Forensic single-cell analysis is crucial for resolving complex DNA mixtures, especially when a suspect’s minor DNA contribution is masked by multiple contributors. Capturing single cells prior to DNA analysis enables the physical deconvolution of biological mixtures and allows individual single-cell electropherograms (scEPGs) to be linked to cell types and persons of interest (POIs). However, interpretation of scEPGs remains challenging due to the limited amount of DNA and associated stochastic effects. 

Thus, approaches that cluster partial scEPGs from the same donor based on (dis-)similarities in allele information, and combine them into a consensus profile, are supposed to reveal consistent genetic signals. However, haploid sperm cells contain only one allele per locus. Therefore, scEPGs from the same donor display diverse allele combinations resulting from random meiotic recombination, requiring an adapted clustering approach. Thus, this study aimed to develop an automated and robust clustering approach to group sperm cell-derived scEPGs from multiple donors, enabling more reliable and comprehensive genotyping and contributor number estimation without the need for reference profiles, which are often unavailable in forensic casework.  

To this end, we employed a graph theory-based clustering approach that accounts for the full range of DNA profile compositions originating from a single POI, i.e., from completely identical to entirely different genetic information. To achieve this, each sperm cell is represented as a point (node), and connections (edges) are established between scEPGs that share common genetic features (alleles). Hence, allele-sharing relationships between neighboring nodes are represented by edges, while non-neighboring nodes may be genetically distinct but remain indirectly connected within the network through intermediate nodes. The network structure then enables the linkage of scEPGs originating from a POI, by building bridges between nodes, even when they share only a few or even no alleles. Person-based grouping was performed by applying two different clustering algorithms directly to the resulting graph structure, 
namely Louvain and Leiden. For performance comparison, empirical scEPG data from five unrelated donors, obtained using DEPArray™ technology, were used, and artificial data representing 1,000 donors were generated based on the observed empirical parameters. Graph-based clustering with two to five person mixtures (with 3-48 haploid single cell STR profiles per individual) yielded consensus profiles with approximately 85% profile completeness at an observed dropout rate of 42%, without reliance on any reference profiles. In contrast, individual single cell STR profiles typically contain around nine alleles (~28% profile completeness), highlighting the gain achieved through clustering and consensus building.

Authors

• Janne Berger (IRM Basel, Switzerland)
• Janine Schulte (IRM Basel, Switzerland)
• Amke Caliebe (Health and Medical University Erfurt, Germany)
• Eva Scheurer (IRM Basel, Switzerland)
• Iris Schulz (IRM Basel, Switzerland)