Genes of the pig, Sus scrofa, reconstructed with EvidentialGene

Introduction

Precision genomics is essential in medicine, environmental health, sustainable agriculture, and research in biological sciences (Goldfeder et al., 2016). Yet the popular genome informatics methods lag behind the high levels of accuracy and completeness in gene construction that is attainable with today’s accurate RNA-seq data.

To demonstrate the accuracy and completeness of gene set reconstruction from expressed gene pieces (RNA-seq) alone, excluding chromosome DNA or other species genes, the pig is a good choice. The pig has well-constructed, partly curated gene sets produced by the major genomics centers NCBI and Ensembl, and is one of seven RefSeq top-level model organisms. Gene sets are based on extensive expressed sequences dating from the 1990s. Pig has a well-assembled chromosome set (Groenen et al., 2012), improved in 2017, and contributions of experimental gene evidence from many projects of agricultural and biomedical focus. Published RNA-seq from over 2,000 samples puts the pig among the top 10 of model animals and plants. Yet there is just one public transcript assembly from these many pig studies, from blood samples only.

If successful, this demonstration can be used to improve reference genes for this species. It will demonstrate to others how to produce reliably accurate gene sets. Unreliability in gene sets is a continuing problem, measurable from missing and fragment orthologs (Trachana et al., 2011; Tekaia, 2016; Simao et al., 2015; Waterhouse et al., 2018). Reasons for unreliability are many, including errors from sequencing and assembly, mis-modeling of complex genes, and errors propagated from public databases. The EvidentialGene project aims for a reliable solution that others can use, simple in concept, to obtain accurate gene sets from a puzzle box full of gene pieces.

Gene sets reconstructed by the author are more accurate by objective measures of homology and expression recovery, than those of the same species produced by popular methods. EvidentialGene has been developed in conjunction with reconstruction of gene sets for popular and model animals and plants such as arabidopsis, corn plant, chocolate tree, zebrafish, atlantic killifish, mosquitos, jewel wasp, and water fleas (Gilbert, 2012, 2013, 2016, 2017). These are more accurate than the same species sets produced by NCBI Eukaryotic Genome Annotation Pipeline (EGAP, Thibaud-Nissen et al., 2013), Ensembl gene annotation pipeline (Curwen et al., 2004), MAKER genome-gene modeling (Holt & Yandell, 2011), Trinity RNA assembly (Grabherr et al., 2011), Pacific Biosciences long RNA assemblies, and others.

Gene sets reconstructed by others using EvidentialGene methods are also more accurate (Nakasugi et al., 2014, Mamrot et al., 2017), in independent assessments. However, some investigators do not apply necessary details of EvidentialGene methodology, or modify portions in ways that reduce accuracy. One impetus for this work is engineering a full, automated pipeline that others can more readily use, for these validated methods. Herein is reported a good quality gene set published in databases to aid future research and database improvements. The automated pipeline of SRA2Genes is introduced here with accurate genes for the pig model. EvidentialGene methodology will be detailed in other papers.

Materials and Methods

Materials for this pig gene set reconstruction are primarily four RNA source projects that were selected from over 100 in the SRA database, based on tissue sampling, methodology, and other factors. These four and gene evidence data materials are listed below in “Data and Software Citations.” The NCBI SRA web query system allows search and retrieval by species and attributes including tissue, sex, sample type, and instrument type. Queries used to select these materials are listed in Data Citations. Evidence data sets for validation of this pig gene set include reference proteomes of human, mouse, cow and zebrafish, vertebrate conserved single copy gene proteins, and pig chromosome assembly. Data sets for comparison include recent NCBI and Ensembl pig genes, and expressed pig sequences (see Data Citations).

The selected RNA projects are pig1a (PRJNA416432, China Agricultural University), pig2b (PRJNA353772, Iowa State University, USDA-ARS), pig3c (PRJEB8784, Univ. Illinois), and pig4e (PRJNA255281, Jiangxi Agricultural University, Nanchang, China). These four include 26 read sets of 1,157,824,292 read pairs, or 106,654 megabases. All these are paired-end reads from recent-model Illumina sequencers, ranging from 75 to 150 bp read length. Pig3c includes adult female and male tissues of muscle, liver, spleen, heart, lung, and kidney. Pig1a includes adult female tissues of two sample types. Pig2b includes tissue samples of brain, liver, pituitary, intestine, and others. Pig4e provides embryonic tissue RNA. Notably missing were head sensory organs, one result being that some eye, ear, nose and taste receptor genes are under-represented or fragmented in this reconstruction.

