A TSS analysis pipeline based on 5’ single-cell RNA-seq data

To facilitate the application and comparison of TSS prediction and quantification methods on 5’ scRNA-seq data, we propose a pipeline below which has been implemented in the scTSS software. The pipeline has two major steps. The first step is TSS prediction and quantification; the second step is differential TSS usage (DU) analysis. Our pipeline considers a general scenario where multiple single-cell samples (e.g., each sample is from a different donor or biological condition) are available, and a joint analysis is to be performed.

For the first step, predictions can be performed for both individual TSSs and clusters of TSSs. In this work, we focus on the prediction of TSS clusters to facilitate multi-sample analysis and analyzing patterns of TSS distribution. Another critical consideration in TSS prediction is the type of sequencing data used. We classify 5’ scRNA-seq data into two categories: paired-end and single-end (Fig 1A). Paired-end data include cDNA information in both Read 1 and Read 2, with Read 1 being particularly useful for accurately identifying genuine TSSs. In contrast, in single-end data, only Read 2 contains cDNA information, which does not directly indicate the exact position of genuine TSSs. Read 1 only includes other information such as the cell barcode. Nonetheless, it’s feasible to infer the position of genuine TSSs based on Read 2’s position and the length of cDNA fragments. In the remainder of the manuscript, we designate the paired-end data as “on-site data” and the single-end data as “near-site data”, in order to emphasize their difference in the context of TSS prediction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. An overview of the scTSS toolkit. (A) A schematic comparison between the paired-end data (on-site data; left) and single-end data (near-site data; right). In the on-site data, Read 1 directly captures the genuine TSS. In the near-site data, Read 1 does not directly capture the genuine TSS, but there exists a predictable relationship between Read 2 and the genuine TSS, facilitating its estimation. (B) The workflow of the scTSS toolkit. scTSS consists of two major steps, TSS cluster prediction and quantification, followed by differential TSS usage (DU) analysis. In the prediction and quantification step, the toolkit takes in either predicted TSS clusters based on on-site reads, or predicted near-site clusters based on near-site reads, and eventually outputs cell-level count matrices across samples with unified TSS clusters. In the DU analysis step, we test for changes of TSS usage between biological conditions (e.g., cell types). (C) A toy example of the DU test given TSS counts from two cell types across five donors. We suppose TSS cluster 1 displays differential usage between the two cell types, with higher usage observed in cell type A, whereas the usage of TSS cluster 2 does not present significant difference between cell types. scTSS computes a P-value for each TSS cluster, indicating the significance of its differential usage between the two cell types.

https://doi.org/10.1371/journal.pcbi.1012878.g001

Beginning with multiple single-cell samples profiled as 5’ on-site data, the pipeline initiates with TSS prediction and clustering on individual samples using a bioinformatics tool compatible with on-site sequencing data. In the subsequent section, we compare four candidate tools. Once a set of TSS clusters is generated from each sample, our scTSS tool performs TSS cluster merging to consolidate a unified set of TSS clusters (Fig 1B; see Methods). This ensures that the TSS clusters are comparable across samples. Subsequently, read counts of the TSS clusters are summarized for each individual cell. After TSS cluster quantification, scTSS produces a count matrix of unified TSS clusters per cell for each sample.

In scenarios involving multiple single-cell samples profiled as 5’ near-site data, conventional bioinformatics tools tailored for on-site data are not directly applicable. In such cases, our pipeline initially employs a conventional tool to detect clusters of the 5’ ends of Read 2, termed as the near-site clusters (Fig 1B; see Methods). Subsequently, scTSS performs an adjustment to infer the corresponding TSS clusters from these near-site clusters. The subsequent steps in the pipeline remain consistent with those applied to the on-site data.

For the second step, the objective of DU analysis is to detect differential usage of TSS clusters associated with biological conditions, such as variations between different cell types or disparities between healthy and diseased donors (Fig 1B). TSS cluster usage is defined as the percentage of RNA transcripts originating from a gene that have a TSS within a specific cluster. To conduct DU analysis, scTSS utilizes a generalized Binomial model to assess the association between TSS cluster usage and biological conditions, while controlling for potential gene expression changes (see Methods). In an illustrative scenario depicted in Fig 1C, TSS counts from two cell types across five donors are examined. We suppose TSS cluster 1 displays differential usage between the two cell types, with higher usage observed in cell type A, whereas the usage of TSS cluster 2 remains consistent across both cell types. scTSS computes a P-value for each TSS cluster, indicating the significance of its difference between the two cell types.

Accuracy of TSS prediction on 5’ single-cell on-site data

For the TSS prediction task, we considered four existing tools developed for on-site data by incorporating them into the scTSS pipeline (see Methods). Three of these tools were originally designed for CAGE-seq data (CAGEr [17], TSSr [18], and TSRexploreR [19]), and one tool was specifically designed for 5’ scRNA-seq on-site data (SCAFE [26]). A detailed feature comparison between these methods is summarized in S1 Table. For the numerical comparison, we used a COVID-19 (paired-end; on-site) dataset generated from the 10x Genomics Single Cell 5’ Library [33]. This dataset contains 22 single-cell samples of peripheral blood mononuclear cells (PBMCs) from 13 different donors, including 5 samples from healthy donors, 4 samples from COVID-19 patients with severe symptoms, 7 samples from COVID-19 patients with moderate symptoms, and 6 samples from convalescent patients.

