The first process in DNA-Seq data analysis involves alignment and assembly. Alignment is the process of mapping DNA-Seq reads to a reference genome. Many sequence alignment software tools that are available today use two main algorithms: the hash-based and the Burrows-Wheeler Transform methods.

Some hash-based algorithms build their hash table on the set of input reads (MAQ [6], Illumina’s ELAND unpublished algorithm, SHRiMP [7], ZOOM [8]). Another set of tools build their hash table on the reference genome (SOAPv2 [9], BFAST, http://genome.ucla.edu/bfast/, MOSAIK http://bioinformatics.bc.edu/marthlab/Mosaik/, Novoalign http://www.novocraft.com/main/index.php, PERM [10]). After building the hash-table these methods can either use the reference genome to scan the hash table of input reads, or use the set of input reads to scan the hash table of the reference genome.

Many recent algorithms rely on the theory of string matching using Burrows-Wheeler Transform (BWT). BWT algorithms (BOWTIE [11], BWA [12], SOAPv2 [9]) typically create a suffix array from the BWT transformed sequence, rather than from the original sequence. In the first step, the sequence order of the reference genome is modified using the BWT, a reversible process (i.e., the original genome sequence can be reconstructed backwards) that reorders the genome grouping together in the data structure the sequences that appear multiple times. Next, the final index is created and is used for rapid read placement on the genome. The main advantage of BWT algorithms is their speed, as they are much faster than hash-based algorithms at the same sensitivity level [3].

Assembly starts from aligned DNA-Seq reads to reconstruct the original DNA sequence computationally, which generates large, continuous regions of DNA sequence [3]. Many alignment software provide tools to perform the assembly after the read alignment (e.g., MAQ), or standalone resources can be used (SAMTOOLS [13], Emboss [14]) or commercial packages like Geneious (http://www.geneious.com) and CLC-Bio (http://www.clcbio.com). For organisms without a sequenced reference genome, it is not possible to perform any reference genome guided assembly of the reads, thus de novo assembly is always an essential step for data analysis. The majority of de novo assemblers that have been released follow two basic approaches: overlap graphs [15] and de Bruijn graphs [16]. The overlap graph method calculates all the pair-wise overlaps between the reads and report this information in a graph. The manipulation of the same overlap graph leads to a layout of reads and then to a consensus sequence of contigs using Celera Assembler [17] or Arachne [18] among others. This traditional approach is computationally intensive as the overlap graph is extremely large even for simple organisms. De Bruijn graphs algorithm is used by most assemblers (Velvet [19], SOAPdeNOVO [20], ABySS [21]) and reduces the computational charge by breaking reads into smaller sub-sequences of DNA, called k-mers, where the k parameter describes the length in bases of these sequences [22]. The de novo assembly can be used also to resolve complex genomic region (e.g., rapidly evolving or rich in repetitive elements) of organisms with a reference genome. In this case the contigs are aligned back to the reference genome and can undergo all the next analytical steps here described.

There are many issues that must be considered when dealing with NGS data, beginning with the alignment of short reads. As an example, since each read is aligned independently, many reads spanning Indels may be misaligned. The per-base quality scores (i.e., the probability that the called base in the read is the true base [23]) may also be inaccurate due to systematic errors in sequencing technology, machine cycle, and sequence context [24,25,26]. Thus, following the alignment and/or assembly of reads, quality control steps are implemented before continuing in the downstream analyses (in Section 2.1).

Once the reads have been aligned and calibrated, SNPs, Indels and CNVs can be called. This step requires sensitive and specific statistical models and tools [6,9,12,13,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48], named in Table 1.

Additional software such as PLINKseq (http://atgu.mgh.harvard.edu/plinkseq/) implement statistical models to analyze variants called from NGS experiments, testing for association with continuous or dichotomous traits and assessing an unusual distribution for rare variation across different functional categories [52]. Some other tools like PolyPhen2 [53] and VAAST [54] can be used afterwards for functional variant annotation and prioritization providing hints on the biology and pathophysiology of psychiatric disorders. Also, alternative annotation tools and strategies have been proposed [55] to retrieve information and test hypotheses about the functional role of variants present by chance in any single human genome or enriched in the genome of people affected by a psychiatric disease [56,57].