Selection of these four projects was done with the objective of collecting all transcribed genes of the pig, within constraints on effort, using the author’s experience in reading RNA database information on samples. A full range of tissues, sex, development stages of samples and accuracy of sequencing instruments are important criteria for transcript reconstruction. Collecting public RNA samples that include all expressible genes requires some trial and error, or very large computational effort for well-studied organisms. Of the many RNA source projects for pig, most samples are for specific tissues, often for mutant strains, with limited sample documentation, and for less accurate sequencer instruments.

EvidentialGene methods use several gene modeling and assembly components, annotates their results with evidence, then classifies and reduces this over-assembly to a set of loci that best recovers the gene evidence. Each component method has qualities that others lack, and produces models with better gene evidence recovery. Gene reconstruction steps are (Gilbert, 2012): 1. produce several predictions and transcript assembly sets with quality models; 2. annotate models with available gene evidence (transcript introns, exons, protein homology, transposon, and other); 3. score models with weighted sum of evidence; 4. remove models below minimum evidence score; 5. select from overlapped models at each locus the highest score, and alternate isoforms, including fusion metrics (longest is not always best); 6. evaluate resulting best gene set (i.e., compare to other sets, examine unrecovered gene evidence); 7. re-iterate the above steps with alternate scoring to refine. Evidence criteria for genes are, in part, protein homology, coding/non-coding ratio, RNA read coverage, RNA intron recovery, and transcript assembly equivalence.

For RNA-only assembly, this paradigm is refined at step 2–4 to introduce a coding-sequence classifier (Gilbert, 2013), which reduces large over-assembly sets (e.g., 10 million models of 100,000 biological transcripts) efficiently, using only the self-referential evidence of coding sequence metrics (protein length and completeness, UTR excess).

CDS overlap by self-alignment identifies putative gene loci and their alternate transcripts, similarly to how CDS overlap by alignment to chromosomal DNA is used in traditional genome-gene modeling to classify loci. This CDS classifier, in tr2aacds.pl pipeline script, uses the observed high correlation between protein completeness and homology completeness, making a computationally efficient classifier that will reduce the large over-assembly set to one small enough that the additional evidence classifications are feasible to refine this rough gene set to a finished one, using evidence of protein homology, expression validity, chromosomal alignment, and others.

An automated pipeline, SRA2Genes, with methods outlined above, is used for this pig gene reconstruction. It includes RNA-seq data fetching from NCBI SRA, over-assembly of these data by several methods and parameters, transcript assembly reduction with coding-sequence classifier, protein homology measurement, sequencing vector and contamination screening, gene annotation to publication quality sequences, and preparation for submission to transcript shotgun assembly archive (TSA). Each project RNA data table is returned by SRA web query, as srapig-PRJNA416432.csv for pig1a. The program is run with this data table, as for pig1a, to fetch data and produce cluster batch scripts for compute steps: “$evigene/scripts/evgpipe_sra2genes.pl -SRAtable srapigPRJNA416432.csv -runname=pig1a -NCPU=8 -log.” Each of the four projects’ RNA set was assembled this way, to the step of non-redundant gene set with alternate isoforms. Then a single gene set is produced, by combining these four reduced assemblies as input transcripts to SRA2Genes. Supplemental Archive 4 contains scripts used for this. Details of the components, their options and configurations for gene reconstruction that are used in this automated pipeline are contained in the public release of software and data (see ‘Results’), as built during development with many animal and plant gene sets.

Assemblers used for all four RNA samples are Velvet/Oases (Schulz et al., 2012), idba_tran (Peng et al., 2013), SOAPDenovoTrans (Xie et al., 2013), and Trinity (Grabherr et al., 2011). K-mer sizes are computed to span the read sizes in, usually, 10 steps. As observed in this and other RNA assembly studies, k-mer of 1/2 read size produces the single most complete set, however, most k-mer sizes produce some better gene assemblies, due to wide variation in expression levels and other factors (Zhao et al., 2011; Peng et al., 2013; Bankevich et al., 2012). Strongly expressed, long genes tend to assemble well with large k-mer. Both non-normalized and digitally normalized (Crusoe et al., 2015) RNA sets were used; each method produces a somewhat different set of accurate genes.

