TRANSIT provides several statistical methods capable of accomplishing two common types of tasks:

Identifying essential genes in a single growth condition

Bayesian/Gumbel Method

All TnSeq analysis methods are sensitive to insertion density or library saturation (to different degrees). It is important to have sufficient diversity (or saturation) of the Tn mutant library, so that as many of the TA sites in non-essential regions are represented as possible. While saturation rarely achieves 100%, good libraries often have density greater than 50%, whereas libraries with lower density in the 20-30% range are more challenging to analyze and give less confident predictions (because a sequence of TA sites could be missing insertions due to chance).

TRANSIT takes .wig files as input, which contain counts of reads (or unique templates) observed at each each TA site. In this way, TRANSIT accepts datasets prepared with any pre-processing or custom protocol. An optional pre-processor called called TPP is provided with the software distribution for extracting these counts from raw sequencing files (typically in .fastq format, paired-end reads in two files called “read1” and “read2”). Note that this pre-processing procedure is designed for libraries prepared in adherence with the protocol described in [14]. However, other labs might want to apply their own custom pre-processing procedure, particularly if they use an alternative protocol for preparing TnSeq samples for sequencing or if they use a different transposon.

TPP uses BWA (Burroughs-Wheeler Aligner; [21]) to map reads into the genome. The workflow performed by TPP (Fig 4) can be briefly summarized as follows: First, read 1 is analyzed to identify the subset of reads that have a prefix matching the terminus of the Himar1 transposon (ACTTATCAGCCAACCTGTTA). The transposon prefix is stripped off, and the genomic suffix is mapped onto the genome to identify the TA site (and strand) represented by each read. Then, read 2 is analyzed to extract both a random nucleic-acid barcode (See Fig 4) and genomic suffix. The genomic suffix is also mapped onto the genome and represents the end-point of the original DNA fragment (typically a few hundred bp away). All the reads mapping to a given TA site are reduced to unique “template” counts by discarding duplicates that have the same barcode and end-point, and these template counts are written out in .wig format (for input to TRANSIT). (If barcodes were not applied during sample prep or only single-ended data was collected, TPP can be run optionally without providing read 2, in which case raw read counts are output, without reduction to unique template counts.) In our experience, this barcoding data-reduction technique has proved useful in reducing PCR effects, which can artificially bias the read counts depending on which fragments amplify more efficiently. We have observed that this is especially true for noisier datasets, where read counts are highly variable (though recent optimizations in the PCR protocol have mitigated this problem somewhat [10]) and the resulting template counts better reflect the true abundance of distinct mutants in the population. It can take on the order of an hour to process each dataset (depending on size of dataset as well as speed of computer), which is dominated by the time required to align the genomic parts of the reads to the genome using BWA. Ideally, it is recommended to collect datasets with 5-10 million pairs of reads, which often gets reduced to ∼ 2 million unique templates, in order to have sufficient dynamic range of counts for analysis (for example, aiming for a nominal mean of ∼ 50 templates per TA sites, estimated based on ∼ 75,000 TA sites in the M. tuberculosis genome, with library saturation of around 50%).

Multiple statistics are calculated by TPP for diagnostic purposes. The primary metrics used to assess the quality of a TnSeq dataset are insertion density and mean read count, which should be P_ins > 30% and NZmean > 10. Additional statistics are reported, such as number of reads with valid Tn prefixes, number of reads mapping to genome (broken down into read 1, read 2, and both), correlation of reads at each TA sites on forward versus reverse strand, ratio of reads to templates, etc. These metrics are important for diagnostic purposes. In addition, specific nucleotide sequences representing the vector or primer are counted. If the number of mapped reads is low, then the user could check to see if there is a large fraction of reads matching these sequences, which could indicate phage contamination in the library (left over from the original transfection in constructing the library) or excessive primer-dimers lacking genomic inserts generated during sample preparation.

The main TRANSIT interface (Fig 5) allows the user to select an annotation file (in a tab-separated format called “.prot_table”), which contains the definitions of genes and their coordinates. It must match the reference genome to which reads were mapped by TPP. Several .prot_tables are provided online for commonly used genomes, and a script for converting annotations for other organisms from Genbank to .prot_table format is also available.

From the main interface the user can also load datasets in .wig format (e.g. output by TPP), which contain the coordinates and template counts at TA coordinates throughout the genome. Once a dataset is loaded, the corresponding table will be populated with diagnostic information about the dataset, like density and mean read count. This information can be used to compare datasets and identify potential problems. TRANSIT also provides a way to create a scatter plot of the read counts in two datasets from the menu-bar at the top of the interface. In addition, the user can visualize read-counts throughout the genome using TRANSIT’s Track View (also found in the menu-bar; see Fig 1).

Once the user has picked the desired datasets for comparison, they can choose which analysis they wish to perform from the drop-down menu to the right. As soon as a method is selected, the right-hand panel of TRANSIT’s interface is automatically populated with the appropriate parameters for the method selected. The user may use the default parameters (which are intended to work well on most datasets) or change individual parameters as needed. Parameter definitions are provided in the documentation included with TRANSIT.

After TRANSIT completes an analysis, it will create output files in the specified location and automatically add them to a list in the results window to keep track of the results files created in a session. Output files are tab-separated so they can be opened in the user’s preferred spreadsheet software (e.g. Excel). TRANSIT also has the capability of opening results files in a new window, by selecting them from the list of results files and clicking on the “Display Table” button. This list also allows the user to generate custom graphs of the results, such as volcano plots (which plots log-fold change in read counts and adjusted p-values; See Fig 6).

