Current techniques for deconvoluting mixed Short Tandem Repeat (STR) profiles are significantly constrained by analytical efficiency bottlenecks. Traditional probabilistic genotyping methods require an exhaustive search of all possible genotype combinations, resulting in an exponential increase in computational complexity as the number of contributors grows, which renders them inadequate for complex mixture scenarios. Although Transformer-based methods possess strong long-sequence modeling capabilities, the quadratic computational overhead of their self-attention mechanisms leads to heavy memory consumption and low inference efficiency when processing extended profile sequences. To address these issues, Selective State Space Models (SSMs), represented by the Mamba architecture, demonstrate significant advantages. By breaking the limitation of fixed parameters in traditional models, this architecture introduces an input-aware selection mechanism that enables state transition parameters to adapt dynamically to the input data. Simultaneously, by efficiently compressing global sequence information into hidden state variables, it successfully reduces the computational complexity of long-sequence modeling to a linear scale. Leveraging these advantages, this paper proposes an efficient mixed STR profile deconvolution network based on the Selective SSM. Taking the encoder-extracted latent vector sequence of the mixed STR profile as input, the network divides the long sequence into overlapping chunks and decouples multi-scale features through a dual-path architecture: the intra-chunk path focuses on local features to capture short-range dependencies between adjacent loci, while the inter-chunk path performs global modeling across all chunks to capture long-range dependencies. Both paths embed Bidirectional Mamba modules, integrating the SSM's selective mechanism with specific profile characteristics. By dynamically optimizing model parameters based on biological features such as allele position, peak height, and peak area, the network precisely focuses on key regions and realizes global feature interaction. Finally, the fused feature representations are reconstructed by a decoder into single-contributor STR profiles, accomplishing end-to-end individual genotype resolution. Experimental verification shows that the method proposed in this paper achieves an average split accuracy of 95% for the major contributor in two-person mixed samples, with an average response time of 1.18 seconds; in 3-6 person mixed samples, the average split accuracy for the major contributor is 85%, with an average response time of 2.26 seconds. Compared to existing probabilistic genotyping and Transformer-based methods, the proposed approach significantly reduces computational resource consumption and inference latency while maintaining high deconvolution accuracy, providing a novel solution that effectively balances precision and efficiency for mixed STR profile deconvolution.

Authors

• Xinyu Yang (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Liang Zeng (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Zhen Xu (Institute of Forensic Science, Ministry of Public Security, China)
• Fan Yang (Institute of Forensic Science, Ministry of Public Security, China)