To create a single species gene set, secondary runs of SRA2Genes are performed, starting from the combined transcript set of assembly/reductions on four samples. Several intermediate runs were used in this way, assessing gene set completeness. For instance, pig4e sample was added after the merging of pig1a, 2b, 3c samples (all adult tissues) found a deficit for embryonic genes. After merging four project over-assemblies, reference protein assessment identified fragment models. These were targeted for further assembly with rnaSPAdes (Bankevich et al., 2012) to finish fragment assemblies. Prior work with several methods of assembly extension has proven to be unreliable, including assemblers Oases, SOAP and idba. These typically extend fragments by sequence overlap alone, but rarely produce longer coding sequences, instead indel errors, and fused genes are frequent artifacts. rnaSPAdes, unlike the others, uses a graph of paired reads to extend partial transcripts, and may prove more reliable. Five merges of all four samples were done, with addition of new transcript assemblies to replace missing or fragmented models.

Results

Discussion

The main result of this demonstration compared with the NCBI RefSeq pig gene set is, on average, they are equally valid by homology measures, but differ at many gene loci, with Evigene adding many alternate transcripts. The Evigene set also retains more putative loci, lacking measured homology but with other evidence, that further study will clarify their value. Improvements to the pig gene set are numerous enough to warrant updating RefSeq with those from this work. These include 1,500 missing or poorly modeled genes with homology to human, and improved vertebrate conserved genes. Between RefSeq and Evigene sets, all highly conserved vertebrate genes of the BUSCO set exist in pig. Another 3,000 improvements are mostly alternate transcripts with greater alignment to other species, by changes in an exon or two.

This Evigene set has demonstrated objectively accurate gene assemblies that improve the reference gene set of the pig model organism. It has been submitted for that purpose to NCBI as a third party annotation/assembly (TPA) of a TSA, which are International Nucleotide Sequence Database Collaboration (INSDC) classifications. There are policy reasons to limit inferential or computational TPA entries, and there are also policy reasons to accept these. On one hand, objectively accurate gene and chromosome assemblies of experimental RNA and DNA fragments are the desired contents of public sequence databases. On the other hand, having many assemblies of the same RNA or DNA fragments is confusing and could overwhelm databases devoted to experimentally derived genome sequences. This pig gene set adheres to the described policy of TPA in that (a) it is assembled from primary data already represented in the INSDC databases (SRA sequence read section); (b) it is indirectly experimentally supported by reference gene homology measures; (c) it is published in a peer-reviewed scientific journal. Additionally this gene set provides thousands of improvements to the reference gene set. The author produced no wet-lab experimental evidence, but has assembled gene sequence evidence from several sources into a gene set that substantially improves upon NCBI EGAP and Ensembl gene sets. Review of this data set, by NCBI and independent peers, weighs the above dilemma: improve public genome sequences or limit independent computational assemblies.

Animal gene set reconstructions by Ensembl and NCBI RefSeq are widely used and considered high quality, though recent published comparisons of these two are uncommon (for the special case of human genes, see Zhao & Zhang, 2015). This author often compares NCBI and Ensembl genes when constructing Evigene models, observing that NCBI methods now commonly surpass those of Ensembl, as is found in these pig gene sets. Table 4 indicates this for reconstruction of conserved genes in five model animals, all with errors above the human set that are correctable (e.g., for pig NCBI+Evigene). The NCBI improvement is likely due in part to NCBI EGAP’s more extensive use of RNA-seq and transcript assemblies, and use of transcript evidence-based exceptions to a rule that gene models must align fully to chromosome assemblies. Neither NCBI EGAP nor Ensembl produce de novo assemblies of RNA-seq. These projects, however, can and do use assembled transcripts from INSDC public databases. Evigene’s de novo assembled genes can thus improve these other widely used gene sets.

Combining and selecting by evidence criteria the assemblies of several methods improves gene reconstruction to a higher level of accuracy. The individual methods return from 77% (Velvet) down to 50% (Trinity) of the best gene models, and a hybrid PacBio+Illumina assembly is intermediate at 66%. K-mer sizes are an important parameter, as noted by others: “smaller values of k collapse more repeats together, making the assembly graph more tangled. Larger values of k may fail to detect overlaps between reads, particularly in low coverage regions, making the graph more fragmented” (SPAdes, Bankevich et al., 2012). Alternate isoforms of each gene, which share exons and differ in expression levels, are more accurately distinguished from other genes at large k-mer sizes (idba_tran, Peng et al., 2013). These results are consistent with multi-method reconstructions for arabidopsis, corn, zebrafish, mosquitos, and water fleas.

