Ameliorated de novo transcriptome assembly using Illumina paired end sequence data with Trinity Assembler

Abstract

Advent of Next Generation Sequencing has led to possibilities of de novo transcriptome assembly of organisms without availability of complete genome sequence. Among various sequencing platforms available, Illumina is the most widely used platform based on data quality, quantity and cost. Various de novo transcriptome assemblers are also available today for construction of de novo transcriptome.

In this study, we aimed at obtaining an ameliorated de novo transcriptome assembly with sequence reads obtained from Illumina platform and assembled using Trinity Assembler. We found that, primary transcriptome assembly obtained as a result of Trinity can be ameliorated on the basis of transcript length, coverage, and depth and protein homology. Our approach to ameliorate is reproducible and could enhance the sensitivity and specificity of the assembled transcriptome which could be critical for validation of the assembled transcripts and for planning various downstream biological assays.

Specifications
Bioproject accession	PRJNA244590
Organism	Helicoverpa armigera Hübner, 1809
Common name	Cotton bollworm
Sex	–
Project data type	Transcriptome or gene expression
Platform	Illumina HiSeq 2000
Data format	SRA
Library details	Strategy: RNA-seq
Source: Transcriptomic
Selection: cDNA
Layout: Paired
Experimental factors	The data consist of 10 RNA-seq cDNA libraries. Helicoverpa armigera (Hübner) strains were cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis. Sensory organs i.e. taste organs in adults (male and female) and in larvae also.
Experimental features	Transcriptome survey for identifying genes relevant to chemoreception
Consent	–
Sample source location	CSIRO Ecosystem Sciences, Black Mountain, Canberra ACT 2601, Australia
Link	ftp://ftp-trace.ncbi.nlm.nih.gov/sra/sra-instant/reads/ByStudy/sra/SRP/SRP041/SRP041166

1. Data files

Accession	Title	Source name	Organism	Treatment
SRR1238087	Mixed-feeding 5th instar-antennae	SRX516834	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1238089	Mixed-feeding 5th instar-mouthpart	SRX516871	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1238090	Male-adult-tarsus	SRX516872	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239328	Female-adult-tarsus	SRX518085	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239329	Mixed feeding 5th instar fat body	SRX518086	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239330	Female adult abdomen	SRX518087	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239331	Female adult head wei	SRX518088	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239333	Female adult head wei DSN	SRX518089	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239334	Male adult head wei	SRX518090	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis
SRR1239335	Male adult head wei DSN	SRX518091	Helicoverpa armigera	Cultured on artificial diet containing the Cry1Ac protoxin of Bacillus thuringiensis

2. Material and methods

Deep sequencing based whole transcriptome data for reanalysis was obtained from NCBI SRA (Sequence Read Archive) with the link ftp://ftp-trace.ncbi.nlm.nih.gov/sra/sra-instant/reads/ByStudy/sra/SRP/SRP041/SRP041166 [1]. Raw data was obtained in SRA format and further converted to FASTq format using SRA Tool kit (version 2.4) (http://www.ncbi.nlm.nih.gov/Traces/sra/) [2].

2.1. Raw data summarization

Paired End FASTq files were subjected to standard quality control using NGSQC Tool Kit [3] with the following command:

• perl IlluQC_PRLL.pl -c 10 -pe SRR1238087_R1.fastq SRR1238087_R2.fastq Adapter.txt A –o “Output Folder Name”.

For each paired end data 10 core threads of processing with 2.4 GHz speed with default memory allocation were provided.

2.2. Transcriptome assembly

All the 10 HQ filtered paired end libraries were subjected to pooled de-novo transcriptome assembly as followed in the original manuscript [1]. Evaluation of multiple assemblers for de novo transcriptome assembly was already done and results are available [4], [5], [6]. For this study we chose to go with De brijn graph based Trinity Assembler [7] based on the criteria of a) default K-mer, b) less memory foot print, c) optimized for Illumina paired end data, d) reproducibility and e) configurable for all computing capacities. The following command was used to initiate the pooled assembly using Trinity.

Trinity\

• –seqType fq\
• –JM 600G\
• –left/data/Projects/RNASeq/Pooled_Reads/R1.fastq\
• –right/data/Projects/RNASeq/Pooled_Reads/R2.fastq\
• –CPU 50\
• –min_contig_length 200\
• –output/data/Projects/RNASeq/Pooled_Reads/Helicoverpa\
• –min_kmer_cov 2\
• –bflyHeapSpaceMax 50G\
• –bflyHeapSpaceInit 10G

For the pooled assembly 50 core threads of processing with 2.4 GHz speed and a maximum Heap Space of 50 GB were allotted.

