Arkas: Rapid reproducible RNAseq analysis

Arkas-Quantification Implementation

Arkas is a two-step cloud pipeline. Arkas-Quantification is the first step, which reduces the computational steps required to quantify and annotate large numbers of samples against large catalogs of transcriptomes. Arkas-Quantification calls Kallisto for on-the-fly transcriptome indexing and quantification recursively for numerous sample directories. Kallisto quantifies transcript abundance from input RNAseq reads by using pseudoalignment, which identifies the read-transcript compatibility matrix⁴. The compatibility matrix is formed by counting the number of reads with the matching alignment; the equivalence class matrix has a much smaller dimension compared to matrices formed by transcripts and read coverage. Computational speed is gained by performing the Expectation Maximization (EM) algorithm over a smaller matrix.

For RNAseq projects with many sequenced samples, Arkas-Quantification encapsulates expensive transcript quantification preparatory routines, while uniformly preparing Kallisto execution commands within a versionized environment encouraging reproducible protocols. The quantification step automates the index caching, annotation, and quantification associated while running the Kallisto pseudoaligner integrated within the BaseSpace environment. The first step in the pipeline can process raw reads into transcript and pathway collection results within Illumina’s BaseSpace cloud platform, quantifying against default transcriptomes such as ERCC spike-ins, ENSEMBL non-coding RNA, or cDNA build 88 for both Homo sapiens and Mus musculus; further, the first step supports user uploaded FASTA files for customized analyses. Arkas-Quantification is packaged into a Docker container and is publicly available as a cloud application within BaseSpace.

Arkas-Analysis Implementation

Previous work⁷ has revealed that filtering transcriptomes to exclude lowly-expressed isoforms can improve statistical power, while more-complete transcriptome assemblies improve sensitivity in detecting differential transcript usage. Based on earlier work by Bourgon et al.⁸, we included this type of filtering for both gene- and transcript-level analyses within Arkas-Analysis. The analysis pipeline automates annotations of quantification results, resulting in more accurate interpretation of coding and transcript sequences in both basic and clinical studies by just-in-time annotation and visualization.

Arkas-Analysis integrates quality control analysis for experiments that include Ambion spike-in controls, multiple normalization selections for both coding gene and transcript differential expression analysis, and differential gene-set analysis. If ERCC spike-ins, defined by the External RNA Control Consortium⁹, are detected then Arkas-Analysis will calculate Receiver Operator Characteristic (ROC) plots using 'erccdashboard'¹⁰. The ERCC analysis reports average ERCC Spike amount volume, comparison plots of ERCC volume amount, and normalized ERCC counts (Figure 1).

Figure 1. Arkas-Analysis ERCC spike-in Controls Report.

A) The Receiver Operator Characteristic plot. The X-axis shows the False Positive Rate, the Y-axis shows True Positive Rate. B) and D) show the spike-in total RNA amounts with a linear model fit, and quantified ERCC transcript counts. C) shows a dispersion of mean transcript abundance counts and the estimated dispersion.

Subsequent analyses import the data structure from SummarizedExperiment (Morgan, 2016) and create a sub-class titled KallistoExperiment that preserves the S4 structure and is convenient for handling assays, phenotypic and genomic data. KallistoExperiment includes GenomicRanges¹¹, preserving the ability to handle genomic annotations and alignments, supporting efficient methods for analyzing high-throughput sequencing data. The KallistoExperiment sub-class serves as a general-purpose container for storing feature genomic intervals and pseudoalignment quantification results against a reference genome called by Kallisto. By default KallistoExperiment couples assay data such as the estimated counts, effective length, estimated median absolute deviation, and transcript per million count where each assay data is generated by a Kallisto run; the stored feature data is a GenomicRanges object from¹¹, storing transcript length, GC content, and genomic intervals.

Given a KallistoExperiment containing the Kallisto sample abundances, principal component analysis (PCA) is performed¹² on trimmed mean of M-value (TMM) normalized counts¹³ (Figure 2A). Differential expression (DE) is calculated on the library normalized transcript expression values, and the aggregated transcript bundles of corresponding coding genes using limma/voom linear model¹⁴ (Figure 3A). Further, an additional PCA and DE analysis of both transcripts and coding genes is performed using in-silico normalization using factor analysis¹⁵ (Figure 2B, Figure 3B, Figure 3C). In each DE analysis FDR filtering method is defaulted to 'Benjamini-Hochberg', if there are no resultant DE genes/transcripts the FDR methods is switched to 'none'. Arkas-Analysis consumes the Kallisto data output from Arkas-Quantification, and automates DE analysis using TMM normalization and in-silico normalization on both transcript and coding gene expression in a defaulted two group experimental design, which allows end-users to select the normalization type best suited for their needs.

