Introduction

As infection with Zika virus (ZIKV) is associated with increasing cases of congenital microcephaly and adult Guillain-Barre Syndrome, a characterization of its pathophysiology becomes crucial. A characterization of the molecular effects of infection may help in the development of fetal diagnostics and can accelerate the identification of genes and pathways critical in disease progression. RNA-Sequencing (RNA-Seq) is an effective technology for probing the transcriptome and has been applied to study the effects of ZIKV infection of human neuroprogenitor cells (hNPCs) [1].

While initial analyses of the data conducted a general survey of transcriptome changes upon infection [1–3], they [1,2] used a method, Cufflinks/Cuffdiff [4], that failed to take advantage of the experimental design used in Tang et. al [1]. They [1–3] also did not examine transcriptome dynamics at the isoform level.

We applied the recently-developed kallisto [5] and sleuth [6] programs to improve the accuracy of quantification and to extract information from the data that was previously inaccessible. We found that sleuth’s improved control of false discovery rate [6] resulted in differential transcript and gene lists that are much more specific and that are significantly enriched in neurodevelopmental pathways. They reveal ZIKV’s neural tropism and the host’s response to viral infection. Furthermore, we demonstrate that the combination of accurate kallisto quantification, assessment of inferential variance and the sleuth response error model allows for the detection of post infection isoform-specific changes that were missed in previous analyses.

The sleuth Shiny app drives a freely available website that allows for reproducibility of our analyses, and provides tools for interacting with the data. This makes the dataset useful for analysis by infectious disease experts who may not have bioinformatics expertise.

Methods

We ran kallisto and sleuth on a total of eight RNA-seq samples of ZIKV-infected and mock-infected hNPCs (GEO: Series GSE78711) (See Table 1 for experimental design and description of samples). We used kallisto to pseudoalign the RNA-seq reads and perform bootstraps, using an index based on the ENSEMBL GRC38 Homo sapiens release 85 transcriptome. For single-end read quantification, we used default parameters (kmer size = 31, fragment length = 187 and sd = 70). For each of the eight samples, kallisto quantified transcript abundances and performed 100 bootstraps.

Download:

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

Table 1. Experimental design.

Tang et al. infected two samples with ZIKV and two with a mock infection. Library preparation was performed for each sample to make four cDNA libraries. Each library was then sequenced with MiSeq using paired-end reads and NextSeq using single-end reads.

https://doi.org/10.1371/journal.pone.0175744.t001

The response error model of sleuth was then used to identify differentially expressed transcripts. Sleuth used the bootstraps performed by kallisto to estimate the inferential variance of each transcript, and an adjusted variance was used to determine differential expression for that transcript. This data set had a unique experimental design, however. For each sequencing library corresponding to a biological sample, Tang et al. performed both paired-end and single-end sequencing. To take advantage of the technical replicates performed by Tang et al., we modified sleuth to perform a weighted average of the inferential variance with the number of fragments sequenced (Table 1).

Principle component analysis of the transcript abundances provided a quick verification of the accuracy of our methods, as the first principle component separated the samples by infection status and the second principle component separated the samples by sequencing method (Fig 1).

Download:

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

Fig 1. Principle component analysis.

PCA of the eight samples shows that the primary contributor to variance is ZIKV infection status (ZIKV vs mock), while the secondary component is sequencing method (paired-end vs single-end).

https://doi.org/10.1371/journal.pone.0175744.g001

The data analysis pipeline was performed on a laptop and can be repeated using the provided scripts at http://www.github.com/pachterlab/zika/. The kallisto quantifications, the modified version of sleuth, as well as a script for the pipeline, are available on the github. One can use the script to start the Shiny app, which recreates the statistics and figures referenced throughout this paper, along with interactive data visualization tools. Alternatively, the preloaded sleuth Shiny app can be found via http://128.32.142.223/tang16/.

Results

Using a false discovery rate of 0.05 as the threshold for differential expression, we detected 4610 transcripts across 3646 genes that are differentially expressed between ZIKV-and mock-infected samples. (Fig 2, S1 and S2 Tables) For the 3969 genes that Cuffdiff found differentially expressed but sleuth did not, sleuth reported an average false discovery rate of 0.55.

Download:

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

Fig 2. Venn diagram of differential expression analysis.

Sleuth identified 3646 differentially expressed genes. Cuffdiff identified 6864 differentially expressed genes. 2895 of the 3646 differentially expressed genes were also reported in Tang et. al [1], but they reported an additional 3969 genes that we failed to identify. Furthermore, we found 751 differentially expressed genes corresponding to 5426 transcripts not detected by Cuffdiff.

https://doi.org/10.1371/journal.pone.0175744.g002

It was not surprising that the many differentially expressed genes discovered by Cuffdiff were considered false positives by sleuth. In simulations by Pimentel et al [6], sleuth provided the most accurate false discovery rates, whereas other methods including DESeq2, edgeR, and Cuffdiff2 underestimated their false discovery rates. In other words, these methods provided differential gene lists that had many more false positives than what was suggested by their p-values. The fundamental idea underlying sleuth is that, by using bootstraps to estimate inferential variance, it does not assume a parametric distribution to account for uncertainty in isoform mapping.

