Managing TempO-Seq, RNA-Seq and Drug-Seqs

A Comparative Assessment of Performance, Reproducibility, and Biological Insight

Introduction

Transcriptomic profiling has become a cornerstone of modern biological research, providing insights into cellular state, mechanism of action, toxicity, disease progression, and therapeutic response. Advances in next-generation sequencing have enabled the development of diverse transcriptomic technologies that can broadly be divided into targeted and untargeted approaches, each with distinct advantages and limitations. [1]

Untargeted methods, such as conventional RNA sequencing (RNA-Seq), aim to measure as many transcripts as possible within a sample. By sequencing cDNA fragments derived from RNA molecules, these approaches provide an expansive view of the transcriptome and can identify known and novel transcripts, splice variants, expressed pseudogenes, fusion transcripts, and non-coding RNAs. Because untargeted sequencing does not require prior selection of targets, it is particularly valuable for discovery-based research in which the relevant biology is unknown. [1-2]

More recently, highly multiplexed 3′ counting methods such as Drug-Seq have emerged as cost-effective alternatives to conventional RNA-Seq. Drug-Seq maintains many of the benefits of untargeted transcriptomics while dramatically increasing throughput through early sample barcoding and pooling, allowing hundreds to thousands of samples to be processed simultaneously. This design reduces workflow complexity and sequencing costs, making the technology attractive for large-scale compound screening applications.

In contrast, targeted transcriptomic methods restrict sequencing to predefined genes or transcript regions selected prior to the experiment. TempO-Seq represents one example of this strategy, using pairs of detector oligonucleotides that hybridize directly to specific transcripts. Following hybridization and ligation, only the resulting probe products are amplified and sequenced. Because sequencing effort is focused exclusively on predefined targets, reads are used more efficiently and substantially fewer sequencing reads are required to quantify gene expression across large panels, including whole-transcriptome panels covering nearly all protein-coding genes. TempO-Seq improves measurement efficiency by attenuating highly abundant transcripts, freeing sequencing capacity for lower-abundance genes without increasing read depth. Its direct hybridisation strategy is independent of poly(A)-based capture which is common for many untargeted transcriptomics platforms, making it less reliant on intact transcript ends and therefore suitable for degraded or challenging samples. Because sequencing is restricted to predefined probe products rather than native transcript sequences, it also avoids generating genetic sequence information such as polymorphisms or unexpected variants.

The primary limitation of targeted transcriptomics is that only transcripts represented within the panel can be measured. While modern whole-transcriptome targeted panels encompass the vast majority of protein-coding genes, transcripts that are not explicitly targeted, including novel transcripts, many non-coding RNAs, cryptic splice variants, fusion transcripts, and expressed pseudogenes, may not be detected. Consequently, untargeted approaches remain the preferred choice for transcript discovery and comprehensive transcriptome characterization.

As transcriptomics expands into screening, toxicology, translational research, and precision medicine, it is important to understand how targeted and untargeted strategies affect data quality and biological interpretation. Here, we compare TempO-Seq, Drug-Seq, and RNA-Seq to assess differences in signal distribution, sensitivity, reproducibility, and recovery of meaningful biological information.

Datasets and Comparative Framework

An ideal comparison of transcriptomic platforms would involve applying all technologies to the same biological samples generated from the same experimental design and processed under identical conditions. To date, however, no publicly available study has generated large-scale dose-response transcriptomic datasets using TempO-Seq, Drug-Seq, and RNA-Seq on the same samples. Consequently, a direct one-to-one comparison was not possible.

To minimise the impact of biological and experimental variability, datasets were selected according to three guiding principles. First, priority was given to studies containing a large number of samples, enabling robust statistical evaluation of transcriptomic performance. Second, preference was given to perturbation-based experiments involving multiple compounds and dose levels, which provide a broad range of biological responses and facilitate assessment of differential expression. Third, datasets were selected to represent typical use cases of each technology while avoiding comparisons based solely on a single cell type or treatment condition.

