A Study on Hybrid STR Profile Analysis Model Based on Distance-Weighted Self-Attention Mechanism

Forensic DNA analysis is a critical component of judicial practice, where identifying individuals from mixed DNA samples poses significant challenges. Short tandem repeat (STR) analysis serves as the core technique in forensic DNA identification. However, peak height overlap, allele loss, and background noise in mixed samples lead to insufficient identifiability and cumulative errors in traditional probability models during subject identification and proportion estimation. This study aims to construct a deep learning-based cascade analysis framework to systematically enhance the resolution accuracy and robustness of mixed STR profiles. This paper proposes a multimodal fusion algorithm based on the Distance-Weighted Self-Attention (DWSA) mechanism. To address the problem of inferring the number of contributors, a dual-branch network architecture is designed. By integrating sequence features and macro-statistical features, the DWSA mechanism captures spatial correlations among alleles. For contributor proportion estimation, a nonlinear distance-decaying self-attention network is constructed. Jensen-Shannon divergence is employed as the loss function, and multi-locus outputs are integrated in logit space. Experimental results demonstrate that the proposed model achieves 100% accuracy in predicting the number of individuals across 50 test samples. The overall average ensemble MAE for proportion prediction is 0.0809, significantly outperforming traditional methods. This study provides a novel solution for analyzing forensic DNA admixture samples, holding significant theoretical implications and practical application value.