The development and release of algorithms and software for analysis of NGS data has seen exponential growth in the last two years, requiring a huge investment in terms of time, expertise and computer infrastructure.

From this brief review of tools (Table 1), it is evident that analyzing NGS data is a challenging and time consuming operation for scientists. Ideally, these tools must be up to date and easy to use, and their sequential combination should optimize performance and accuracy, with each program producing output files compatible with the input requirements of the software performing the following operation. Such processes sequentially linked together build what is generally called a workflow. An increasingly large number of workflows are available today to manage high-throughput genomics sequencing data, from basic data processing to high-quality visualization of results. Examples include shell-scripts [58,59], tool-specific graphical interfaces [60,61], and graphical workflow environments [62,63,64,65]. The graphical workflow environment are emerging as useful for constructing, modifying, interconnecting and executing computational genomics protocols using data processing workflows, also described as “pipelines” once the processes have been connected (Table 2).

There are several additional features in graphical workflow environments that simplify the data management. Table 2 lists some of the commonly used workflow environments and compares their core features. Each of the graphical environments described in Table 2 allows design and submission of new workflows.

The most commonly used systems in NGS analysis are Taverna [68] and Galaxy [64]. The latest beta-version of the Galaxy (http://galaxy.psu.edu/) platform offers a NGS computational framework that embeds single processing units to be invoked as web-services but it still lacks the functionality of interlinking the outputs of one process into a subsequent module. The available Taverna and Galaxy services have restrictive upper limits on storage and per-process RAM, when they are deployed on web-services, creating bottlenecks with data staging to/from the servers and computational runs. Also, the library of available Galaxy routines doesn’t allow adding new tools and is limited to a few alignment software tools (e.g., BWA and Bowtie), or on the quality control side, an incomplete beta-version of the PICARD suite.

The LONI (Laboratory of Neuro Imaging) pipeline displays unique features for the implementation of new interactive and robust NGS analysis workflow protocols using a graphical environment. The design of novel workflow solutions within the LONI pipeline environment is simple and flexible [69], and allows embedding and connecting heterogeneous software within the same computational protocol without the need of advanced bioinformatics skills (Table 2).

The LONI pipeline architecture [70,71] is a distributed environment utilizing a client-server interface for design, validation, execution and dissemination of computational protocols as graphical workflows. Individual applications are represented as modules and may be linked to form complex network implementation of the desired analytical processes. Using a flexible, user friendly and customizable data processing and visualization system, the LONI pipeline environment provides access to distributed datasets, heterogeneous software tools and diverse web-services. Additional details about the LONI pipeline environment are available in [65].

Based upon the review of the NGS software (Table 1) and the graphical workflow environments to manage workflow (Table 2), we chose to develop within the LONI environment a graphical environment for genomics, called the Graphical Pipeline for Computational Genomics (GPCG) that covers many informatics analytical steps on NGS data. This effort was a joint collaboration between LONI at UCLA, BIRN (Biomedical Informatics Research Network) at UCI, Information Sciences Institute (ISI) at USC.

The GPCG is a set of workflows that may simplify and speed-up approaches for sequencing projects performed on Illumina/Solexa Genome Analyzer-HiSeq platform [72,73,74]. We have implemented workflows in GPCG that: (1.1) aligns reads (both in single and paired end) and (1.2) performs de novo assembly with multiple algorithms; (1.3a) performs basic formatting and quality control, followed by a (1.3b) more advanced and complex quality control to correct for sequencing biases; (2.1a) performs SNP-Indels calling and annotation, and (2.1b) CNVs calling. The Supplementary Materials S1 provides detailed information about our experiments and results.