After applying the above four methods to the COVID-19 dataset via our scTSS pipeline, we first compared their accuracy in TSS cluster prediction. To assess the accuracy of predicted TSS clusters, we compared them with annotated TSSs from the FANTOM5 dataset [34]. It is important to note that the FANTOM5 annotation encompasses samples from various tissue and cell types, whereas the COVID-19 dataset exclusively consists of PBMCs. To ensure a more reliable evaluation, we retained an annotated TSS if it was detected in more than 30 FANTOM5 samples of immune cells. Using the FANTOM5 annotation as a reference, we computed the genomic distances between the dominant TSS (the most frequently used TSS within a TSS cluster) in each predicted TSS cluster and the closest annotated TSS, particularly focusing on genes with only one annotated TSS (Fig 2A). Analyzing the distributions of these distances, we observed a center around 0 bp regardless of the TSS prediction method being used, suggesting consistent agreement between the predicted TSS clusters and annotated TSSs. In addition, it’s worth noting that the standard deviation of distance was 46 bp and 85 bp for TSSr and SCAFE, respectively, significantly smaller than those observed for CAGEr (181 bp) and TSRexploreR (183 bp).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. TSS prediction results of the COVID-19 dataset. (A) The distribution of genomic distance between predicted dominant TSSs and the corresponding annotated TSSs using all genes with one annotated TSS. A positive distance means the predicted dominant TSS is on the 5’ side of the annotated TSS and vice versa. The analysis was restricted to TSS clusters whose predicted dominant TSSs fell within 500 bp of the annotated TSSs, since a larger distance more likely indicates missing annotations. (B) Accuracy of TSS cluster prediction compared with the complete FANTOM5 annotation. The precision, recall, and F1 scores were first calculated for each sample, and then averaged across samples. The error bars indicate the standard deviation of the scores. (C) The proportion of different types of predicted dominant TSSs based on SCAFE. The classification of TSS locations was based on the GRCh38 reference.

https://doi.org/10.1371/journal.pcbi.1012878.g002

Next, we assessed the accuracy of TSS prediction by calculating the precision, recall, and F1 scores of the four methods based on the complete set of annotated TSSs in the FANTOM5 project (Fig 2B; Methods). CAGEr and TSRexploreR showed higher recall rates but lower precision compared to TSSr and SCAFE, resulting in lower overall F1 scores. This result suggests that CAGEr and TSRexploreR may be more prone to false positive predictions on single-cell data. We also calculated accuracy focusing on genes with only one annotated TSS in FANTOM5 (S1 Fig). Consistent with our previous observations based on the all-genes evaluation, SCAFE still achieved the highest F1 score among all methods, followed by TSSr. The high precision of SCAFE may be attributed to the fact that it utilizes a logistic regression model to filter genuine TSSs based on extracted TSS cluster features. We also evaluated the nucleotide proportions at the -1 and +1 positions of the dominant TSSs for each method (S1 Fig). Previous studies suggest a strong preference for PyPu nucleotide (a pyrimidine followed by a purine) at these positions [35]. Notably, both SCAFE and TSSr exhibited higher PyPu proportions at -1 and +1 positions of their predicted dominant TSSs, suggesting they are more precise in predicting TSSs from single-cell on-site data.

We then summarized the statistics of predicted TSS clusters based on the results of SCAFE. On average, SCAFE identified 27,163 TSS clusters from 14,481 genes in each single-cell sample, with a standard deviation of 2,819. An average of 5,494 genes had multiple predicted TSS clusters (S1 Fig), and the average width of the TSS clusters for each sample ranged from 76 bp to 83 bp (S1 Fig). Furthermore, we studied the categories of the TSS clusters based on the genomic locations of their dominant TSSs in the GRCh38 reference (Fig 2C). We observed that the proportions of dominants TSSs located in the 5’-most exons versus other exons were consistent across all samples, irrespective of the disease condition. Then, we also quantified the relative usage of TSS clusters in each category based on the number of reads mapped to them (S1 Fig). The abundance of intronic TSS clusters tended to be higher in healthy samples compared to samples under other conditions. This observation suggests that alternative TSS usage may contribute to the response to disease conditions.

Feasibility of TSS prediction on 5’ single-cell near-site data

To predict TSS clusters from 5’ single-cell near-site data, we treated Read 2 from the COVID-19 dataset as the near-site data by disregarding the availability of Read 1. We included CAGEr, TSSr, and TSRexploreR in our analysis by accounting for the near-site data through the scTSS pipeline (Fig 1A; see Methods). SCAFE was not considered as it is not applicable to near-site data. First, we assessed the distribution of the genomic distance between predicted TSS cluster centers and the corresponding annotated TSSs using genes with one annotated TSS in FANTOM5 (S2 Fig). Compared with the on-site data analysis using Read 1 of the same dataset, we still observed the centering around 0 bp with all three methods, with TSSr leading to the most noticeable peak. The standard deviation of the distances for TSSr, CAGEr, and TSRexploreR was 146 bp, 204 bp, and 185 bp, respectively. Second, we calculated prediction accuracy scores using all genes in the FANTOM5 annotation as a reference (S2 Fig). Although TSSr achieved the highest F1 score, its accuracy was significantly lower than that achieved on the on-site data.

