A rigorous systematic benchmarking study of short-read single-cell RNA sequencing pipelines has demonstrated that while older transposable element insertions can be quantified with high fidelity across individual genomic loci, young repetitive elements remain intensely prone to false-positive detections and mapping errors. Published in Genome Biology by researchers Veronica Finazzi, Catalina A. Vallejos, and Antonio Scialdone, the study introduces an open-source, reproducible Snakemake workflow that pairs empirical data from human and mouse cells with synthetic single-cell datasets containing read-level ground truth. Across multiple computational packages including SoloTE, Stellarscope, and STARsolo, attempting to assign short reads to specific genomic insertions revealed that multi-mapping ambiguities fundamentally distort locus-level estimates for evolutionarily recent transposons. While transposable elements contribute substantially to cellular identity and gene regulation, the authors found that current software tools incorporating multi-mapping reads often inflate false detections without yielding meaningful gains in locus-level biological accuracy.
Although the vast majority of historical insertions have lost their mobilization capacity through evolutionary mutation, the residual sequences frequently serve as cis-regulatory elements or transcribe regulatory RNAs. Their activation is known to be tightly orchestrated across cellular differentiation; for instance, sharp pulses of transposable element transcription mark the acquisition of totipotency during mammalian preimplantation development. Despite this importance, single-cell transcriptomic atlases routinely omit transposable elements from standard workflows due to severe mapping ambiguities. Because transposable elements are repetitive and interspersed across chromosomes, sequencing reads frequently map to dozens or hundreds of genomic positions simultaneously. Furthermore, transcribed insertions frequently overlap host genes, creating chimeric transcripts that confound whether transcription initiated at a transposon or inside an annotated gene body.
To evaluate how existing computational tools cope with these challenges in droplet-based single-cell assays, the researchers built an automated benchmarking framework centered on short-read platforms like 10x Genomics. They focused on three locus-level quantification pipelines: SoloTE, which discards multi-mapped reads by default to focus exclusively on uniquely mapped reads; Stellarscope, an adaptation of the Telescope framework that employs an expectation-maximization algorithm to iteratively reassign multi-mapping reads based on candidate locus abundance; and STARsolo, an aligner and quantification tool that incorporates maximum-likelihood expectation-maximization directly during read mapping to output fractional counts. The team initially surveyed four real single-cell datasets, encompassing mouse embryonic stem cells transitioning into 2-cell-like cells, human in vitro 8-cell-like cells, and healthy mouse and human peripheral blood mononuclear cells. Using SoloTE, they uncovered that transposable elements account for a substantial fraction of all sequenced reads, reaching nearly fifty percent of total reads in the human samples.
Crucially, analyzing the empirical datasets revealed that transcriptomic profiles constructed solely from transposable element expression could accurately reconstruct cellular heterogeneity and separate distinct immune and developmental cell types. In human peripheral blood mononuclear cells, clustering based purely on transposable element counts closely mirrored standard gene-expression clusters, with minor discrepancies restricted to closely related immune lineages like CD14+ versus FCGR3A+ monocytes. In both embryonic stem cells and peripheral blood, transposable element counts even distinguished subtle cell substructures that were not apparent in conventional gene-expression analyses. However, when comparing the specific loci identified as active across tools, agreement was surprisingly poor. SoloTE identified the largest overall repertoire of expressed transposable elements, whereas STARsolo diverged sharply from the other methods. Furthermore, nearly half of the loci uniquely detected by Stellarscope overlapped annotated gene exons, highlighting that differing strategies for handling gene bodies substantially alter which transposons appear active.
Because empirical data lack an absolute ground truth of which loci are truly transcribing, the team engineered a comprehensive simulation framework using Splatter to generate baseline count matrices and Minnow to simulate realistic 100-base-pair droplet-based reads. Minnow modeled droplet-specific polymerase chain reaction amplification biases, sequencing errors, fragmentation patterns, and mapping ambiguities trained directly on mouse embryonic stem cell parameters. The authors created multiple synthetic benchmarks of controlled complexity: matrices containing purely old transposable elements, matrices containing purely young elements, a mixed scenario combining both evolutionary ages, and a combined matrix joining transposons with two thousand highly variable genes. Transposon loci were categorized by evolutionary age based on sequence divergence from family consensus using the Jukes-Cantor model, designating elements younger than two million years as young and those above that boundary as old.
The simulated ground truth exposed stark performance contrasts driven entirely by evolutionary age and sequence homogeneity. In simulations containing only older transposable elements, over eighty percent of generated reads mapped uniquely to a single genomic location, and every tested method exhibited near-perfect precision and recall, yielding Spearman rank correlation coefficients approaching 1.0 against true expression values. In sharp contrast, in simulations restricted to young elements, over eighty percent of sequencing reads mapped to multiple genomic locations. All computational tools suffered from massive over-detection of young loci, reporting thousands of false-positive features and showing limited mutual agreement. SoloTE in unique-mapper mode achieved the highest detection precision by strictly discarding multi-mapping reads, but this selectivity came at the expense of recall. Incorporating multi-mapped reads through expectation-maximization algorithms in Stellarscope and STARsolo recovered more expressed candidates but dramatically amplified false-positive calls, failing to elevate overall F1 scores above modest baselines.
