Making AI accessible for forensic DNA profile analysis.

Area of Science:

• Forensic Science
• Bioinformatics
• Computer Vision

Background:

• Deep learning offers potential for automating allele calling in forensic DNA analysis.
• Current methods often use custom model architectures and extensive manual annotations, hindering reproducibility.
• There is a need for accessible and standardized deep learning approaches in forensic genetics.

Purpose of the Study:

• To investigate the efficacy of training a deep learning model using casework data and a standard U-Net architecture for forensic DNA allele calling.
• To evaluate the performance of the developed model, DNANet, against human analysts and established ground truths.
• To promote accessibility by making code, model weights, and data publicly available.

Main Methods:

• Developed DNANet, a U-Net based deep learning model, for classifying electropherogram scan points as allele or non-allele.
• Trained the model using allele annotations derived from casework data.
• Evaluated DNANet on unseen casework data and independent mixture research data, comparing its performance to analyst annotations and actual donor alleles.

Main Results:

• DNANet achieved high F1 scores: 0.971 on analyst-annotated case data and 0.982 on research data.
• Performance on actual donor alleles (F1 score 0.962) was comparable to that of human analysts.
• The model demonstrated that good performance can be achieved with standard data and architectures.

Conclusions:

• DNANet's performance is on par with human analysts in forensic DNA allele calling, validating the use of standard deep learning architectures and casework data.
• The study highlights the potential for automating complex forensic tasks with accessible deep learning tools.
• A call is made for establishing standardized benchmarks to facilitate quantitative comparisons of allele calling systems.