To further compare TSS predictions from on-site and near-site data derived from the COVID-19 dataset, we calculated sample-wise accuracy scores for predictions on the near-site data, using predictions on the on-site data as the reference (S2 Fig). This analysis revealed relatively high precision but low recall rates, indicating that most TSS clusters identified from the near-site data (Read 2) were also identified from the on-site data (Read 1). The lower recall rate was likely due to the smaller number of TSS clusters identified from near-site data: using TSSr, we predicted an average of 16,235 TSS clusters per sample from on-site data, compared to 7,567 TSS clusters per sample from near-site data. These results suggest that adapting the TSSr method within the scTSS pipeline for predicting TSS clusters from 5’ single-cell near-site data analysis is feasible, although it yields lower overall accuracy compared to on-site data. Lastly, we summarized the proportions of different categories of TSS clusters based on their center positions (S2 Fig). When compared to TSS clusters identified from the on-site data, we observed an increase in the proportion of clusters in the “other exon” and “intergenic” categories in the near-site data. This discrepancy is likely attributable to the exclusion of near-site clusters originating in annotated intronic regions, as required by the adjustment approach.

Next, we applied the scTSS pipeline (with CAGEr, TSSr, or TSRexploreR) to a single-end (near-site) dataset generated from the 10x Genomics Single Cell 5’ Library [36]. This dataset includes 16 single-cell samples of either synovial fluid (SF) or peripheral blood (PB) from eight arthritis patients, with one SF sample and one PB sample from each patient. We refer to this dataset as the Arthritis dataset. We first evaluated the accuracy of TSS clustering based on the distance between the center of the predicted TSS clusters and the annotated TSSs for genes with a single annotated TSS (Fig 3A). The same FANTOM5 annotation as in the COVID-19 on-site data analysis was used. We still observed the centering around 0 bp with all three methods, with TSSr led to the most noticeable peak. Nevertheless, compared to the on-site data, predictions on near-site data resulted in higher variability. The standard deviation of the distances for TSSr, CAGEr, and TSRexploreR was 132 bp, 205 bp, and 194 bp, respectively. We further evaluated the accuracy of TSS clustering using all genes based on the precision, recall, and F1 scores (Fig 3B), and observed a similar pattern as in the on-site data analysis. Compared with CAGEr and TSRexploreR, TSSr had higher precision but lower recall rates. When restricting to one-annotated-TSS genes, TSSr outperformed the other two methods in both precision and recall (S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. TSS prediction results of the Arthritis dataset. (A) The distribution of the genomic distance between predicted TSS cluster centers and the corresponding annotated TSSs using all genes with one annotated TSS. A positive distance means the predicted center is on the 5’ side of the annotated TSS and vice versa. The analysis was restricted to TSS clusters whose centers fell within 500 bp of the annotated TSSs. (B) Accuracy of the TSS cluster prediction compared with the FANTOM5 annotation. The precision, recall, and F1 scores were first calculated for each sample, and then averaged across samples. The error bars indicate the standard deviation of the scores. (C) Gene numbers with varying number of predicted TSS clusters based on the prediction of TSSr. The bars represent the mean value across samples, and the error bars represent the the standard deviation. (D) The proportion of different types of predicted TSS cluster centers based on TSSr. The classification of center locations was based on the GRCh38 reference. When the original near-site clusters were located in an intronic region, indicating missing annotations, we could not confirm the locations of genuine TSS centers. Therefore, we excluded those clusters from the categorization.

https://doi.org/10.1371/journal.pcbi.1012878.g003

Based on the near-site clusters generated by TSSr, the scTSS pipeline identified an average of 7,378 TSS clusters from 4,765 genes per single-cell sample, with a standard deviation of 510. In addition, an average of 1,650 genes had multiple predicted TSS clusters (Fig 3C). The average length of these predicted TSS clusters ranged from 130 bp to 142 bp (S3 Fig). For the Arthritis dataset, we also summarized the categories of TSS clusters based on the genomic locations of their centers (Fig 3D). In both PB and SF samples, the majority of the cluster centers were located on non-5’-most exons, followed by 5’-most exons and intergenic regions. We then assessed the average usage of different categories of TSS clusters, which revealed that approximately 50% of RNA transcripts originated from the 5’-most exons, while the remaining 50% originated from other exons (S3 Fig).

Comparison of DU analysis based on simulated single-cell data

After TSS clusters are predicted and quantified, an important task is to perform differential usage (DU) analysis to identify TSS clusters whose relative usage significantly differs between biological groups. There have been methods developed for differential usage of transcripts based on bulk sequencing data, such as DEXSeq [37], which can be applied to perform DU analysis on pseudo-bulk data transformed from single-cell samples. There have also been methods for differential gene expression analysis, such as MAST [38], IDEAS [39], and the Wilcoxon rank sum test (both at single-cell and pseudo-bulk levels), which may be adapted for DU analysis (see Methods). However, their actual performance in the context of DU analysis given single-cell samples from multiple donors has not been comprehensively evaluated. To evaluate the accuracy of these methods in DU analysis, we designed a simulation study considering sample-level variability with different sample sizes. In this study, we also considered two Binomial generalized linear mixed models (GLMMs): one at the single-cell level and the other at the pseudo-bulk level. The Binomial GLMM is capable of modeling gene expression changes as a confounding factor while simultaneously accounting for sample-level variability (see Methods).

In the simulation, we considered the differential TSS usage analysis between two biological groups (i.e., two cell types). We generated TSS-cluster-level expression data corresponding to different number of single-cell samples per group (5, 10, 15, 20, and 25), combined with varying cell numbers per sample (100 and 500). In each single-cell sample, there were 2,000 genes. Of these, 1,000 genes had TSS clusters following the null hypothesis (same mean usage in the two cell types), while the other 1,000 genes had TSS clusters following the alternative hypothesis (different mean usage in the two cell types). In addition to these simulation settings, we considered the existence of outliers and evaluated method performance given different degrees of outlier frequency (represented by π) and outlier deviation (see Methods). After applying the aforementioned methods for DU analysis, we calculated their type I errors and statistical power, using a threshold of 0.05 on P-values corrected by the Benjamini-Hochberg procedure.