We incorporated into the LONI pipeline some of the most used alignment software tools (Figure 3). Through the “edit module” functionality of the graphical user interface (GUI) it is possible to visually manage all the parameters of the alignment software (e.g., number of allowed mismatches, read trimming, gap extension) encapsulated by the module itself. Each algorithm in a workflow (other than BWA-SW that runs only in single end) has a switch that allows the user to perform an alignment on either single end (SE) or paired end (PE) reads. The input reads are generally in FASTQ format. BWA accepts input read sequences also in BAM format. For paired-end data the pair reads need to be grouped together in one single BAM file. The final output is an alignment file is in SAM format (see Supplementary Materials S1, Figure A2). This file is ready to undergo quality control procedures. The choice of the alignment software is critical as different software might lead to different results in term of mapping and downstream variants calling. It is strongly suggested to perform the alignment using more than one software and identify strategies that may provide better fit for the user’s data. BWA and BOWTIE, for example, are among the fastest alignment algorithms for both single and paired end reads, and are particularly suitable for whole genome alignments. They generally report by only one alignment hit, and in the default mode do not guarantee finding the best hit nor if the found hit is unique. In particular, for users dealing with highly repetitive genomic regions, it is possible to force these algorithms to output all the possible hits with a high cost in terms of speed, or to use a probabilistic software like PERM that outputs all the found hits. The user could perform a de novo assembly (see next paragraph) for a better resolution in such regions. If the main interest is the detection of CNVs with paired-end reads, Novoalign is strongly recommended because of the high mapping accuracy, even if slower than the previously mentioned Burrows-Wheeler (BW) based software.

For selecting parameters settings, users can find hints about the meaning and possible values of each single parameters by double-clicking on a parameter icon or via the execution info menu-accessible as a right-click on the module of interest. For the parameters tuning the users have to refer to the software documentation as all the algorithms works differently. The LONI Pipeline workflow environment enhances the user’s abilities to optimize the selection of the parameters for each tool by allowing specification of a list containing a range of parameters spanning the support of each parameter of interest. This parameter-optimization strategy is exactly the basis of the development of meta-algorithms as workflow protocols [76,77,78].

The quality control process we used to determine the tools to be incorporated is based on an up to date and detailed investigation of the available NGS tools, including research papers with technical and analytical details, and the online discussions panel/forums if available to better follow the analytical trends and open issues. The testing with simulated and real data provided us further hints into the software reliability. Some information and guidelines to help in identifying the best set of software to analyze different type of data can be found in some about reference with comparisons among tools we provided [22,79].

When a reference genome is not available for the species under investigation it is necessary to perform de novo assembly of the reads. De novo assembly is also the suggested approach for reads mapping in regions prone to rearrangements, rapidly evolving, or where the reference genome might not be informative. We have developed three workflows embedding the most common de novo assembly software based on de Bruijn graphs (Figure 4), which works both with single and paired end data. One of the most important parameters to be tuned in de novo assembly is the fixed-length of the subsequences used to build the graphs length, which is abbreviated as k-mer. The most efficient k-mer size for a particular assembly is determined by the read length as well as the error rate, and it may be estimated a priori using “ad hoc” tools provided by the software developer or by evaluating the results obtained through varying the k-mer size across different runs. Some assemblers are also able to output scaffolds (i.e., a set of contigs with known relative orientation and distance) or supercontigs (i.e., contigs in which gaps are allowed). When de novo assembly is used in complex region of organisms with a known reference genome, the contigs can be aligned back to it with the BWA-SW workflow, and then undergo the subsequent analytical steps.

We evaluated all the previously described GPCG workflows (Table 2 and Supplementary Materials S2) with 30 million simulated reads (both SE and PE) generated with dwgsim simulation tool, a publically available utility for simulating whole-genome Illumina reads [83]. Runtimes and information about the performances of the workflow with simulated data are summarized in Supplementary Table S2. We ran all the workflows with default parameters.

We reported for all the modules a better time performance of the GPCG compared to Galaxy, together with the absence for GPCG of the time required for the data upload.

Using the same simulated data files, we tested the common modules embedded both in our GPCG and on the Galaxy webserver interface. Due to the structure of Galaxy and the reduced number of processes available, we couldn’t compare its performances on a workflow scale. The module shared by GPCG and Galaxy were: conversion of solexa into sanger format, BWA and Bowtie paired end alignment, the PICARD utilities to fix the mate information, mark the duplicates, collect the alignment, GC bias, and insert size metrics. We reported the GPCG and Galaxy time performances in Table 4.

