Accurate and consistent identification of allele peaks in STR electropherograms (EPGs) is fundamental to forensic DNA interpretation, particularly in complex and low-template cases. Manual peak annotation and allele calling remain labour-intensive and are subject to inter- and intra-analyst variability.

Deep learning offers a potential route towards more objective and efficient allele peak identification, but existing approaches often rely on bespoke architectures and extensive, specially curated training data, which can limit adoption in routine forensic practice.

In this study, we investigate whether a widely used deep-learning architecture (U-Net) can be trained on data derived directly from routine casework to perform automated allele peak detection in STR EPGs.

Analyst-called alleles from casework profiles were converted into segmentation labels at the scan-point level, enabling the resulting model, “DNANet”, to classify each scan point in an EPG as “allele” or “non-allele”. This formulation is analogous to pixel-wise segmentation in image analysis, but applied to EPG signal traces.

We evaluated DNANet on unseen casework profiles as well as on independent mixture research data, using analyst annotations as ground truth.

In addition, for the research mixtures we compared both DNANet and analyst annotations against the known donor genotypes, providing an allele-level performance benchmark.

DNANet achieved an F1 score of 0.962, which was equivalent to the F1 score calculated from analyst annotations, indicating that its performance in peak detection is on par with trained forensic DNA analysts following standard laboratory procedures. 

These results demonstrate that robust automated allele peak detection can be achieved using: (i) data that are already generated in routine forensic casework, and (ii) a standard, well-documented deep-learning architecture.
To facilitate evaluation and further development by the forensic genetics community, we provide open access to the DNANet code, trained model weights and the research mixture data.

Ongoing work includes:

• generating more detailed, manually curated peak annotations and assessing the added value of these richer labels, supported by a custom annotation tool that will also be made openly available.
• generating synthetic EPGs that allow training AI models on more data which has an actual ground truth from which the EPG is synthesized.
• making integration of custom datasets more accessible to other laboratories around the world can extend the work in this paper.

https://www.sciencedirect.com/science/article/pii/S1872497325001255

Authors

• Abel de Wit (Netherlands Forensic Institute, Netherlands) 
• Claire Wagenaar (Netherlands Forensic Institute, Netherlands) 
• Nathalie Janssen (Netherlands Forensic Institute, Netherlands) 
• Brechtje Hoegen (Netherlands Forensic Institute, Netherlands) 
• Judith van de Wetering (Netherlands Forensic Institute, Netherlands) 
• Huub Hoofs (Netherlands Forensic Institute, Netherlands) 
• Simone Ariëns (Netherlands Forensic Institute, Netherlands) 
• Corina Benschop (Netherlands Forensic Institute, Netherlands) 
• Rolf Ypma (Netherlands Forensic Institute, Netherlands)