Single-cell RNA sequencing (scRNA-seq) has revolutionized the way scientists study gene expression at the individual cell level. However, designing an scRNA-seq experiment requires careful planning, particularly when balancing the number of cells analyzed and sequencing depth while staying within budget. A new study from the Medical College of Wisconsin introduces FastQDesign, a novel framework that helps researchers optimize their study design using raw FastQ files.

Unlike traditional simulation-based methods that rely on processed Unique Molecular Identifier (UMI) matrices, FastQDesign works directly with raw sequencing reads from publicly available datasets. This approach allows researchers to develop study designs rooted in real-world data rather than theoretical models. By applying FastQDesign to synthetic and real-world datasets, the authors demonstrate how their method can improve cost-effectiveness while preserving data quality.

The framework of FastQDesign

Step 1, we prepare both the FastQ reference dataset and the pseudo-design dataset downsampled from the FastQ reference dataset. The reference dataset can be from a publicly available resource, such as GEO. After the cellranger’s alignment process, the barcode and the alignment detail (BAM file) are transferred to our proposed algorithm fastF to generate the pseudo-design dataset. It first processes the cell barcode one at a time, selecting it only if the current random number n is less than the given subsampling cell number ratio N until finished the entire barcodes list; then processes the BAM file one read at a time, and checks: i) if the current random number r is less than the given subsampling read depth ratio R, ii) if the current read is confidently mapped to the transcriptome (i.e., whether it is noise), iii) if the read belongs to the selected cell barcodes. Note that both n, r ∈ Uniform(0, 1) are simulated from a random number generator (RNG). The filtered reads are then encoded to the SQLite database to generate the UMI matrix from the pseudo-design dataset. Step 2, we compare the stability of the pseudo-design sample from three aspects, cell clustering, marker genes, and pseudotime, by the adjusted rand index(ARI), Jaccard index, and Kendall’s τ index (details are in Methods). We define the similarity as the average of these three indices. We obtain the grid of similarity by varying cell number and read depth, where each dot is the average of 10 repeated measurements. A shape-constrained additive model(SCAM) is fitted to smooth the surface. Step 3, cost-benefit analysis for optimal designs. The colored-coded curves stand for different flow cell capacities. In particular, the purple curve is the budget function, any designs under it are feasible(black), otherwise, it is not attainable(grey). The design with a diamond shape surpasses the similarity threshold(red straight line) and has a minimal cost, which is optimal cost design. The design with a star shape under the budget(blue vertical line), achieves the optimal similarity. The designs are one-to-one correspond in both scatter plots.

This study highlights the importance of proper experimental planning in scRNA-seq research and provides a practical tool for scientists investigating heterogeneous cell populations. With FastQDesign, researchers can make informed decisions that maximize the impact of their sequencing experiments while minimizing unnecessary costs.

Availability – A high-efficiency software suite is available at https://github.com/yuw444/FastQDesign.