The main flaw in this Evigene pig set is incomplete reconstruction of many genes, especially longer ones. While this is not always a problem with RNA-only assemblies, it is a common one. Importantly, there does not appear to be a reliable method for improving gene assemblies identified as fragmentary, using de novo RNA assembly. While there are several methods that attempt to address this, those tested by the author are unreliable. A trial of rnaSPAdes to extend fragments did improve some genes, but not as many as the RNA data warrants.

A second flaw in EvidentialGene’s method of classifying loci from self-referential alignment of coding sequences is that some paralogs are confused as alternate transcripts of the same locus. With high sequence identity, paralogs align to each other similarly to transcripts of one locus (a class termed “altpar” or “paralt”), though with mismatches that chromosome alignment can resolve. This has been measured at a rate of about 5% for reference gene sets of mouse and zebrafish, and 3% for arabidopsis; a smaller 0.5% portion of alternates at one locus are misclassified as paralog loci. Several de novo gene assembly methods that classify loci have similar altpar confusion, as RNA-seq reads are often shared among paralogs as well as alternate transcripts. These altpar transcripts have not been resolved for this pig gene set, though it is an improvement in development.

This demonstration excluded the use of chromosomes and other species genes to assemble or extend assemblies. Both methods can be employed to advantage to reconstruct genes, where there are few errors in these additional evidences. An important reason to limit initial gene reconstruction to RNA-only assembly is to avoid compounding errors from several sources. This limited-palette reconstruction is validated with independent evidence from genomic DNA and other species sources; genes identified as mis-assembled, or missing, in such RNA-only sets can be improved with these other methods. Many discrepancies between RNA-only reconstruction and the other evidences are from flaws in chromosome assemblies or other species genes that can be identified with careful evaluations.

Gene transcripts from any source, such as EST and PacBio, may be added into SRA2Genes pipeline. Excluded from this reconstruction are the extensive public set of pig ESTs, and the PacBio+Illumina assembly from the same study as pig2b. These contribute a small number of improved transcripts not in this EvidentialGene set (8 missed human orthologs in ESTs, 12 EST, and 24 PacBio with significant improvements), and are used in the RefSeq set. However, as these are already in the public databases, this demonstration reconstruction adds no value to them.

While these gene data and paper were in review at repositories, Zhao et al. (2018) pre-published a reconstruction of pig genes, with newly sampled proteomic and transcriptomic sequences. The authors provide public access to these under BioProject PRJNA392949 for SRA RNA-Seq, and a bioRxiv preprint with sequences of 3,703 novel protein isoforms. The experimental design of this work is well suited to gene set reconstruction, as it sampled 34 tissues of adult male, female and juvenile pigs. Unlike the samples winnowed from prior SRA entries by this author, each from a pig portion, this new work is comprehensive in collecting expressed and translated genes.

This pre-published gene set is compared to the same RefSeq gene set and chromosome assembly as this paper. In brief, of the 3,700 novel proteins, most align to other gene sets and chromosome assembly: 74% are contained in this paper’s transcripts, 65% are contained in the RefSeq transcript set, and 61% are contained in the pig chromosome set, at 75% or greater alignment (protein to RNA/DNA aligned with tBLASTn). Nonetheless, most of these novel proteins do not have a protein equivalent in the gene sets: about 800 novel proteins align to Evigene proteins, and about 600 to RefSeq proteins. A main difference here lies in measures from RNA to protein, including new alternate transcripts and discrepancies in RNA to protein reconstruction, rather than in newly identified gene loci, and is beyond scope of this note to resolve. A rough draft with SRA2Genes of this recent RNA-Seq, assembling only well-expressed genes, contains about the same 74% of novel proteins as for this paper’s set. An application suited to SRA2Genes is to update with these completely sampled pig genes, including depositing an improved version to Transcriptome Shotgun Assembly public database for further uses.

Conclusions

The SRA2Genes pipeline is demonstrated, for the pig model organism, as a reliable gene reconstruction method, useful to other projects and for improving public reference gene sets. The resulting complete transcriptome assembly of pig fills a void at public repositories. Reconstruction from RNA only provides independent gene evidence, free of errors and biases from chromosome assemblies and other species gene sets. Not only are the easy, well known ortholog genes reconstructed well, but harder gene problems of alternate transcripts, paralogs, and complex structured genes are usually more complete with EvidentialGene methods.

