Single cell RNA sequencing allows scientists to study gene expression in thousands or even millions of individual cells at once. To make experiments more efficient, researchers often pool samples from multiple donors into a single sequencing run. After sequencing, the data must be separated back into the correct individual sources, a process called demultiplexing.

However, as the number of donors increases, this step becomes much more difficult. Genetic similarities between individuals, overlapping signals, and the presence of doublets, where two cells are captured together, can lead to errors. These challenges can reduce the accuracy of downstream analysis and make it harder to interpret results.

Researchers at Auburn University have developed an improved computational method called Souporcell3 to address these issues. Their goal was to create a more reliable way to separate pooled single cell RNA sequencing data, even when many donors are included.

Souporcell3 builds on earlier approaches by improving how cells are grouped based on genetic variation. It uses a clustering strategy called K Harmonic Means, which is better at avoiding poor solutions that can trap traditional methods. The system also includes an iterative refinement process, where low quality groupings are reworked while high confidence clusters are preserved.

A-F Method overview: A. Randomly initialize with k x 10 clusters to ensure broad coverage of the genotype space. B. Merge nearby cluster centers using a distance metric defined as the sum of allele fraction differences, weighted by allele counts, until k clusters remain. C. Clustering using K-Harmonic Means (KHM) to ensure enhanced convergence and reduced sensitivity to initialization. D. Identify low and high quality clusters based on the number of assigned cells and the loss value per cluster. E. Reinitialize low-quality clusters F. Rerun KHM on the refined set while locking the high-quality clusters. G-I Results: G. ARI values for vireo, vireo with overclustering, souporcell, and souporcell3 for 10% doublet, 64-donor dataset. H. Number of incorrectly merged clusters for the same methods and dataset. I. Cluster maps using UMAPs for the x and y axes, showing the result of 10% doublet, 64-donor dataset after running souporcell3 (left) and vireo with overclustering (right) with poor clustering results highlighted with circles.

One important advantage of this approach is scalability. While previous tools often struggled when analyzing more than about 16 donors, Souporcell3 can handle datasets with up to 64 donors. This makes it especially useful for large studies that aim to capture diverse populations or rare cell types.

The researchers tested the method against existing tools and found that it significantly improved accuracy. In particular, it eliminated cases where clusters from different donors were incorrectly merged, a common problem in complex datasets. The method also performed well across different levels of doublet contamination, maintaining strong agreement with true sample identities.

Accurate demultiplexing is essential for ensuring that RNA sequencing data reflects the correct biological sources. By improving this step, Souporcell3 helps researchers generate cleaner datasets and more reliable insights into gene expression.

As single cell studies continue to scale up, tools like Souporcell3 will play an important role in enabling large, multi donor experiments while maintaining data quality and confidence in the results.

Availability: Souporcell3 is freely available under the MIT open-source license at https://github.com/wheaton5/souporcell.