Intelligent deconvolution algorithm for mixed STR profiles based on locus association modeling.

Area of Science:

• Forensic Science
• Genetics
• Computational Biology

Background:

• Short tandem repeat (STR) typing is crucial for forensic identification and mass disaster victim identification.
• Current probabilistic genotyping models for mixed DNA samples assume locus independence, potentially limiting accuracy.
• Advanced forensic DNA analysis requires methods that capture complex inter-locus correlations.

Purpose of the Study:

• To develop and validate a deep learning-based approach for forensic STR mixture deconvolution that utilizes inter-locus dependencies.
• To improve the accuracy of identifying individuals from complex DNA mixtures compared to existing methods.
• To assess the model's performance, including its generalizability across different sequencing platforms.

Main Methods:

• A deep learning model was trained on single-contributor STR profiles to learn inter-locus dependencies.
• These learned dependencies were integrated with a fully continuous probabilistic genotyping model.
• The refined model was validated using the PROVEDIt dataset and a real-world case study.

Main Results:

• The deep learning approach achieved significant accuracy improvements in deconvoluting 2-, 3-, and 4-person DNA mixtures, outperforming conventional models by up to 30 percentage points.
• Validation on a real case demonstrated closer agreement with manual identification than the fully continuous model.
• The model showed platform-specific performance, with multi-platform training enhancing generalization.

Conclusions:

• The proposed deep learning model offers a robust and accurate solution for interpreting forensic STR mixtures by leveraging inter-locus correlations.
• Accounting for locus dependencies is critical for advancing DNA mixture analysis in forensic science.
• Further development should focus on improving cross-platform generalizability for wider forensic application.