For Drug-Seq, analyses were performed using the largest publicly available datasets, including the Novartis U2OS compound screening study (mapped read counts were not publicly available and were therefore empirically estimated by dividing mapped UMI counts by the average UMI rate observed in other Drug-Seq experiments) and the GPDx4 HEK293 perturbation dataset. These studies provide extensive collections of chemically perturbed samples generated using the Drug-Seq workflow and represent some of the most comprehensive resources currently available for evaluating this technology.

For TempO-Seq, four independent datasets were analysed. These included two internal, unpublished dose-response studies performed in MCF7 and MDA-MB-231 cell lines, the publicly available GSE152128 dataset comprising rat primary hepatocytes exposed to 14 chemicals and measured using both the TempO-Seq Whole Transcriptome and S1500 panels (approx. 3000 targeted genes), and the recently released GSE274318 dataset, a large-scale U2OS dose-response study containing more than 30,000 samples. Together, these datasets span multiple species, cell types, panel formats, and experimental scales.

For RNA-Seq, analyses were conducted using GSE179347, which contains multiple human skeletal muscle-, liver-, and adipose-relevant cell models exposed to 21 perturbations, and healthy liver samples from the GTEx consortium. These datasets were selected to provide both perturbation-based and baseline expression profiles generated using conventional RNA sequencing.

Because the datasets originate from different biological systems and experimental designs, this study does not seek to compare absolute gene expression values across platforms. Instead, the analysis focuses on performance characteristics that are largely independent of cell type and treatment context. Specifically, comparisons were performed using four key metrics:

• Gene detection rate: the number of genes detected per sample and the relationship between sequencing depth and transcript detection.
• Read distribution efficiency: the distribution of sequencing counts across genes, including the proportion of reads allocated to highly abundant transcripts versus lower abundance transcripts.
• Replicate consistency: the reproducibility of gene expression measurements across biological and technical replicates.
• Differential expression sensitivity: the ability of each platform to detect statistically significant transcriptional responses following perturbation.

These metrics were selected because they directly reflect the efficiency with which sequencing resources are converted into biologically informative measurements, irrespective of the specific cell type, species, or treatment condition under investigation.

Detection Is Not Expression: Measuring the Usable Transcriptome

To evaluate transcript detection performance, we first generated sequencing saturation curves showing the number of detected genes as a function of sequencing depth for each dataset (Figure 1). Because RNA-Seq datasets were generated at substantially greater sequencing depths than the TempO-Seq and Drug-Seq datasets, the results are displayed on separate scales to facilitate visual comparison.

As expected, untargeted approaches consistently detected the highest number of transcripts. RNA-Seq achieved the broadest transcriptome coverage, with Drug-Seq ranking second, probably as a consequence of its much lower overall sequencing depth. These observations are consistent with the design principles of untargeted sequencing and confirm its superior capacity for transcript discovery.

However, transcript detection alone provides an incomplete measure of transcriptomic performance. A gene represented by a single sequencing read contributes little practical value to downstream analyses and cannot be considered equivalent to a gene measured with sufficient statistical confidence. In differential expression workflows, including DESeq2, genes with very low counts are routinely filtered prior to analysis because they provide limited statistical power and contribute disproportionately to noise. Consequently, the biologically relevant question is not simply how many genes are detected, but how many genes are detected at levels sufficient to support meaningful downstream interpretation. This distinction is particularly important for technologies such as Drug-Seq and TempO-Seq, which claim to capture the full transcriptional signal using only a few million reads.

Figure 1: Saturation curves showing the number of genes detected at a minimum threshold of one read. Data were binned in 100,000-read increments for TempO-Seq and Drug-Seq and 1-million-read increments for RNA-Seq. The temposeq_152128_st dataset used a surrogate panel targeting approximately 3,000 genes, which accounts for its lower number of detected genes.

To examine this distinction, we repeated the analysis using a stricter detection threshold of at least 10 counts per gene and restricted it to datasets containing samples with sequencing depths of 1–3 million reads. This threshold approximates the lower limit commonly used in transcriptomic pipelines to identify genes that can be reliably quantified and may support downstream statistical analysis. Table 1 shows a clear change in gene detection when the minimum threshold is raised from 1 to 10 reads, at which TempO-Seq generally detects considerably more genes. These observations suggest that while untargeted approaches maximise transcript discovery, targeted sequencing can utilise sequencing resources more efficiently by concentrating reads on biologically relevant transcripts rather than distributing sequencing capacity across a much larger collection of low-abundance features.

