The interpretation of mixed Short Tandem Repeat (STR) profiles has long been challenged by noise interference and peak overlap, particularly when accurately extracting the true genotype of the major contributor from low signal-to-noise ratio (SNR) samples. Traditional statistical analyses and conventional machine learning methods are often limited by their noise-robustness, rendering deconvolution results susceptible to baseline drift and minor peak artifacts. To address these challenges, this paper proposes a noise-robust deconvolution method for the major contributor in mixed STR profiles based on a conditional diffusion model. The proposed method utilizes a forward diffusion process to construct a continuous mapping space where clean single-source profiles degrade into random noise. In the reverse diffusion stage, the extraction of the major contributor's profile is formulated as a process of progressive denoising and signal reconstruction from pure noise, conditioned on global features of the mixed STR profile, such as peak height and area. Furthermore, a conditional guidance mechanism incorporating allele count constraints and peak height ratio priors is introduced to impose physical constraints on the denoising trajectory. This ensures that the reconstructed allele peaks of the major contributor maintain structural integrity and that peak height ratios adhere to biological priors. Experiments conducted on real STR datasets and simulated complex noise data validate the efficacy of the proposed approach. Results indicate that in two-person mixtures, the average deconvolution accuracy for the major contributor reaches 96% with an average response time of 1.4 seconds. In 3-6 person mixtures, the average accuracy is 88% with an average response time of 2.9 seconds. The method demonstrates superior anti-interference capability and stability under low SNR conditions, providing a novel technical pathway based on conditional diffusion models for the reliable deconvolution of low-quality DNA mixed profiles, such as those from forensic casework samples.

Authors

• Shanping Yu (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Zhehua Mao (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)
• Liang Zeng (School of Cyberspace Science and Technology, Beijing Institute of Technology, China)