Validation of a neural network approach for STR typing to replace human reading

Area of Science:

• Forensic Science
• Biotechnology
• Machine Learning

Background:

• Traditional forensic DNA analysis involves two independent human reviewers for interpreting short tandem repeat (STR) capillary electrophoresis profiles.
• Discrepancies between human reviewers necessitate a comparison and resolution process, adding time and resources.

Purpose of the Study:

• To evaluate the accuracy and efficacy of an artificial neural network (ANN) tool for interpreting STR DNA profiles.
• To determine if the ANN can replace one human reviewer in a forensic laboratory setting.

Main Methods:

• An artificial neural network (ANN) was developed to classify fluorescence signals in raw capillary electrophoresis data.
• The ANN was integrated into commercial software (FaSTR™ DNA) for reading GlobalFiler™ DNA profiles.
• Validation studies were conducted at Forensic Science South Australia (FSSA) to assess the ANN's performance.

Main Results:

• The FaSTR™ DNA software, utilizing the ANN, demonstrated a 99.7% accuracy in detecting allele peaks in reference profiles.
• The ANN's high accuracy suggests its suitability for automated DNA profile interpretation.

Conclusions:

• The artificial neural network (ANN) reader in FaSTR™ DNA is a highly accurate tool for interpreting STR capillary electrophoresis profiles.
• The implementation of a one-reader workflow, supported by the ANN, is feasible and could streamline forensic DNA analysis processes at FSSA.