The results therefore highlight an important distinction between transcript discovery and usable transcriptome coverage. Although untargeted approaches excel at identifying the largest possible number of RNA species, a substantial fraction of these transcripts may be represented at levels insufficient for robust downstream analysis. By focusing sequencing effort on predefined targets, TempO-Seq generates a larger number of genes with quantifiable expression levels while requiring considerably fewer reads per sample.

Table 1: Number of genes detected at minimum read thresholds of 1 and 10 across three sequencing depths in two Drug-Seq and three TempO-Seq datasets.

The Economics of Sequencing Reads

In transcriptomics experiments, sequencing depth is a finite resource. When a large proportion of reads is consumed by a small number of highly abundant transcripts, fewer reads remain available to quantify low- and medium-abundance genes. As a result, many transcripts may be formally detected but represented by only a handful of reads, limiting their usefulness for differential expression and downstream interpretation.

To assess how efficiently each platform distributes sequencing signal across the transcriptome, we first calculated the NSig80 metric. Originally introduced by the US EPA, NSig80 is defined as the number of genes required to account for 80% of all reads in a sample. Higher NSig80 values indicate that sequencing reads are distributed across a larger number of genes, whereas lower values indicate that much of the sequencing space is dominated by a relatively small group of highly expressed transcripts.

As shown in Figure 2, TempO-Seq datasets often exhibited higher NSig80 values than the other platforms. This finding indicates a more balanced allocation of sequencing reads across the transcriptome and suggests that a larger proportion of genes contribute meaningfully to the overall sequencing signal. Notably, NSig80 for the Drug-Seq datasets was calculated from published UMI-based count tables, which generally yield higher values for this metric. The corresponding read-based NSig80 values are therefore likely lower than those shown in Figure 2.

Figure 2: NSig80 distribution across the analyzed datasets

One potential limitation of NSig80 is that the metric may be influenced by the biological characteristics of a given cell type. Certain cells naturally express a small number of transcripts at very high levels and therefore may be expected to display lower NSig80 values regardless of the analytical platform employed. However, a particularly informative comparison can be made between the Drug-Seq Novartis dataset and the TempO-Seq GSE274318 dataset. Both studies were performed using the U2OS cell line, thereby reducing the influence of cell-type-specific transcriptional architecture. Despite this similarity, the TempO-Seq dataset exhibited NSig80 values that were more than twice those observed in the Drug-Seq dataset, providing strong evidence that the observed differences arise from platform characteristics rather than cellular biology alone.

We next examined gene-count distributions in samples with approximately three million mapped reads (Figure 3). Drug-Seq showed a strong accumulation of genes with low counts, meaning many detected transcripts had limited value for reliable downstream statistical analysis. By contrast, TempO-Seq showed a rightward shift toward moderate and high counts, indicating that sequencing reads were distributed across more genes at levels useful for differential expression, pathway enrichment, and related analyses.

Figure 3: Read distribution across genes in samples with sequencing depths close to 3 million reads. Orange indicates TempO-Seq and blue indicates Drug-Seq. RNA-seq was excluded because it requires substantially greater sequencing depth.

Taken together, the NSig80 analysis and count-distribution profiles demonstrate that sequencing efficiency is not solely determined by the total number of reads generated or the total number of genes detected. Rather, the manner in which sequencing space is allocated across transcripts strongly influences the amount of usable biological information obtained from each sample. The more balanced read distribution observed in TempO-Seq enables a greater proportion of the transcriptome to be quantified with meaningful depth, providing a plausible explanation for its strong performance in downstream analyses despite requiring substantially fewer sequencing reads.

Consistency between replicates

A fundamental measure of transcriptomic assay performance is the degree of agreement observed between biological replicates. High replicate-to-replicate correlation indicates that the technology is able to capture biological signals consistently while minimizing technical variation introduced during sample processing, library preparation, sequencing, and data analysis.

