What it does

StringTie is a fast and highly efficient assembler of RNA-Seq alignments into potential transcripts. It uses a novel network flow algorithm as well as an optional de novo assembly step to assemble and quantitate full-length transcripts representing multiple splice variants for each gene locus. Its input can include not only the alignments of raw reads used by other transcript assemblers, but also alignments of longer sequences that have been assembled from those reads. In order to identify differentially expressed genes between experiments, StringTie's output can be processed by specialized software like Ballgown, Cuffdiff or other programs (DESeq2, edgeR, limma etc.).

Inputs

StringTie takes as input a BAM (or SAM) file of paired-end RNA-seq reads, which must be sorted by genomic location (coordinate position). This file contains spliced read alignments and can be produced directly by programs such as HISAT2. We recommend using HISAT2 as it is a fast and accurate alignment program. Every spliced read alignment (i.e. an alignment across at least one junction) in the input BAM file must contain the tag XS to indicate the genomic strand that produced the RNA from which the read was sequenced. Alignments produced by HISAT2 (when run with --dta option) already include this tag, but if you use a different read mapper you should check that this XS tag is included for spliced alignments.

NOTE: be sure to run HISAT2 with the --dta option for alignment (under Spliced alignment options), or your results will suffer.

Also note that if your reads are from a stranded library, you need to choose the appropriate setting under Specify strand information above. As, if Forward (FR) is selected, StringTie will assume the reads are from a --fr library, while if Reverse (RF) is selected, StringTie will assume the reads are from a --rf library, otherwise it is assumed that the reads are from an unstranded library (The widely-used, although now deprecated, TopHat had a similar --library-type option, where fr-firststrand corresponded to RF; fr-secondstrand corresponded to FR). If you don't know whether your reads are from are a stranded library or not, you could use the tool RSeQC Infer Experiment to try to determine.

As an option, a reference annotation file in GTF/GFF3 format can be provided to StringTie. In this case, StringTie will prefer to use these "known" genes from the annotation file, and for the ones that are expressed it will compute coverage, TPM and FPKM values. It will also produce additional transcripts to account for RNA-seq data that aren't covered by (or explained by) the annotation. Note that if option -e is not used the reference transcripts need to be fully covered by reads in order to be included in StringTie's output. In that case, other transcripts assembled from the data by StringTie and not present in the reference file will be printed as well.

NOTE: we highly recommend that you provide annotation if you are analyzing a genome that is well-annotated, such as human, mouse, or other model organisms.

Outputs

StringTie's primary output is

• a GTF file containing the Assembled transcripts

Optionally, it can output

• a TSV (tab-delimited) file of Gene abundances

If a reference GTF/GFF3 file is used as a guide, StringTie can also output:

• a GTF file containing all fully-covered reference transcripts in the provided reference file that are covered end-to-end by reads
• Files (tables) for Ballgown and/or DESeq2/edgeR/limma-voom, which can use them to estimate differential expression

StringTie's primary GTF output

The primary output of StringTie is a Gene Transfer Format (GTF) file that contains details of the transcripts that StringTie assembles from RNA-Seq data. GTF is an extension of GFF (Gene Finding Format, also called General Feature Format), and is very similar to GFF2 and GFF3. The field definitions for the 9 columns of GTF output can be found at the Ensembl site here. The following is an example of a transcript assembled by StringTie as shown in a GTF file:

• seqname: Denotes the chromosome, contig, or scaffold for this transcript. Here the assembled transcript is on chromosome X.

• source: The source of the GTF file. Since this example was produced by StringTie, this column simply shows 'StringTie'.

• feature: Feature type (e.g., exon, transcript, mRNA, 5'UTR).

• start: Start position of the feature (exon, transcript, etc), using a 1-based index.

• end: End position of the feature, using a 1-based index.

• score: A confidence score for the assembled transcript. Currently this field is not used, and StringTie reports a constant value of 1000 if the transcript has a connection to a read alignment bundle.

• strand: If the transcript resides on the forward strand, '+'. If the transcript resides on the reverse strand, '-'.

• frame: Frame or phase of CDS features. StringTie does not use this field and simply records a ".".

• attributes: A semicolon-separated list of tag-value pairs, providing additional information about each feature. Depending on whether an instance is a transcript or an exon and on whether the transcript matches the reference annotation file provided by the user, the content of the attributes field will differ. The following list describes the possible attributes shown in this column:

  • gene_id: A unique identifier for a single gene and its child transcript and exons based on the alignments' file name.
  • transcript_id: A unique identifier for a single transcript and its child exons based on the alignments' file name.
  • exon_number: A unique identifier for a single exon, starting from 1, within a given transcript.
  • reference_id: The transcript_id in the reference annotation (optional) that the instance matched.
  • ref_gene_id: The gene_id in the reference annotation (optional) that the instance matched.
  • ref_gene_name: The gene_name in the reference annotation (optional) that the instance matched.
  • cov: The average per-base coverage for the transcript or exon.
  • FPKM: Fragments per kilobase of transcript per million read pairs. This is the number of pairs of reads aligning to this feature, normalized by the total number of fragments sequenced (in millions) and the length of the transcript (in kilobases).
  • TPM: Transcripts per million. This is the number of transcripts from this particular gene normalized first by gene length, and then by sequencing depth (in millions) in the sample. A detailed explanation and a comparison of TPM and FPKM can be found here, and TPM was defined by B. Li and C. Dewey here.