In Fig 4, we present the simulation results when the number of cells in each sample was 500. The results show that MAST and the Wilcoxon test (both at single-cell and pseudo-bulk levels) were not able to control the type I error at the target level (Fig 4A). We observed a type I error rate close to 1 for MAST in all simulation settings. Since MAST directly models read counts, it cannot account for potential gene expression changes between biological groups, which lead to changes in observed TSS-cluster-level counts even when TSS usage remains constant. While the Wilcoxon test had a smaller type I error rate than MAST, its error was still highly inflated in most settings. Furthermore, the single-cell-level Wilcoxon test experienced more severe type I error inflation compared to the pseudo-bulk-level test. In contrast, DEXSeq, IDEAS (modified for DU analysis), and Binomial GLMMs demonstrated effective type I error control. For Binomial GLMM, the single-cell-level and pseudo-bulk-level models led to almost identical results, so the two lines overlap in Fig 4. When the number of cells per sample was reduced to 100 (S4 Fig), we obtained similar results, with DEXSeq, IDEAS, and Binomial GLMM more effectively controlling the type I error compared with MAST and the Wilcoxon tests.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Simulation study of DU analysis with 500 cells per sample. (A) Comparison of type I error rates between seven TSS DU testing methods given various sample size, outlier frequency (π), and degree of outlier deviation. (B) Comparison of statistical power between the seven TSS DU testing methods. The results of bulk-level and cell-level Binomial GLMMs were virtually identical, resulting in overlapping lines.

https://doi.org/10.1371/journal.pcbi.1012878.g004

Regarding statistical power, we found that the Binomial GLMM and DEXSeq, both parametric models, demonstrated better statistical power than IDEAS, a nonparametric model, particularly when the number of samples was small or the outlier deviation was high. This observation held constant across different cell counts per sample (Fig 4B and S4 Fig). In most simulation settings, the statistical power of the Binomial GLMM and DEXSeq was highly comparable, with DEXSeq showing slightly lower power under conditions of small sample size or high outlier deviation. Using the Binomial GLMMs, we observed again that the tests at the single-cell and pseudo-bulk levels yielded almost identical results, with the difference in P-values for each TSS cluster being less than . In addition, we observed the expected trend of decreasing power with increasing outlier frequency or deviation for all three models.

We also compared the running time of these methods (S5 Fig). The pseudo-bulk level Binomial GLMM was the fastest method when the cell number per sample was 500, followed by DEXSeq and the pseudo-bulk-level rank test, all of which demonstrated rapid execution speeds. In summary, our simulation analysis identifies pseudo-bulk-level Binomial GLMM and DEXSeq as the two best methods for DU analysis for their higher accuracy, robustness to outliers, and computational efficiency.

scTSS reveals differential TSS usage in real data

After evaluating the accuracy of DU analysis on simulated data, we employed the scTSS pipeline to identify differential TSS usage between B cells and activated T cells from the COVID-19 dataset, using samples from healthy donors. On average, there were 579 activated T cells and 387 B cells in each sample. We used SCAFE for the prediction of TSS clusters, and the pseudo-bulk-level Binomial GLMM for differential testing. In the DU analysis, we filtered out genes expressed in fewer than 10% cells (combining activated T and B cells) per sample and genes with only a single identified TSS cluster. After the filtering step, we performed the differential TSS usage test on 3,243 TSS clusters in 1,397 genes, and obtained 981 differential TSS clusters with BH-adjusted P-values  <  0.05.

Based on the P-values yielded by scTSS, we selected the top 200 TSS clusters up-regulated in each cell type, and compared their average usage in different samples. A clear pattern of differential TSS usage can be observed between B cells and activated T cells (Fig 5A). For example, three TSS clusters were identified on the IFITM2 gene, and all of them presented obvious differential usage between the two cell types (BH-adjusted P-values  <  ). Among them, one TSS cluster (chr11:308,133-308,161:+) was up-regulated in activated T cells and the other two (chr11:314,141-314,155:+ and chr11:314,040-314,063:+) were up-regulated in B cells (Fig 5B-C). This new observation is complementary to the previous finding that the expression of IFITM2 is up-regulated during T-cell activation [40]. Another example is the NCF1 gene, on which we identified two TSS clusters, and both of them demonstrated differential usage (P-values  <  ) (Fig 5D). NCF1 encodes neutrophil cytosolic factor 1, which was shown to regulate T cell activation in autoimmunce diseases like arthritis [41]. To further analyze the top differentially used TSSs identified by scTSS, we conducted a motif analysis on their promoter sequences (S6 Fig; Methods). The motifs with the smallest P-values included those of ETS1 and RUNX1, which have been reported as important factors in the development of T cell lineages [42,43], as well as motifs of ERG and PU.1, which are responsible for controlling B cell programs [44,45]. It is also worth noting that, based on a differential gene expression (DGE) analysis of all 1,397 genes used in the DU analysis, differences in TSS usage may arise from regulatory mechanisms independent of gene expression changes. Among the genes analyzed, 37.3% were not differentially expressed but exhibited differentially used TSS clusters, while 9.81% were differentially expressed without corresponding changes in TSS cluster usage (S7 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Differential TSS usage analysis between the activated T (ACT) cells and B cells in the COVID-19 dataset. (A) Heatmap of normalized average TSS usage in ACT and B cells. For each TSS cluster, its sample-specific usage was first calculated by taking the average across all cells in that sample. Then, the average usage was normalized across samples using the min-max normalization. (B) tSNE plot of all ACT and B cells among healthy donors. (C) tSNE plots of ACT and B cells colored by the usage of different TSS clusters on IFITM2. Cells in which IFITM2 was not detected were excluded. (D) tSNE plots of ACT and B cells colored by the usage of different TSS clusters on NCF1. Cells in which NCF1 was not detected were excluded.