To assess reproducibility, we calculated pairwise Spearman correlations between biological replicates for each dataset and platform (Figure 4). Across all evaluated studies, TempO-Seq consistently achieved the highest correlation coefficients, with median values typically exceeding 0.95 and frequently approaching 0.98. In contrast, DRUG-seq datasets showed lower correlations, generally ranging between approximately 0.83 and 0.87, while the RNA-seq dataset GSE179347 exhibited substantially greater dispersion and lower overall concordance between replicates.

The narrow distribution observed across TempO-Seq datasets is particularly noteworthy. Not only are median correlations higher, but the interquartile ranges are markedly smaller, demonstrating that replicate consistency is maintained across multiple independent studies, tissue types, and experimental conditions. This suggests that the observed reproducibility is an intrinsic feature of the technology rather than a dataset-specific effect.

Figure 4: Spearman correlation distributions across the analysed datasets. The GTEx liver RNA-seq dataset was excluded because it does not include replicate samples.

The high level of concordance observed between TempO-Seq biological replicates is further supported by the coefficient of variation (CV) analysis (Figure 5). As expected, datasets exhibiting stronger replicate-to-replicate correlation generally also display lower within-group variability. Across the studies evaluated here, TempO-Seq consistently achieved the lowest median gene-level CV values, while DRUG-seq and RNA-seq showed substantially greater variability.

Most TempO-Seq datasets exhibited median CV values between approximately 0.24 and 0.40, compared with approximately 0.55-0.65 for DRUG-seq. The RNA-seq dataset GSE179347 displayed the highest overall variability, with median CV values approaching 0.85. These results independently confirm the conclusion drawn from the correlation analysis: TempO-Seq generates highly consistent gene expression measurements across biological replicates.

Figure 5: Coefficient of variation distributions across the analysed datasets. The GTEx liver RNA-seq dataset was excluded because it does not include replicate samples.

Several features of the TempO-Seq workflow likely support this reproducibility. Unlike conventional RNA-seq, it avoids reverse transcription, second-strand synthesis, fragmentation, end repair, adaptor ligation, and extensive pre-capture amplification, steps that can introduce stochastic variability and amplification bias. Direct hybridisation and ligation of predefined detector oligonucleotides therefore reduces opportunities for technical variation to accumulate. Targeted sequencing may further improve precision by focusing reads on informative genes rather than spreading them across the whole transcriptome. This increases counting precision at a given depth, reduces sampling noise, and likely explains the stronger replicate concordance observed for TempO-Seq compared with DRUG-seq and RNA-seq.

Sensitivity to Differential Gene Expression

The ultimate value of a transcriptomic technology lies in its ability to detect biologically meaningful changes in gene expression. While sequencing depth contributes to sensitivity, the distribution of sequencing reads across genes is equally important. When a large fraction of genes is supported by only a small number of reads, statistical power is reduced because low-count measurements are inherently more susceptible to sampling noise. Consequently, differential expression analysis becomes biased towards genes that are already highly expressed.

To reduce the influence of treatment-specific effects, we performed repeated randomised comparisons within each dataset, comparing randomly selected treated samples with their matched solvent controls. While the magnitude of transcriptional responses can vary substantially between compounds, doses, and experimental conditions, the use of large datasets and many randomised comparisons should help minimise the influence of treatment-specific effects and reveal broader technology-dependent trends. The DRUG-seq datasets consistently identified the fewest DEGs, with averages of approximately 56 genes (DrugMatrix) and 15 genes (Novartis) per comparison. In contrast, the TempO-Seq datasets identified substantially larger numbers of transcriptional changes, ranging from approximately 189 to 1,699 genes depending on the study. The RNA-seq dataset (GSE179347) identified approximately 904 DEGs per comparison, likely benefiting from its substantially greater sequencing depth (Table 2).

dataset	Average number of differentially expressed genes	Average expression (CPM) of differentially expressed genes
Drug-Seq_gdpx4	56.2	923.56
Drug-Seq_novartis	14.9	711.23
temposeq_wt_mcf7	379.8	180.95
temposeq_wt_mda	527.2	163.33
temposeq_gse152128_st	276.7	908.49
temposeq_gse152128_wt	1698.6	143
temposeq_gse274318_wt	188.8	98.65
rnaseq_gse179347	903.5	--

Table 2: Average number of differentially expressed genes identified across 100 random comparisons and their mean expression in CPM across the analysed datasets. RNA-seq CPM values were excluded because CPM does not account for gene-length-dependent differences, which require alternative normalization methods.

