained labels with known NOC and was split into a training, test and hold-out set. The training set was used to optimize ten different algorithms with selection of profile characteristics. Per profile, over 250 characteristics, denoted 'features', were calculated. These features were based on allele counts, peak heights and allele frequencies. The features that were most related to the NOC were selected based on partial correlation using the training set. Next, the performance of each model (=combination of features plus algorithm) was examined using the test set. A random forest classifier with 19 features, denoted the 'RFC19-model' showed best performance and was selected for further validation. Results showed improved accuracy compared to the conventional maximum allele count approach and an in-house nC-tool based on the total allele count. The method is extremely fast and regarded useful for application in forensic casework.