https://doi.org/10.1371/journal.pcbi.1012878.g005

Since DEXSeq and Binomial GLMM demonstrated similar performance in the simulation study, we also compared their effectiveness in detecting differentially used TSS clusters between activated T cells and B cells. The comparison suggested a strong agreement between the two sets of identified differential TSS clusters, with 835 TSS clusters reported by both methods (S8 Fig). Another 146 and 25 TSS clusters were uniquely identified by Binomial GLMM and DEXSeq, respectively. We ranked the top 200 differentially used TSS clusters based on the P-values generated by Binomial GLMM and compared their ranks as determined by DEXSeq (S8 Fig). Our analysis revealed a high level of agreement between the two methods. Notably, for the top 115 TSS clusters, there was complete concurrence in their rankings between these methods.

Next, we employed the scTSS pipeline to identify differential TSS usage between T cells from the SF and PB samples in the Arthritis dataset. We used TSSr for the prediction of near-site TSS clusters, and pseudo-bulk-level Binomial GLMM for differential testing. After we filtered out genes expressed in fewer than 50% samples and genes with only a single predicted TSS cluster, the differential test was performed on 9,060 TSS clusters from 3,019 genes, and led to 74 differential TSS clusters with BH-adjusted P-values  <  0.05. The usage of the 40 most significant TSS clusters are displayed in S9 Fig. A clear pattern of differential TSS usage between the PB and SF samples can be observed. For example, a TSS cluster on the CHCHD5 gene (chr2:112,585,869-112,585,937:+) presented higher abundance in PB samples compared with SF samples across the eight patients (S9 Fig). In contrast, another TSS cluster on the UBR2 gene (chr6:42,563,957-42,564,104:+) demonstrated higher abundance in SF samples than PB samples. These two TSS clusters were both among the 10 most significant TSS clusters identified by the Binomial GLMM. Similar to the observation from the COVID-19 analysis, we observed little overlap between genes with differential expression and differential TSS usage in the Arthritis data analysis (S7 Fig).

We also compared the differential testing results of the Binomial GLMM and DEXSeq on this dataset. The two methods identified 31 common differential TSS clusters, while DEXSeq and Binomial GLMM uniquely identified 6 and 43 differential TSS clusters, respectively (S10 Fig). Furthermore, we ranked the top 50 differentially used TSS clusters based on the P-values generated by Binomial GLMM and compared these ranks with those determined by DEXSeq (S10 Fig). Similar to the COVID-19 data analysis, we observed a high level of concordance between the two methods among the most significant TSS clusters. We then closely examined the differential TSS clusters identified exclusively by each method. For Binomial GLMM, 19 of the unique TSS clusters were up-regulated in SF and the other 24 TSS clusters were down-regulated in PB (S11 Fig). The two most significant TSS clusters were chr16:29,806,824-29,806,868:+ (on gene MAZ) and chr1:28,369,595-28,369,608:+ (on gene PHACTR4), both showing higher usage in the PB samples (S12 Fig). For DEXSeq, all 6 TSS clusters uniquely identified by it were up-regulated in PB (S11 Fig). Among these, the top two TSS clusters were chr2:88,857,651-88,857,685:- (on gene IGKC) and chr14:9,002,994-9,003,015:+ (on gene ERI1) (S13 Fig).

To further evaluate the performance of the bulk-level Binomial GLMM and DEXSeq, we conducted additional sub-sampling analyses using both the COVID-19 and Arthritis datasets. For each dataset, two types of sub-sampling were performed: (1) sub-sampling cells within each sample and (2) sub-sampling samples within each condition. In the cell-level sub-sampling, we randomly selected 80%, 60%, and 40% of cells from each sample and conducted DU testing using both methods. In the sample-level sub-sampling, we randomly selected subsets of samples within each condition for DU testing. A TSS cluster was considered differentially used if its BH corrected P-value was  <  0.05. Each sub-sampling procedure was repeated 10 times. In the sub-sampling analysis of the COVID-19 dataset (S14 Fig), consistent with the findings from the original data, we observed a significant overlap between the differential TSS clusters identified by the two methods. However, the bulk-level Binomial GLMM consistently identified more differential TSS clusters than DEXSeq in both cell-level and sample-level sub-sampling analyses, suggesting superior statistical power. Similarly, in the sub-sampling analysis of the Arthritis dataset (S14 Fig), across most parameter settings, the bulk-level Binomial GLMM consistently identified a greater number of differential TSS clusters compared to DEXSeq, with the majority of TSS clusters identified by DEXSeq also detected by the bulk-level Binomial GLMM. In the small sample size scenario with only three samples in each condition, DEXSeq outperformed bulk-level Binomial GLMM but its performance was very unstable across independently repeated experiments.