Furthermore, we found a few hundred genes with differentially expressed transcripts not identified by Cuffdiff. We ascribe these to the accounting of experimental design and the isoform-level analysis.

Zika induced isoform divergence

Differentially regulated genes may be missed in gene-level analysis for several reasons. Noise in the measurement of highly expressed transcripts can mask expression changes in lowly expressed transcripts. In the case of isoform switching, upregulation in one isoform and downregulation in another may “cancel out.” We identified 108 genes that contain transcript(s) that are significantly upregulated and other transcript(s) that are significantly downregulated, a phenomenon we coin “isoform divergence” (S3 Table). Of these 108 isoform diverging genes, 57 were not considered differentially expressed by Cuffdiff analysis.

We performed a pathway analysis on the 108 genes using Reactome [7]. Enrichment was identified in neuronal system (specifically transmission across chemical synapses and protein-protein interactions at the synapses), developmental biology (specifically axon guidance), immune system, DNA repair, chromatin modifying enzymes, gene expression (rRNA and transcriptional regulation), metabolism, signal transduction, transmembrane transport and vesicle-mediated transport.

One of these 57 isoform diverging genes not picked up by Cufflink is NRCAM, neuronal cell adhesion molecule, which is putatively involved in neuron-neuron adhesion and axonal cone growth. Another is CHRNA7, cholinergic receptor nicotinic alpha 7 subunit. [8] Figs 3 and 4 shows transcript abundances in NRCAM and CHRNA7 across different samples, highlighting isoform-specific changes.

Download:

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

Fig 3. NRCAM is an example of an isoform divergent gene involved in neuron-neuron adhesion.

For a specific gene, the sleuth Shiny app plots the counts corresponding to each transcript and sample. Visualized here are counts for four transcripts of NRCAM across the eight samples, colored by infection status.

https://doi.org/10.1371/journal.pone.0175744.g003

Download:

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

Fig 4. The counts of CHRNA7, another isoform diverging gene, plotted by the sleuth Shiny app.

Visualized here are counts for three transcripts of CHRNA7 across eight samples, colored by infection status.

https://doi.org/10.1371/journal.pone.0175744.g004

A gene ontology (GO) analysis of sleuth-discovered genes showcase neural and head development networks

We performed a side-by-side gene ontology (GO) analysis with the differential genes identified by sleuth and Cuffdiff, using ClueGO [9, 10] over the Biological Processes ontology network, using GO Term Fusion. We set the network specificity to global (GO tree interval: 1–4), using pathways with a minimum of 50 genes and kappa score of 0.5. We highlighted enriched nodes of particular interest and their enrichments in Fig 5.

Download:

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

Fig 5. GO pathways enrichment.

The enriched nodes of particular interest include neuron projection guidance (p-value = 2.7E-3 vs >0.05 with Cuffdiff), cerebral cortex development (1.6E-7 vs >0.05), neuron development (9.9E-6 vs 3.9E-4), neuron projection development (1.8E-6 vs 5.0E-5), nervous system development (3.0E-10 vs 1.0E-9), central nervous system development (6.9E-9 vs 1.0E-4), brain development (2.8E-9 vs 8.0E-4), forebrain development (1.9E-7 vs 4.1E-2), telecephalon development (2.7E-5 vs 5.2E-3), head development (1.3E-6 vs 3.2E-4), and cellular response to stress (9.4E-26 vs 7.3E-22).

https://doi.org/10.1371/journal.pone.0175744.g005

Provided in the supplementary materials are the side-by-side GO analysis results tables (S4 and S5 Tables).

Discussion

RNA-Seq can provide rapid and high resolution probing of infection response, and initial studies of Zika infection highlight isoforms, genes and pathways that may play an important role in disease etiology. However, the simplicity of RNA-Seq library prep and cDNA sequencing belies the complexity of analysis. We have shown that a careful analysis of previously published data can reveal novel targets with higher confidence, and in the process rendering a valuable dataset usable by the community of Zika researchers.

The kallisto and sleuth tools we have used in our analysis are particularly powerful when coupled with the interactive sleuth Shiny application, and our publicly available server provides access to numerous interactive plots and figures that cannot be reproduced in a static publication. This highlights the utility and importance of data sharing [11], and we hope that our analysis, aside from its usefulness for the Zika scientific community, can also serve as a blueprint for future data sharing efforts.

sleuth is a fast and accurate pipeline for analyzing RNA-Seq data that allows for testing at the isoform level. The alignment and quantification pipeline is feasible and compatible with a standard desktop computer. The interactive Sleuth application, made publically available, allows for informative data visualization, including those of library prep fragment size distributions, principle component analysis, and gene and transcript expression changes. We invite the scientific community studying Zika to utilize this toolkit.

Supporting information