Figure 2. Arkas-Analysis Normalization Report: Normalization Analysis Using TMM and RUV.

A) TMM normalization is performed on sample data and depicts the sample quantiles on normalized sample expression, PCA plot, and histogram of the adjusted p-values calculated from the DE analysis. Orange is the comparison group and green is the control group. B) A similar analysis is performed with RUV in-silico normalization.

Figure 3. Arkas-Analysis Differential Expression Report: DE using TMM and RUV.

A) DE analysis using TMM normalization. The X-axis is the sample names (test data), the Y-axis are Gene symbols (HUGO). Expression values are plotted in log₁₀ 1+TPM. B) Similar analysis using RUV normalization. C) The design matrix with the RUV adjusted weights. The sample names are test data used in demonstrating the general analysis report output.

Gene set differential expression, which includes gene-gene correlation inflation corrections, is calculated using Qusage¹⁶. Qusage calculates the variance inflation factor, which corrects the inter-gene correlation that results in high type 1 errors using pooled or non-pooled variances between experimental groups. The gene set enrichment is conducted using Reactome pathways constructed using ENSEMBL transcript/gene identifiers (Figure 4 and Table 1); REACTOME gene sets are not as large as other databases, so Arkas-Analysis outputs DE analysis in formats compatible with more exhaustive databases such as Advaita. The DE files are compatible as a custom upload into Advaita iPathway guide, which offers an extensive Gene Ontology (GO) pathway analysis. Pathway enrichment analysis can be performed from the BaseSpace cloud system downstream from parallel differential expression analysis and can integrate with other pathway analysis software tools.

Figure 4. Arkas-Analysis Gene-Set Enrichment Plot.

Gene-Set enrichment output report, each point represents the differential mean activity of each gene-set with 95% confidence intervals. The X-axis are individual gene-sets. The Y-axis is the log₂ fold change.

Data variance between software versions

We wished to show the importance of enforcing matching versions of Kallisto when quantifying transcripts because there is deviation of data between versions. Due to updated versions and improvements of Kallisto software, there obviously exists variation of data between algorithm versions (Figure 5, Supplementary Table 1, Supplementary Table 2). We calculated the standardized mean differences, and the variation of the differences between data output from Kallisto versions 0.43 and 0.43.1 (Supplementary Table 2), and found large variation of differences between raw values generated by differing Kallisto versions, signifying the importance of version analysis of Kallisto results.

Figure 5. Quantile-Quantile Plots of Data Variation Comparing Differences in Kallisto Data from Versions 0.43.1 and 0.43.0.

The X-axis depicts the theoretical quantiles of the standardized mean differences. The Y-axis represents the observed quantiles of standardized mean differences.

The Dockerization of Arkas BaseSpace applications versionizes the Kallisto reference index to enforce that the Kallisto software versions are identical, and further documents the Kallisto version used in every cloud analysis. The enforcement of reference versions and Kallisto software versions prevents errors when comparing experiments.

Operation

Arkas-Quantification instructions are provided within BaseSpace (details for new users can be found here). The input are RNA sequencing samples, which may include SRA imported reads, and the outputs include the Kallisto data, .tar.gz files of the Kallisto sample data, and a report summary. Users may select for species type (Homo sapiens or Mus musculus), optionally correct for read length bias, and optionally select for the generation of pseudoBAMs. More significantly, users have the option to use the default transcriptome (ENSEMBL build 88) or to upload a custom FASTA of their choosing. For users that wish for local analysis, they can download the sample .tar.gz Kallisto files and analyze the data locally.

The Arkas-Analysis instructions are provided within the BaseSpace environment. The input for the analysis app is the Arkas-Quantification sample data, and the output files are separated into corresponding folders. The analysis also depicts figures for each respective analysis (Figure 1–Figure 4) and the images can be downloaded as a HTML format.

Annotation of coding genes and transcripts

The extraction of genomic and functional annotations directly from FASTA contig comments, eliding sometimes-unreliable dependencies on services such as BioMart, are calculated rapidly. The annotations were performed with a run time of 2.336 seconds (Supplementary Table 3) which merged the previous Kallisto data from 5 samples, creating a KallistoExperiment class with feature data containing a GenomicRanges¹¹ object with 213782 ranges and 9 metadata columns. The system runtime for creating a merged KallistoExperiment class for 5 samples was 23.551 seconds (Supplementary Table 4).