The results window (Fig 7) allows the user to sort on the desired column (e.g. p-values) thus facilitating the identification of genes of interest. The user can right-click on a gene to display the gene in Track View to examine the insertion patterns (Fig 1), or get other method-specific options (like histograms of the permutations obtained with the resampling method; See Fig 3).

To identify essential genes we analyzed the datasets grown on glycerol using the Bayesian/Gumbel Method, which performs an analysis on an individual condition. After loading the glycerol replicates and the annotation file into TRANSIT, and running the Gumbel method with default parameters, we obtained an output file with results.

A total of 674 genes was found to be essential (Z ‾ i > θ e, where θ_e is a threshold determined by the method used to control the FDR) by the Gumbel method (16.3%), matching expectations that typically 15% of bacterial genomes are necessary for growth [19, 23]. The Gumbel method also identified 2670 non-essential genes (Z ‾ i < θ n), with the remainder being labeled as Uncertain (because the posterior probability did not exceed the significance thresholds), or were too short for reliable analysis. Table 1 contains a summary of the classifications obtained by the Gumbel method.

Well-known essential genes like GyrA (DNA gyrase A) and RpoB (DNA-directed RNA-polymerase) were identified as essential by the Gumbel method, both achieving a posterior probability of essentiality of 1.0. Those genes are completely devoid of insertions (aside from a few insertions at the N/C termini). However, one of the strengths of the Gumbel method is that it can also identify genes which contain both essential and non-essential regions, indicative of essential domains. An example of such a gene is Rv3910, which codes for an essential MviN domain [24]. TRANSIT identifies this gene as essential, as it contains a large gap of 32 TA sites in a row without insertions, despite the fact that it has insertions on 10 out of the the remaining 17 sites. DeJesus et al. [11] discusses concordance of these results with previous essentiality analysis using the hybridization-based TraSH method [25].

Another approach to analyzing an individual condition is the Hidden Markov Model. Like before, glycerol replicates were loaded into TRANSIT and the HMM method was run using default parameters.

The HMM analysis classified 16.3% of the sites in the genome as belonging to the “Essential” state, 5.4% belonging to the Growth-Defect state, 77.1% to the Non-Essential state, and 1.2% to the Growth Advantage state (See Table 2). One advantage of this method is that it is not limited to gene-boundaries but instead can assess essentiality of entire regions. For example, the PDIM locus (fadD26, ppsABCDE, mas; which spans ∼ 38kb, 594 TA sites, and 10 genes) is required for virulence in vivo but has high metabolic costs for the organism in vitro and therefore results in a Growth-Advantage for the organism when disrupted. Indeed, sites in this region (e.g. Rv2930-Rv2939) are labeled “GA” (the mean read count in this region is 502.0, a 1.9 fold increase from the global mean), thus identifying that disruption of the PDIM locus when growing on standard in vitro conditions affords an advantage to the organism.

The resampling method can be used for comparative analysis of different growth conditions. After adding glycerol replicates as Control datasets and cholesterol replicates as Experimental datasets, the resampling method was run with default parameters.

A total of 28 genes were identified as differentially essential (adjusted p-value < 0.05; see Table 3 for the number of conditionally essential genes identified). Several of these genes are known to be uniquely required for growth in glycerol or cholesterol. For example, glycerol kinase (GlpK) is necessary for growth on glycerol but should not be necessary for growth on other carbon sources like cholesterol. Indeed, GlpK had a total of 1968 reads in the cholesterol condition, and only 22 total reads in glycerol, achieving an adjusted p-value (or q-value) of 0.0 with the resampling method.

Among those genes identified as necessary for growth in cholesterol were several of the Mce-family of proteins, which is believed to be involved in lipid catabolism [26]. These included Mce4A, Mce4C, Mce4D, and Mce4F, which had 3,453, 1817, 4896, and 5168 reads in glycerol respectively, and 302, 61, 180 and 32 reads in cholesterol.

Several of the genes identified as differentially essential actually contain read-counts that are significantly suppressed in one condition compared to the other, indicating a selection against insertions, hence suggesting a fitness cost to the organism. An example of such a gene is Rv3200c, which contains 167 reads in glycerol and 1,755 in cholesterol, despite having a substantial number of TA sites with insertions in both conditions (7 out of 13 in glycerol and 9 out of 13 in cholesterol), thus showing the gene can tolerate insertions in both conditions. The relative suppression in read-counts alone is enough to achieve a q-value of 0, suggesting that it is conditionally essential.

To illustrate the comparative analysis on datasets from a different organism, TRANSIT was used to perform a comparative analysis of TnSeq datasets of H. influenza. Gawronski et al. [3] compared two datasets of H. influenza grown in vitro and one dataset derived from lung samples after passaging through mice. The libraries were relatively sparse, with 39% mean density in vitro and 26% for the lung dataset. Gawronski et al. identified a total of 136 genes necessary for growth in lung using a combination of the log ratio of read-counts between conditions and the insertion density in vitro. Using TRANSIT’s comparative analysis, 342 genes are obtained using a q < 0.05 cutoff. Out of the 136 genes identified by Gawronski et al., 133 (98%) are identified by TRANSIT as differentially essential.