A particularly informative comparison is provided by the Novartis DRUG-seq study and the TempO-Seq GSE274318 study. Both represent dose-response experiments performed in the same cell line, making them more comparable than datasets generated in different biological systems. Despite inevitable differences in compounds and treatment conditions, TempO-Seq identified approximately 189 DEGs per comparison, compared with only 15 DEGs for DRUG-seq. Although treatment-specific effects cannot be completely excluded, the magnitude of this difference is consistent with the hypothesis that read distribution influences the sensitivity of differential expression detection.

This trend was confirmed by direct comparison of the same compound and concentration across the two U2OS datasets: 1 µM retinoic acid induced 65 differentially expressed genes in TempO-Seq, but none in DRUG-seq. Notably, the TempO-Seq response included several genes expected to be perturbed by retinoic acid, including RARB, a commonly used readout of retinoic acid signalling, as well as CYP26A1, STRA6, DHRS3, HOXA2, HOXA5, and NRIP1. [3]

To further test this hypothesis, we calculated the average CPM of genes identified as differentially expressed. If a platform primarily detects changes among highly expressed genes, then the significant genes themselves should exhibit relatively high expression levels. This is precisely what was observed. Differentially expressed genes identified by DRUG-seq showed average CPM values between approximately 711 and 924, whereas most TempO-Seq datasets ranged between approximately 99 and 181 CPM. These results suggest that DRUG-seq preferentially detects expression changes among highly abundant transcripts, while TempO-Seq remains sensitive across a broader range of expression levels.

CPM values from RNA-seq were not directly compared because CPM normalisation does not account for transcript length. Since RNA-seq quantifies full-length transcripts whereas TempO-Seq and DRUG-seq interrogate predefined transcript regions, direct comparison of CPM values across these platforms would be misleading.

The ability to detect differential expression in lowly expressed genes is particularly relevant because biological importance is often poorly correlated with expression level. Many key regulatory genes, including transcription factors, nuclear receptors, signalling molecules, and chromatin regulators, are expressed at relatively low abundance under basal conditions. Despite their low expression, these genes frequently function as master regulators that control extensive downstream transcriptional programmes. Examples include toxicologically relevant regulators such as AHR, NRF2 (NFE2L2), HIF1A, PPARs, STAT family members, and numerous ligand-activated transcription factors. Early perturbations in these regulators may represent the first molecular events following chemical exposure and often provide critical mechanistic insight into pathway activation and adverse outcome pathways.

Conclusion

The analyses presented in this paper demonstrate that differences in transcriptomic performance are driven not only by sequencing depth, but also by how efficiently reads are distributed across genes. Across multiple independent datasets, TempO-Seq consistently exhibited a more balanced allocation of sequencing reads, higher replicate concordance, and lower within-group variability than the other non-targeted transcriptomics platforms. These advantages improved sensitivity in differential expression analysis, increasing the biological insight that scientists can extract from expression data.

Non-targeted transcriptomics retains value for discovery-focused questions, including novel transcripts, unexpected biology, and uncharacterised responses. For the broader and more routine applications of transcriptomics, including screening, toxicology, translational research, and precision medicine, targeted approaches such as TempO-Seq may provide a more practical balance of precision, scalability, reproducibility, and biological sensitivity.

References

[1] Word LJ, Willis CM, Judson RS, Everett LJ, Davidson-Fritz SE, et al. (2025. PLOS ONE 20(5) e0320862. 

[2] Bushel PR, Paules RS and Auerbach SS (2018). Front. Genet. 9:485.

[3] Lavudi et al. Front Cell Dev Biol. 2023 Aug 11;11:1254612