Detailed read-level tracking demonstrated the precise computational mechanisms behind these errors. In young transposable element simulations, 4,257 false-positive loci were reported by all tools simultaneously, whereas 194 genuine loci were missed across the board. The systematically overlooked elements were heavily enriched for L1 retrotransposons, which exhibited low sequence divergence and suffered from severe read dropouts. Conversely, the shared false-positive calls were significantly enriched for long terminal repeat retrotransposons, specifically ERVL, ERVK, and ERV1 families, displaying longer sequence lengths and higher divergence. This revealed a pervasive directional bias: short reads originating from young, highly repetitive insertions frequently misaligned to older, more divergent paralogous copies across the genome. Even tools restricted strictly to uniquely mapping reads suffered from this cross-assignment, showing that relying exclusively on unique mappers does not completely inoculate an analysis against multi-mapping artifacts when young repeats are present.
To isolate the biological and structural determinants that govern quantification accuracy, the authors fitted generalized estimating equations models to true-positive detections. Across computational tools, transposon sequence length and divergence from consensus emerged as the strongest determinants of accurate quantification. Longer transposable elements were significantly less likely to achieve above-median Spearman correlation with ground truth, reflecting a regression coefficient of negative 0.823. Conversely, higher sequence divergence was strongly associated with superior correlation accuracy, displaying a positive coefficient of 0.372. Structural completeness also exerted a measurable effect: transposable elements with 3-prime truncations were significantly more likely to be accurately quantified than full-length insertions, showing an odds ratio increase with a positive coefficient of 0.642. While global expression levels did not uniformly dictate accuracy, an exploratory interaction model revealed that higher average expression significantly boosted Stellarscope performance, reflecting its pseudobulk expectation-maximization approach where abundant transcripts gather read weight more reliably.
Assessing performance across specific transposon families further underscored that computational efficacy is not uniform across classes. Short interspersed nuclear elements, including Alu and B2 families, demonstrated the highest quantification precision and F1 scores across methods, facilitated by their shorter sequence lengths and higher localized divergence within read-covered segments. In stark contrast, longer elements from the L1 and ERVL families showed pervasive mapping failure and the lowest correlation with ground truth. However, when the authors computationally aggregated individual locus counts into subfamily-level bins, precision and recall surged dramatically across all evaluated tools. Aggregation largely eliminated false-positive noise, demonstrating that while short reads frequently cannot resolve which exact paralogous locus generated a transcript, they reliably capture the general active subfamily. Notably, SoloTE counts aggregated to the subfamily level proved considerably more accurate than the native subfamily output of scTE, which assigns multi-mappers solely based on best-scoring alignments without rigorous downstream filtering.
The benchmark also illuminated pervasive cross-talk between host genes and transposable element annotations. When simulating mixed environments containing both genes and transposons, all tools generated false positives where gene-derived reads were misallocated to overlapping transposon exons, or transposon reads were erroneously credited to annotated genes. Because simulated reads measured 100 base pairs, fragments falling entirely within shared exon overlaps greater than 100 base pairs were mathematically impossible to disambiguate from sequence alignment alone. Stellarscope proved especially vulnerable to this cross-assignment, generating more than 400 false-positive transposable element calls directly corresponding to actively expressed genes, because its pipeline does not filter out reads intersecting gene bodies. Conversely, SoloTE explicitly filters out gene-mapped reads prior to transposon intersection, minimizing false-positive transposable element calls but risking the omission of genuine autonomous transcription occurring within complex gene loci.
Computational scalability and software resource footprints exhibited dramatic disparities across pipelines during testing on virtual cloud instances. Stellarscope proved exceptionally resource-heavy and computationally slow, requiring migration to a 512-gigabyte RAM virtual machine to complete the full-depth young transposable element simulation, which ran for nearly six days. Its default expectation-maximization convergence threshold of 10 to the negative seventh power was never reached across any real or simulated benchmark, requiring users to relax convergence parameters to avoid endless iterations. In contrast, SoloTE completed the exact same dataset in under one hour, and STARsolo finished in under ninety minutes on a standard 256-gigabyte virtual machine. These runtime profiles reveal critical practical considerations for researchers planning large-scale single-cell atlas analyses, where tool efficiency directly governs operational feasibility.
Based on their unified benchmark, the investigators formulated clear operational best practices for researchers profiling transposable elements in single-cell transcriptomics. First, locus-specific analyses should be restricted primarily to evolutionarily older insertions, while single-locus calls for young elements should be treated with extreme skepticism. Second, unique-mapper strategies like default SoloTE represent the most pragmatic starting point for locus-level profiling, maximizing precision and minimizing gene-overlap artifacts. Third, whenever young elements or atlas-scale discovery are the primary focus, researchers should aggregate locus counts into subfamilies to secure robust, biologically meaningful expression estimates. Finally, single-cell workflows must integrate explicit quality-control audits for gene-transposon overlaps to prevent host gene expression from masquerading as mobile element activity, while the wider field anticipates hybrid long-read protocols to definitively span repetitive loci.
Subject of Research: Benchmarking computational tools for locus-specific transposable element quantification in single-cell RNA sequencing
Article Title: Benchmarking computational tools for locus-specific analysis of transposable elements in single-cell RNA-seq datasets
Article References: Finazzi, V., Vallejos, C. A., & Scialdone, A. (2026). Benchmarking computational tools for locus-specific analysis of transposable elements in single-cell RNA-seq datasets. Genome Biology. https://doi.org/10.1186/s13059-026-04349-8
Image Credits: AI Generated
DOI: 10.1186/s13059-026-04349-8
Keywords: transposable elements, single-cell RNA-seq, benchmarking, computational biology, SoloTE, Stellarscope, STARsolo, multi-mapping reads, Snakemake, retrotransposons, repetitive elements, gene expression regulation, computational genomics, bioinformatics, Splatter, Minnow, pseudobulk
News Source: Juliet Wilcox. (November 19, 2026). Benchmark Shows Short-Read Single-Cell RNA Sequencing Stumbles on Young Transposable Elements. Scienmag.