Data and Software Citations

NCBI pig gene set used in comparison, from ftp://ftp.ncbi.nlm.nih.gov/refseq/S_scrofa/mRNA_Prot/pig.1.rna.gbff.gz, accessed on 27 Apr 2018.

Ensembl pig gene set used in comparison, from ftp://ftp.ensembl.org/pub/release-93/sus_scrofa/pep/Sus_scrofa.Sscrofa11.1.pep.all.fa.gz, accessed on 28 Jul 2018.

NCBI RefSeq pig chromosome assembly Sscrofa11.1, accession: GCF_000003025.6, dated 2017-2-7, is used for chromosome mapping.

NCBI RefSeq gene sets used as reference genes are H_sapiens, M_musculus, B_taurus, and D_rerio, accessed at same location and date as pig genes. Ensembl gene sets for comparison of same species and date are from ftp://ftp.ensembl.org/pub/release-93/{species}/pep/.

Evigene zebrafish (D_rerio) gene set is at http://hdl.handle.net/2022/22652 (IUScholarWorks)

RNA data sources with NCBI BioProject ID are

• SRA data pig1a: PRJNA416432 (China Agricultural University),

• SRA data pig2b: PRJNA353772 (Iowa State University, USDA-ARS),

• SRA data pig3c: PRJEB8784 (Univ. Illinois),

• SRA data pig4e: PRJNA255281 (Jiangxi Agricultural University, Nanchang, China).

RNA data query for the pig used to select these four, at http://www.ncbi.nlm.nih.gov/sra query=((“biomol transcript”[Properties]) AND “platform illumina”[Properties]) AND “library layout paired”[Properties] AND “Sus scrofa”[Organism]

RNA long-read SRA query for the comparison data set of Table 3B: query=(((“biomol transcript”[Properties]) AND “platform pacbio smrt”[Properties])) AND “Sus scrofa”[Organism]

The SRA read table of these data sets is the starting point for SRA2Genes, and provided at http://eugenes.org/EvidentialGene/vertebrates/pig/pig18evigene/

Expressed sequences of the pig from dbEST, by Sanger and 454 sequencing (max length 900 bases), from projects reported in PubMedID:14681463, dbEST n = 304,418, and PubMedID: 17407547, dbEST n = 716,260.

Vertebrate conserved single-copy genes, of OrthoDB v9 (http://www.orthodb.org), BUSCO.py software, with hmmer (v3.1, http://hmmer.org/) (Waterhouse et al. 2013, Simao et al., 2015).

Software components of EvidentialGene SRA2Genes:

• fastq-dump, of sratoolkit281, https://www.ncbi.nlm.nih.gov/sra/docs/toolkitsoft/

• blastn, blastp of https://blast.ncbi.nlm.nih.gov/ (Altschul et al., 1990)

• vecscreen at http://ncbi.nlm.nih.gov/tools/vecscreen/ and tbl2asn at http://ncbi.nlm.nih.gov/genbank/tbl2asn2/

• fastanrdb, of exonerate, https://www.ebi.ac.uk/about/vertebrate-genomics/software/exonerate (Slater & Birney, 2005)

• cd-hit, cd-hit-est, of https://github.com/weizhongli/cdhit/ (Li & Godzik, 2006)

• normalize-by-median.py, of khmer, https://github.com/ged-lab/khmer (Crusoe et al., 2015)

• velvet, oases of velvet1210 assembler, https://www.ebi.ac.uk/~zerbino/oases/ (Schulz et al., 2012)

• idba_tran, of idba assembler, https://code.google.com/archive/p/hku-idba/downloads/ (Peng et al., 2013)

• SOAPdenovo-Trans, http://soap.genomics.org.cn/SOAPdenovo-Trans.html (Xie et al., 2013)

• Trinity, of trinityrnaseq assembler, https://github.com/trinityrnaseq/trinityrnaseq (Grabherr et al., 2011)

• rnaSPAdes, of SPAdes assembler, http://cab.spbu.ru/software/spades/ (Bankevich et al., 2012)

International Nucleotide Sequence Database Collaboration policy documents pertaining to these data:

• About TSA, https://www.ncbi.nlm.nih.gov/genbank/TSA

• About TPA, https://www.ncbi.nlm.nih.gov/genbank/TPA

• TPA FAQ, https://www.ncbi.nlm.nih.gov/genbank/tpafaq

• TPA-Inferential, https://www.ncbi.nlm.nih.gov/genbank/TPA-Inf.

Supplemental Information