scTSS enables discovery of cell subpopulations based on TSS-specific expression

As TSS isoforms have been shown to present cell-type-specific expression, we next investigated the possibility to identify cell subpopulations using the abundance of TSS clusters. For this analysis, we selected all the naïve T cells from the five healthy donors in the COVID-19 dataset. On average, there were 1,629 naïve T cells per sample with a standard deviation of 155. After obtaining the TSS usage matrices of these samples, we performed a clustering analysis based on TSS usage (see Methods), and identified five clusters in the naïve T cells (Fig 6A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Clustering analysis of naïve T cells based on TSS-specific expression. (A) tSNE plot of naïve T cells based on TSS-specific expression. (B) Heatmap of sample-level TSS usage between clusters 3 and 4. The top 200 differential TSS clusters were selected based on P-values. For each TSS cluster, the min-max normalization was performed on its usage across samples.

https://doi.org/10.1371/journal.pcbi.1012878.g006

To further evaluate the difference in TSS cluster usage between the five identified clusters, we applied scTSS (bulk-level Binomial GLMM) to perform the DU test. We focused on cluster 3 and cluster 4, both of which include at least 50 cells from each sample. After we filtered out genes expressed in fewer than 10% cells (combining cells in cluster 3 and 4), 3,390 TSS clusters were tested, and 2,034 of them had BH-adjusted P-values  <  0.05. We selected the top 200 TSS clusters based on P-values and visualized their relative usage by sample and cluster (Fig 6B). Hierarchical clustering of the TSS clusters revealed two distinct patterns: one group of 105 TSS clusters had higher abundance in cluster 3, and the other group of 95 TSS clusters had higher abundance in cluster 4. This observed pattern was consistent among the five donors. Our analysis demonstrates that usage of TSS clusters can be used to identify cell subpopulations with distinct TSS expression patterns.

TSS cluster prediction for a single sample

The first step in the scTSS pipeline is to perform TSS cluster prediction on individual single-cell samples. By applying a TSS cluster prediction tool through the scTSS pipeline, we aim to obtain a set of identified TSS clusters for each single-cell sample. We categorize 5’ 10x Genomics Gene Expression Assay into two types: paired-end (on-site) and single-end (near-site) (Fig 1A). In on-site data, both Read 1 and Read 2 contain cDNA information. Significantly, Read 1 can be used to accurately identify the true TSSs, enabling the application of TSS cluster prediction tools originally designed for CAGE-seq data to predict TSS clusters in this type of sequencing. In contrast, in near-site data, only Read 2 carries cDNA information and does not directly pinpoint the exact location of the authentic TSSs. Nonetheless, the gap between the 5’ end of Read 2 and the actual TSS depends on the length of the cDNA fragment.

In the context of on-site data, a variety of TSS cluster prediction tools, originally developed for CAGE-seq data, can be utilized. Examples include CAGEr [17], TSSr [18], and TSRexploreR [19]. These methods are all designed to remove any potential additional G-cap at the 5’ end of Read 1, thereby enabling precise identification of TSSs. However, a common challenge we have observed is memory exhaustion when handling large BAM files, which is common in single-cell data analysis. To mitigate this in the scTSS pipeline, we first convert the BAM file into a mapped TSS table before performing TSS cluster analysis. The mapped TSS refers to a genomic location identified by mapping the 5’ end of Read 1 to the genome. A mapped TSS table is a tab-separated file including the following columns: chromosome, location, count, and strand. Using the mapped TSS table instead of the BAM file can effectively avoid memory exhaustion while retaining the necessary information needed to identify TSS clusters. Notably, two recent tools named SCAFE [26] and CamoTSS [27] are specifically tailored for single-cell on-site data, offering enhanced TSS clustering with potential removal of strand invasion artifacts. In situations where these tools, specifically for those designed for bulk CAGE-seq data, are not able to yield single-cell level expression matrices, scTSS incorporates a solution, enabling the generation of single-cell TSS cluster expressions, which will be discussed in subsequent sections.

For near-site data, scTSS provides a method to estimate the locations of genuine TSS clusters. The method begins by obtaining the locations of near-site clusters, which are identified from the 5’ end of Read 2, using a TSS cluster prediction tool for on-site data. After obtaining the near-site TSS clusters, we adjust their locations to approximate genuine TSS clusters. First, when the original near-site clusters are located in an intronic region, indicating missing annotations, we exclude them from downstream analysis as it is not possible to determine the genuine TSSs. Next, we learn the distribution of the distance between the near-site clusters and corresponding genuine TSSs using genes with a single annotated TSS. For each gene, we calculate the smallest distance between the centers of near-site TSS clusters and the annotated TSS. Distances greater than 1,000 bp are ignored, as they most likely indicate the presence of novel, unannotated TSSs. As an example, the distributions of the Arthritis dataset are shwon in S3 Fig. The weighted mean of all distances is then calculated across the single-annotated-TSS genes, where the weight corresponds to the read number supporting the near-site clusters. This weighted mean is referred to as the adjustment distance. Finally, we adjust the near-site clusters by this adjustment distance to obtain the estimated TSS clusters. The adjustment is performed on the transcriptome to ensure the procedure is intron-aware.

Differential TSS cluster usage analysis between cell populations

(1) TSS cluster merging for multiple samples

When TSS cluster prediction tools analyze samples individually, directly comparing TSS clusters across different samples becomes challenging due to the prediction of non-identical but overlapping clusters. An example of this challenge is when the predicted TSS clusters for the same gene show slight variations in their positions but correspond to the same set of genuine TSS isoforms. Therefore, when our objective involves comparing the utilization of TSS clusters among different samples, it becomes necessary to carry out TSS cluster merging beforehand. To achieve the merging of TSS clusters across various samples, scTSS first pools predicted TSS clusters from all samples. Following this, the pooled TSS clusters are separated into unified, non-overlapping TSS clusters using the “disjoin" function from the R package GenomicRanges [48]. Then, we can perform TSS cluster quantification in each sample for a unified set of TSS clusters.