Complete transcriptomes enrich annotation information, improving downstream analyses

The choice of catalog, the type of quantification performed, and the methods used to assess differences can profoundly influence the results of sequencing analysis. ENSEMBL reference genomes are provided to GENCODE as a merged database from Havana's manually curated annotations with ENSEMBL's automatic curated coordinates. AceView, UCSC, RefSeq, and GENCODE have approximately twenty thousand protein coding genes, however AceView and GENCODE have a greater number of protein coding transcripts in their databases. RefSeq and UCSC references have less than 60,000 protein coding transcripts, whereas GENCODE has 140,066 protein coding loci. AceView has 160,000 protein coding transcripts, but this database is not manually curated. GENCODE is annotated with special attention given to long non-coding RNAs (lncRNAs) and pseudogenes, improving annotations and coupling automated labeling with manual curating. The database selected for protein coding transcripts can influence the amount of annotation information returned when querying gene/transcript level databases.

Although previously overlooked, lncRNAs have been shown to share features and alternate splice variants with mRNA, revealing that lncRNAs play a central role in metastasis, cell growth and cell invasion¹⁷. LncRNA transcripts have been shown to be functional and are associated with cancer prognosis; proving the importance of studying these transcripts, which are included as defaults within the Arkas pipeline.

Each transcript database is curated at different frequencies with varying amounts RNA entries that influences that mapping rate. GENCODE loci annotations contain 9640 loci, UCSC contain 6056 and RefSeq contain 4888. GENCODE annotations have the greatest number of lncRNA, protein and non-coding transcripts, and highest average transcripts per gene, with 91043 transcripts unique to GENCODE, absent for UCSC and RefSeq databases. ENSEMBL and AceView annotate more genes in comparison to RefSeq and UCSC, and return higher gene and isoform expression labeling improving differential expression analyses¹⁸. ENSEMBL achieves conspicuously higher mapping rates than RefSeq, and has been shown to annotate larger portions of specific genes and transcripts that RefSeq leaves unannotated¹⁸. Although ENSEMBL has been shown to detect the same differentially expressed genes as AceView, ENSEMBL/GENCODE annotations are manually curated and updated more frequently than AceView¹⁸. The choice of transcriptome will definitely influence the power of an analysis, thus Arkas cloud analysis applications use ENSEMBL build 88 (ncRNA, and cDNA) by default for Homo sapiens and Mus musculus and also allow users to upload customized FASTA files.

Docker as a cornerstone of reproducible research

Reproducible research should consistently link the works developed by the research community to unique data environments such as clinical, sequencing and other experimental data, used in the construction of the published work. The aim for transparent research methodologies is to clearly define their association with every research experiment, minimizing opaqueness between findings and methods. For clinical studies, re-generating an experimental environment has a very low success rate¹⁹, which is why non-validated preclinical experiments have spawned the development of best practices for critical experiments. Re-creating a clinical study has many challenges, for example the difficult nature of a disease, the complexity of cell-line models in mouse and human that attempt to capture human tumor environment, and limited power through small enrollments in clinical trials¹⁹. Experimental validation is quite difficult and dependent on the skillful performance of an experiment, and an earnest distribution of the analytic methodology, which should contain most, if not all, raw and resultant data sets.

With recent developments for virtualized operating systems, developing best practices for bioinformatic confirmations of experimental methodologies is much more straightforward than duplicating clinical trials' experimental data. Recent technology advancements such as Docker allow for local software environments to be preserved using a virtual operating system. Docker allows users to build layers of read/write access files, creating a portable operating system which exhaustively controls software versions and data, and systematically preserves the complete software environment. Conserving a researcher's developmental environment advances analytical reproducibility if the workflow is publicly distributed. We suggest a global distributive practice for scholarly publications that regularly includes the virtualized operating system containing all raw analytical data, derived results, and computational software. Currently, Docker, compiled software through CMake, and virtual machines are being utilized, showing progress toward a global distributive practice linking written methodologies, and supplementary data, to the utilized computational environment²⁰.

Comparing Docker as a distributive practice to virtual machines seems roughly equivalent. Distributed virtual machines are easy to download, and the environment allows for re-generating resultant calculations. However, this is limited if the research community advances the basic requirements for written methodologies and begins to adopt a large scale virtualized distribution, converging to an archive of method environments which would make hosting complete virtual machines impractical or impossible. If an archive were constructed where each research article would link to a distributed methods environment, then an archive of virtual machines for the entire research community is impossible. However, an archive of Dockerfiles is more realistic because a Dockerfile consists of only a few bytes in size.

Novel bioinformatic software is often distributed as a cross-platform flexible build process independent of compiler, which reaches Apple, Windows and Linux users. The scope of novel analytical code is not to manage nor preserve computational environments, but to have environment independent source code as transportable executables. Docker, however, does manage operating systems, and the scope for research best practices does include gathering sets of source executables into a single collection of minimum space and maximum flexibility. Docker can provide the ability for the research community to simultaneously advance publication requirements and develop the future computational frameworks in cloud.