Gene abundances in tab-delimited format

If StringTie is run with the -A option, it returns a file containing gene abundances. The tab-delimited gene abundances output file has nine fields; below is an example of a gene abundance file produced by StringTie using reference annotation:

• Gene ID: The gene identifier comes from the reference annotation provided with the -G option. If no reference is provided this field is replaced with the name prefix for output transcripts (-l).
• Gene Name: This field contains the gene name in the reference annotation provided with the -G option. If no reference is provided this field is populated with '-'.
• Reference: Name of the reference sequence that was used in the alignment of the reads. Equivalent to the 3rd column in the .SAM alignment.
• Strand: '+' denotes that the gene is on the forward strand, '-' for the reverse strand.
• Start: Start position of the gene (1-based index).
• End: End position of the gene (1-based index).
• Coverage: Per-base coverage of the gene.
• FPKM: normalized expression level in FPKM units (see previous section).
• TPM: normalized expression level in RPM units (see previous section).

Fully covered transcripts matching the reference annotation transcripts (in GTF format)

If StringTie is run with the use reference guide option (-G), it will also return a file with all the transcripts in the reference annotation that are fully covered, end to end, by reads. The output format is a GTF file as described above. Each line of the GTF is corresponds to a gene or transcript in the reference annotation.

Ballgown Input Table Files

An option to output files for Ballgown can be selected under Output files for differential expression? above. If selected, StringTie will return Ballgown input table files containing coverage data for the reference transcripts given with the -G option. These tables have these specific names: (1) e2t.ctab, (2) e_data.ctab, (3) i2t.ctab, (4) i_data.ctab, and (5) t_data.ctab. A detailed description of each of these five required inputs to Ballgown can be found at this link. With this option StringTie can be used as a direct replacement of the tablemaker program included with the Ballgown distribution.

DESeq2/edgeR/limma-voom Input Table Files

DESeq2, edgeR and limma are three popular Bioconductor packages for analyzing differential expression, which take as input a matrix of read counts mapped to particular genomic features (e.g., genes). This read count information can be extracted directly from the files generated by StringTie (run with the -e parameter) by selecting DESeq2/edgeR/limma-voom under Output files for differential expression? above. This uses the StringTie helper script prepDE.py to convert the GTF output from StringTie into two tab-delimited (TSV) files, containing the count matrices for genes and transcripts, using the coverage values found in the output of StringTie -e.

More Information

Evaluating transcript assemblies: A simple way of getting more information about the transcripts assembled by StringTie (summary of gene and transcript counts, novel vs. known etc.), or even performing basic tracking of assembled isoforms across multiple RNA-Seq experiments, is to use the gffcompare program. Basic usage information for this program can be found on the GFF utilities page.

Differential expression analysis:

Together with HISAT and Ballgown (or DESeq2/edgeR/limma-voom), StringTie can be used for estimating differential expression across multiple RNA-Seq samples and generating plots and differential expression tables as described in our protocol paper and shown in a diagram in the StringTie manual here.

Our recommended workflow includes the following steps:

1. For each RNA-Seq sample, map the reads to the genome with HISAT2 using the --dta option. It is highly recommended to use the reference annotation information when mapping the reads, which can be either embedded in the genome index (built with the --ss and --exon options, see HISAT2 manual), or provided separately at run time (using the --known-splicesite-infile option of HISAT2). The SAM output of each HISAT2 run must be sorted and converted to BAM using samtools as explained above.
2. For each RNA-Seq sample, use this StringTie tool to assemble the read alignments obtained in the previous step; it is recommended to run StringTie with the -G option if the reference annotation is available.
3. Run the separate StringTie merge tool in order to generate a non-redundant set of transcripts observed in all the RNA-Seq samples assembled previously. StringTie merge takes as input a list of all the assembled transcripts files (in GTF format) previously obtained for each sample, as well as a reference annotation file (-G option) if available.
4. For each RNA-Seq sample, run this StringTie tool selecting to output files for Ballgown (or DESeq2/edgeR/limma-voom), which will generate tables of transcript and gene estimated abundances (count files). The option -e (Use Reference transcripts only) is not required but is recommended for this run in order to produce more accurate abundance estimations of the input transcripts. Each StringTie run in this step will take as input the sorted read alignments (BAM file) obtained in step 1 for the corresponding sample and the -G option with the merged transcripts (GTF file) generated by stringtie merge in step 3. Please note that this is the only case where the -G option is not used with a reference annotation, but with the global, merged set of transcripts as observed across all samples. (This step is the equivalent of the Tablemaker step described in the original Ballgown pipeline.)
5. Ballgown (or DESeq2/edgeR/limma-voom) can now be used to load the coverage tables generated in the previous step and perform various statistical analyses for differential expression, generate plots etc.

An alternate, faster differential expression analysis workflow can be pursued if there is no interest in novel isoforms (i.e. assembled transcripts present in the samples but missing from the reference annotation), or if only a well known set of transcripts of interest are targeted by the analysis. This simplified protocol has only 3 steps (depicted in the StringTie manual here) as it bypasses the individual assembly of each RNA-Seq sample and the "transcript merge" step. This simplified workflow attempts to directly estimate and analyze the expression of a known set of transcripts as given in the reference annotation file.