(2) TSS cluster quantification at single-cell level

For on-site data, the expression level of a TSS cluster within an individual cell is determined by counting the reads from that cell whose 5’ end of Read 1 is located within the TSS cluster. For near-site data, the expression level of a TSS cluster is derived from its corresponding near-site cluster, by counting Read 2 whose 5’ end is located within the near-site cluster from that cell.

(3) Differential usage analysis by generalized linear mixed models

In this section, we introduce the binomial generalized linear mixed model (GLMM) used in the differential TSS usage analysis, which is also incorporated in the scTSS pipeline and software. We quantify the observed usage of a TSS cluster as the proportion of transcripts from a gene using any TSS in this cluster. For example, for a gene with T TSS clusters, the observed usage of TSS cluster t ( t = 1 , 2 , … , T )  is defined as:

(1)

where is the transcript count of the t-th TSS cluster. In the context of single-cell data with unique molecular identifiers (UMIs), corresponds directly to the read count associated with cluster t. To better understand how data preparation affects DU analysis, we explore two types of modeling. The first type models TSS usage at the single-cell resolution, while the other type models TSS usage at the bulk level by aggregating data from each single-cell sample into a pseudo-bulk sample.

Suppose there are J single-cell samples, K biological conditions, and cells in sample j. We consider one TSS cluster each time. For single-cell level modeling, we denote as the read count of a specific TSS cluster in cell from sample j ( j = 1 , 2 , … , J ) , and we use to denote the condition of this sample. Then we assume that follows a binomial distribution,

(2)

where is the overall read count of all TSS clusters on the same gene for cell i in sample j, and is the (unobserved) true usage of the TSS cluster of interest in sample j. Then, we further assume a linear relationship between the logit of and sample/condition-level variables:

(3)

where is an intercept, and is a vector of the condition variable(s) and other available covariates. In the special case where no other covariates are available, is a binary vector of size  ( K − 1 ) × 1. Suppose we select condition 1 as the baseline condition (without loss of generality). If , the elements of are all zero values; if , the elements of are all zero values except for the -th element. β denotes the corresponding coefficient(s) to be estimated. is the sample-level random effect for sample j, which is used to account for the sample-level heterogeneity. We assume The above model can be estimated using the “glmer” function from the R package lme4 [49]. For the DU analysis, we test the significance of usage change based on the statistical significance of β, using the “anova” function from the lme4 package.

For pseudo-bulk level modeling, we let

(4)

be the pseudo-bulk level read count of the TSS cluster of interest, and

(5)

be the pseudo-bulk level read count of the gene. Then, we assume a Binomial model

(6)

where denotes the true TSS cluster usage of sample j. We further assume the same link function as specified in formula (3), indicating a linear relationship between the logit of and sample/condition-level variables. The pseudo-bulk level model is also estimated and tested by the lme4 package.

Simulation study of DU analysis

(1) Generation of simulated data

In order to generate simulated TSS cluster expression data with ground truth information on TSS cluster usage, we first simulated the read counts of genes, and then simulated the read counts of TSS clusters. We considered the simulation setting of two biological conditions (i.e., cell types), with single-cell samples in each condition, and single-cells in each sample. To simulate gene-level read counts, we utilized the R package scDesign2 [50], which learns distributional characteristics of single-cell gene expression levels from real data. In this study, we used scDesign2 to learn gene expression parameters from the real single-cell gene expression levels of the healthy donors from the COVID-19 dataset [33]. To account for the potential gene expression changes between biological conditions, which is a common counfouding factor in DU analysis, we focused on two different cell types, B cells and activated T cells. Before generating synthetic gene expression data for these two celltypes, we used the approach below to filter genes. First, we only retained genes expressed in at least 1% of cells in all five samples. Second, we selected the top 2,000 genes with the highest detection rates in B cells and activated T cells. Given the filtered real data, we could use scDesign2 to generate synthetic gene expression data for varying sample sizes and cell numbers.

To simulate TSS-cluster-level count data, we assumed there were two TSS clusters on each gene. For the first TSS cluster on a gene, we let represent its usage in cell i  ( i = 1 , 2 , … , I )  from sample j  ( j = 1 , 2 , … , J )  under biological condition k  ( k = 1 , 2 ) . The usage was randomly drawn from a Beta distribution:

(7)

where and β were the two shape parameters. Therefore, the corresponding mean TSS usage was . The values of and were determined depending on the null and alternative hypotheses. We assumed that TSS clusters on 1,000 genes followed the null hypothesis and TSS clusters on the other 1,000 genes followed the alternative hypothesis.