Another advantage for using Docker as the machine manifesting the practice of reproducible research methods, is that there is a trend of well-branded organizations such as Illumina's BaseSpace platform, Google Genomics, or SevenBridges (all of which offer bioinformatic computational software structures), to use Docker as the principal framework. Cloud computational environments offer many advantages over local high-performance in-house computer clusters, which systematically structure reproducible methodologies and democratize medical science. Cloud computational ecosystems preserve an entire developmental environment using the Docker infrastructure, improving bioinformatic validation. Containerized cloud applications form part of the global distributive effort and are favorable over local in-house computational pipelines because they offer rapid access to numerous public workflows, easy interfacing to archived read databases, and they accelerate the upholding process of raw data. The Google Genomics Cloud has begun to make first steps with integrating cloud infrastructure with the Broad Institute, whereas Illumina's BaseSpace platform has been hosting novel computational applications since its launch.

Scholarly publications that choose only a written method section passively make validation gestures, which is arguably inadequate in comparison to the rising trend or well-branded organizations. We envision a future where published work will share conserved analytic environments, with cloud software accessed by web-distributed methodologies, and/or large databases organizing multitudes of Dockerfiles with accession numbers, strengthening links between raw sequencing data and reproducible analytical results.

Cloud computational software does not only wish to crystallize research methods into a pristine pool of transparent methodologies, but also matches the rate of production of high quality analytical results to the rate of production of public data, which reaches hundreds of petabytes annually. In a talk given by Dr. Atul Butte in December 2015, he discussed that with endless public data, the traditional method for practicing science has inverted; no longer does a scientist formulate a question and then experimentally measure observations that test the hypothesis. In the modern area, empirical observations are being made at an unbounded rate, the challenge now is formulating the proper question (more details on his talk can be found here). Given a near-infinite amount of observations, what is the phenomena that is being revealed? Cloud computational software can accelerate the production of hypotheses by increasing the flexibility and efficiency of scientific exploration.

Many bioinformaticians have noted a rising trend in biotechnology, predicting that open data and open cloud centers will help democratize research efforts and create a more inclusive practice. With the presence of cloud interfacing applications such as Illumina's BaseSpace Command Line Interface, DNA-Nexus, SevenBridges, and Google Genomics becoming more popular, cloud environments pioneer the effort for achieving standardized bioinformatic protocols.

Democratization of big-data efforts has some possible negative consequences. Accessing, networking, and integrating software applications for distributing data as a public effort requires massive amounts of specialized technicians to maintain and develop cloud centers that many research institutions are migrating toward. Currently, it is fairly common for research centers to employ high-performance computer clusters which store laboratory software and data locally; cloud computing clusters are beginning to offer clear advantages compared to local closed computer clusters. Collaborations are becoming more common practice for large research efforts, and sequencing databases have been distributing data globally, making cloud storage more efficient. This implies that services from cloud centers will most likely be offered by very few elite organizations because the large scale of cloud services will prevent incentives for smaller companies.

It is very likely that only a few elite organizations will provide services to cloud computing environments, acting as a gateway which directs the global research community toward a narrow set of well established, standardized, computational applications. With regard to recent changes relating to media consumption and e-commerce, democratization allows independent alternative selections far greater exposure, equalizing profits for lower ranked selections “at the tail", however it may be possible that the abundant amount of data distributed over storage archives, which stimulates an economically abundant environment, could shift into a fiercely controlled economic environment of scarcity. For example, if a gold-standard is reached for computational applications, the range of alternative selections could remain non-existent, which may diminish the future roles of bioinformaticians. This possible scenario suggests bioinformaticians could be re-directed to small garages instead of the technocratic places such as Silicon Valley, motivated not from a spirit of entrepreneurialism, but from a lack of funding.

Automative downstream analyses is not without its drawbacks; most computational software is highly specialized for niche groups with a mathematical framework constructed by specialized assumptions, this may require a diverse array of computational developments, and thus a large community of developers. The automation of analytical results seems almost unavoidable, and the benefits seem to outweigh the negative consequences.

Data Used in Testing Variation between Versions

Controls: SRR1544480 Immortal-1

SRR1544481 Immortal-2

SRR1544482 Immortal-3

Comparison: SRR1544501 Qui-1

SRR1544502 Qui-2

Software availability

Latest source code:

https://github.com/RamsinghLab/Arkas-RNASeq

Archived source code as at the time of publication:

DOI: 10.5281/zenodo.545654²¹

License:

MIT license

Reference FASTA Annotation Files

For Homo-sapiens and Mus-musculus ENSEMBL FASTA files were downloaded here for release 88.

ERCC Sequences

The ERCC sequences are provided in a SQL database format located here