All of the performances of the GPCG platform were better compared to the run time of Galaxy. Furthermore, a bottleneck in the timing of analysis is the upload of to the Galaxy webserver, together with the not predictable waiting in queue time for the processes (Table 4). As the status of the process cannot be checked during the execution, the user must wait the end of the run of a process to input the results in to the next step without any hint about the elapsed time, and without the possibility to automatically keep track of the run times.

We also evaluated the performance of our GPCG workflows in a “real life scenario”. Supplementary Table S2 reports the performances of the whole-genome alignment with BWA of an entire Illumina flowcell with an average of 130 million Illumina PE reads per lane. The average time required for the alignment of one lane was 6 hours, and for the whole flowcell was roughly 48 hours using GPCG. The output of the alignment was one SAM alignment file for each lane. In all large scale NGS data analysis it is important to manage the flow of both data and input/output files as they traverse complex workflows. The GPCG provides a simple way to set up the input file section (see magnified data source panel on the right, Figure 9), manage the naming of output files from individual modules using the transformation tool within a module definition (Figure 9).

To translate DNA-Seq data analysis into a graphical pipeline solution within the LONI environment [65] we initially described all global processes in the protocol design or skeletonization step, using a top-down approach. We outlined the general classes of sequence data analysis, then the appropriate sub-classes of analyses, specific tools, test-data, invocation of concrete tools, and a detailed example of executable syntax for each step (http://pipeline.loni.ucla.edu/support/user-guide/building-a-workflow, [65]). The previously mentioned analytical steps are constructed from a series of command line executable processes, referred to as modules or nodes, which are connected to each other to form a visual workflow analysis protocol. All the logically concatenated modules/nodes involved in the same analytical step (e.g., alignment) comprise a workflow.

After all necessary modules that make up a workflow are independently defined and validated through the LONI pipeline GUI interface (http://pipeline.loni.ucla.edu/support/user-guide/creating-modules/), they are integrated into a coherent workflow.

The user can access the current version of the GPCG online at http://pipeline.loni.ucla.edu. The local Client can be set up by the users following the instructions reported at http://pipeline.loni.ucla.edu/support/user-guide/installation/. The GPCG can also be deployed on a server base downloading the distributed pipeline server installer (DPS) (http://pipeline.loni.ucla.edu/support/server-guide/installation/), and following the instructions available at http://pipeline.loni.ucla.edu/DPS). The GPCG workflows can also be directly launched via the pipeline web-start server (PWS) (http://pipeline.loni.ucla.edu/PWS). To search across the entire set of workflows the user can rely on an interactive graphical navigator (http://pipeline.loni.ucla.edu/services/library-navigator/), which enables not only the discovery, but also web-based utilization of this new computational-genomics infrastructure through the web, Figure 10.

Table 1 presents the software we have embedded in the workflows released with GPCG. We evaluated the workflows with both real and simulated data (Supplementary Table S2).

Evaluation with simulated data We used dwgsim [84], a utility for whole-genome Illumina reads simulation, contained in DNAA v0.1.2 (http://sourceforge.net/projects/dnaa/), to generate Illumina-like short sequences, using the default empirical error model illustrated on DNAA’s Whole-Genome Simulation web-site (http://dnaa.sf.net). In total we generated 30 million reads with 100 bp length, using the complete human genome (hg18) as a reference and with default parameters. We developed also a module that allows the users to generate simulation datasets according to their needs (see Table 2 and Supplementary Materials S2).

Evaluation with real data To further compare the behavior of our workflows on real applications, we used an entire flow cell (8 lanes) with an average of 130 million Illumina PE reads per lane with length of 100 bp (fastq produced with the Illumina Pipeline v1.8) to be aligned by BWA-PE against the whole human genome sequences (assembly: NCBI36.1/hg18).

These experiments were produced on an Illumina HiSeq200 DNA sequencer using v5 kits.

All the testing runs were performed on a Linux server Dual Intel Xeon X5650 2.67 Ghz 6 core processors 96 GB’s of RAM 2 × 10 k RPM 150 GB WD VelociRaptor RAID 0 RHEL 5 64-bit.