Under the null hypothesis, we assumed that the expected TSS cluster usage was the same in the two biological groups. Thus, we first generated a baseline usage, a, from Uniform ( 0 . 1 , 0 . 9 ) . In addition, to account for sample-level variability, we drew for sample j in group k:

(8)

where π represented the possibility of a sample being an outlier regarding the usage of this TSS cluster, and and represented the deviation of outliers from the expected values. We considered π ∈ { 0 . 1 , 0 . 2 , 0 . 3 , 0 . 4 }  and (small outlier deviation) or (large outlier deviation). The expected usage of the first TSS cluster in sample j of group k was finally determined as follows:

(9)

Under the alternative hypothesis, we assumed that the expected TSS cluster usage was different in the two biological groups. We generated as the baseline usage in group 1, and as the baseline usage in group 2, where d ∼ Uniform ( 0 . 1 , 0 . 5 ) . As for the sample-level effect, we assume follows a mixture distribution:

(10)

(11)

where took the same values as under the null hypothesis. Then, the expected usage of the first TSS cluster in sample j of group k was determined as follows:

(12)

Once ’s were obtained, we draw TSS usage () for individual cells based on formula (7). The usage of the other TSS cluster on the same gene was automatically obtained as . Given the single-cell level TSS usage, the TSS-cluster-level read counts were generated by a Binomial distribution with the gene-level read count as the sample size.

(2) Alternative methods for DU analysis

Other than the two Binomial GLMM models included in scTSS toolkit, we applied five alternative methods for DU analysis of the simulated data, including DEXSeq [37], IDEAS [39], MAST [38], and the Wilcoxon rank sum test (both at single-cell and pseudo-bulk level). For all methods, we applied the Benjamini-Hochberg procedure to adjust P-values for multiple testing.

DEXSeq was initially designed for detecting alternative transcription usage for bulk RNA-seq data. Thus, we formed a pseudo-bulk sample for each single-cell sample in the simulated data. When applying DEXSeq for DU analysis, we followed its tutorial and set the formula as “sample + TSS + condition:TSS” for the full model and “sample + TSS” for the reduced model (null hypothesis).

IDEAS was originally designed for differential gene expression analysis based on multiple single-cell samples. It first estimates the distribution of gene expression in each individual, and then performs the test by comparing within-group distances of individuals with between-group distances. However, IDEAS cannot be directly applied to TSS usage data, so we made the following modifications for DU analysis. First, we converted the TSS cluster counts into TSS cluster usage. Second, we used the Wasserstein-1 distance to measure the difference between the empirical CDFs of TSS cluster usage between any two single-cell samples. Third, we performed the permutation 2000 times for each TSS cluster to obtain its P-value.

MAST was designed for differential gene expression analysis based on single-cell RNA-seq data. When applying MAST on TSS cluster counts for DU analysis, in the “zlm” function, we set the formula for the full model as “sample+condition”. When conducting the likelihood ratio test to obtain the P-values, we set “doLRT=condition”.

To perform the Wilcoxon rank sum test at the pseudo-bulk level, we treated the cells from each biological group and single-cell sample as a pseudo-bulk sample, and added up the read counts of a TSS cluster across the cells before calculating TSS usage. To perform Wilcoxon rank sum test at the single-cell level, we calculated the TSS cluster usage in individual cells. The Wilcoxon rank sum test was directly applied to the calculated TSS cluster usage.

Real data analysis

The raw sequences in the COVID-19 dataset were processed by Cell Ranger software (“count” mode) to align reads to the reference genome using default parameters and obtain BAM files. The GRCh38 reference genome was downloaded from 10x Genomics’ website (References-2020-A, https://cf.10xgenomics.com/supp/cell-exp/refdata-gex-GRCh38-2020-A.tar.gz). Four methods, including CAGEr [17], TSSr [18], TSRexploreR [19], and SCAFE [26], were applied via the scTSS pipeline to perform TSS cluster prediction and DU analysis. For CAGEr, the “method” parameter was set as “distclu” in the “clusterCTSS” function, and other parameters were set to default. For TSSr, we first ran the “filterTSS” function followed by the “clusterTSS” function with default parameters. For TSRexploreR, we set “threshold=3” and “max_distance=25” in the “tss_clustering” function, which were used in the software tutorial. For SCAFE, we followed the standard workflow in its tutorial, and used the default pretrained logistic regression model in the “tool.cm.filter” function. For the motif analysis on differentially used TSSs between activated T cells and B cells, we used the motif discovery algorithm (findMotif.pl) from Homer [51]. For each top differentially used TSS cluster, we used the 150 bp upstream of the dominant TSS as the region for motif identification. The background region was set as the 150 bp upstream of the dominant TSS in all the predicted TSS clusters used in the DU analysis. The “len” parameter was set to 8. To analyze the subpopulations of naïve T cells in the five healthy samples by TSS usage, we performed TSS prediction on each sample and TSS merging across samples. Then, we converted the TSS counts to TSS usage for each TSS cluster. Lastly, we followed the clustering procedure in Seurat [52] to identify cell subpopulations based on TSS usage.

For the Arthritis dataset, the raw sequences were aligned using the same approach as described for the COVID-19 dataset. CAGEr, TSSr and TSRexploreR were applied via the scTSS pipeline to perform TSS cluster prediction and DU analysis. The parameter settings were the same as described above. SCAFE was excluded as it is not applicable to near-site data.

In DU analysis on both datasets, we applied the Benjamini-Hochberg procedure to adjust P-values for multiple testing.

Calculation of precision, recall, and F1 score

The gene-level precision was defined for each gene based on the proportion of predicted TSS clusters that were observed in the annotation. We considered a predicted TSS cluster to be observed if it overlapped with an annotated TSS on the same gene. The gene-level recall was defined for each gene based on the proportion of annotated TSSs that were predicted by scTSS. We considered an annotated TSS as being predicted if it was included by a predicted TSS cluster on the same gene. Lastly, the overall precision (P) and recall (R) for a single-cell sample were obtained by calculating the average of gene-level precision and recall, respectively. In addition, the F1 score for a single-cell sample was defined as

(13)