Single-cell RNA sequencing has changed how scientists study biology by allowing them to measure gene activity in individual cells instead of averaging signals across millions of cells. This level of detail helps researchers better understand development, disease, and the many different cell types that make up tissues. One remaining challenge, however, is accurately reconstructing the full-length RNA transcripts, also called isoforms, that are present in each cell.

Researchers from The Children’s Hospital of Philadelphia and The Pennsylvania State University have developed a new computational tool called Amaranth to improve transcript reconstruction from single-cell RNA sequencing data.

Overview of experiment workflows

Smart-seq3 scRNA-Seq reads are demultiplexed and then aligned to the reference genome using zUMIs (Parekh et al. 2018) and STAR (Dobin et al. 2013). In the single-cell assembly workflow, read alignments are split into individual bam files. Each of these bam files consists of all alignments from a unique single cell. Assembly tools such as Amaranth take each bam file as input and output assembled transcripts for this specific cell in gtf format. All cells are assembled separately. In the meta-assembly workflow, meta-assemblers extrapolate other cells’ alignments to improve the assembly. All cells’ alignments are collected together and piped to downstream meta-assemblers (e.g. Amaranth-meta). At the end of the workflow, all assembled transcripts are evaluated against the reference transcriptome using gffcompare.

Modern single-cell RNA sequencing methods, such as Smart-seq3, can generate reads that span entire RNA molecules. These methods produce two different types of sequencing reads. One type is linked by unique molecular identifiers, or UMIs, which help identify reads originating from the same RNA molecule. The other type consists of internal reads that fill in missing portions of the transcript. Although both read types contain valuable information, they have different characteristics and biases.

Most existing transcript assembly software treats these reads similarly, which can reduce assembly accuracy. Amaranth was designed specifically to recognize and account for the unique properties of each read type.

The software uses specialized algorithms to correctly determine which DNA strand a read originated from, improve the identification of RNA splicing events, and more accurately locate where transcripts begin. These improvements allow Amaranth to reconstruct full-length transcripts with greater precision.

The researchers also developed a companion tool called Amaranth-meta, which combines information from multiple cells to further improve transcript reconstruction for individual cells. By sharing information across similar cells, the software generates more complete and reliable transcript assemblies.

To evaluate the new approach, the team analyzed Smart-seq3 datasets generated from human HEK293T cells and mouse fibroblasts. Amaranth consistently outperformed several leading transcript assembly tools, producing more accurate transcript reconstructions for individual cells as well as during combined analyses across multiple cells.

Improved transcript assembly is important because many genes produce multiple RNA isoforms, each of which can have different biological functions. Better identification of these isoforms allows researchers to study gene regulation at much higher resolution and may reveal previously overlooked differences between healthy and diseased cells.

As single-cell RNA sequencing technologies continue to improve, computational tools such as Amaranth will play an increasingly important role in extracting meaningful biological information from complex sequencing data. More accurate transcript reconstruction will help researchers better understand cellular diversity, gene regulation, and the molecular mechanisms underlying health and disease.

Availability – Amaranth is implemented in C++ and is freely available at https://github.com/Shao-Group/amaranth under the BSD-3-Clause license.