Using DNA traces obtained from firearms to distinguish between active use and passive contact with weapons

Sammanfattning

Firearm-related crime presents a significant challenge in forensic investigations, particularly when in court indirect contact is invoked as an alternative explanation for the presence of DNA on a weapon. This study aimed to characterise location-specific DNA distribution patterns from different firearm handling scenarios and to assess whether these patterns can support activity-level evaluation in casework. Four experimental scenarios were conducted using semi-automatic pistols: controlled and semi-controlled direct transfer via loading and shooting (handling) and indirect transfer via a personal towel or shirt (storage). DNA traces were collected by swabbing 13 locations at each firearm, quantified by qPCR and profiled by STR analysis. A random forest machine learning (ML) classifier was trained on the combined qPCR and STR parameters across all targeted locations to distinguish between direct and indirect transfer. A simulated crime was performed on two firearms with unknown handling history to validate the experimental design under casework like conditions. Decontamination with the sodium hypochlorite and ethanol protocol applied with a toothbrush was successfully shown to reduce background DNA below the detection limit across all firearm surfaces. In controlled loading and shooting, DNA was concentrated mostly at the grip and magazine surfaces, whereas semi-controlled scenario had distributions dominated by the grip and slide back. Indirect transfer via personal belongings yielded lower total DNA concentrations compared to direct handling. The random forest (RF) classifier achieved 99,4% ± 1,7% accuracy trained and validated on real data and 98,6% ± 3,4% trained and validated on augmented and real data. The high accuracy can be a consequence of using only two classes (binary classification), data provided to RF was from only controlled studies or too many features caused RF to learn outliers and noise of dataset (data overfitting). It should be interpreted as a proof-of-concept results rather than an indication of performance in real casework conditions. ML model showed that the magazine lip and body, and frame front left were the discriminating locations to determine between direct and indirect transfer. The simulated crime examination produced distribution patterns consistent with controlled experiments, with both models correctly classifying all contributors with their handling scenario.