2.3. Assembly validation

Since de novo transcriptome assemblers are capable of producing in fragmented/mis-assembly, validation of the assembled transcriptome is done by mapping back the HQ filtered reads to the ESTs. Bowtie [8] was used to map the HQ filtered reads from each library to the assembled transcriptome using the following parameters.

perl TRINITY_HOME/util/align_and_estimate_abundance.pl \

• –transcripts TrinityMergedAssembly.fasta \
• –seqType fq \
• –left SRR1238087_R1.fastq_filtered \
• –right SRR1238087_R2.fastq_filtered \
• –est_method RSEM \
• –aln_method bowtie \
• –thread_count 10

2.4. Transcript quantitation, coverage and depth analysis

Assembly validated .bam (Binary Sequence Alignment/Map) file was processed using bedtools [9] and samtools [10] for quantitation (read count estimation) for each transcript in a library and also to calculate the total coverage and average depth of the transcriptome in each library.

For quantitation the following parameters/command was used.

• samtools idxstats SRR1238087.bowtie.csorted.bam > SRR1238087.bowtie.csorted.bam.idxstats

For calculating each transcript coverage and its average depth corresponding bedGraph file was generated using the following the parameters/command.

• genomeCoverageBed -ibam SRR1238087.bowtie.csorted.bam -bga > SRR1238087.bedgraph

From the resultant bedGraph file, the following formulae were used to calculate coverage and average depth.

$$\mathrm{Average}\mathrm{Depth}=\frac{\left[\mathrm{Number}\mathrm{of}\mathrm{Reads}\mathrm{Mapped}\right]*\left[\mathrm{Read}\mathrm{Length}\right]}{\mathrm{Length}\mathrm{of}\mathrm{Transcript}}$$

$$\mathrm{Coverage}=\left[\frac{\mathrm{Mappapalbe}\mathrm{Transcript}\mathrm{Length}}{\mathrm{Length}\mathrm{of}\mathrm{Transcript}}\right]*100$$

2.5. Analysis of transcriptome integrity

While doing merged assembly multiple transcripts might arise due to errors in assembly. In our approach, we performed transcriptome integrity analysis based on read count, coverage and average depth on an intra- and inter-library specific manner. Correlation coefficient graphs were plotted to understand the variations between the libraries that could be indicative of whole transcriptome integrity. Isotig analysis of validated transcriptome based on length was also done to estimate the transcriptome integrity.

2.6. Transcript annotation

Homology based annotation for each transcript was done against NCBI nrdb (Dec 2014) protein database using Blastx. Annotation and statistical ranking of the results were done using Blast2GO [11]. Also domain level annotation was performed using the Online InterProScan tool [7] RunIprScan-1.1.0 (http://michaelrthon.com/runiprscan/).

Blastx and Blast2GO parameters used are

• e-Value <= 10 − e4

• Similarity ≥ 35%

• Annotation cutoff ≥ 55

• GO weight cutoff ≥ 5.

2.7. Normalization and expression profiling

Primary advantage of using NGS based transcriptome profiling is to identify sample/condition specific expressed transcripts which is not easy with earlier hybridization methods. Transcripts with a read count ≥ 10 in any one of the library was considered to be as likely expressed. A sub .bam file was created from the master .bam file based on the above criteria. The sub .bam file was subjected to normalization and expression profiling using RSEM software [12]. The following parameters/commands were used to normalize each library.

• rsem-prepare-reference\

• --no-polyA\

• --no-bowtie\

• ValidTranscripts.fasta\

• Harmigera_RSEM

• /data/Program/rsem-1.2.12/rsem-calculate-expression\

• --paired-end\

• -p 8\

• --bam\

• SRR1238087.bowtie.csorted.bam\

• Harmigera_RSEM\

• SRR1238087

RSEM software provides an output for each library with expected normalized read count, TPM (tags per million) and FPKM (fragments per kilobase per million). Log to the base 2 of FPKM was considered as absolute expression or Delta CT equivalent value.

3. Results

3.1. Raw data summarization and transcriptome assembly

Quality control of individual libraries using NGSQC tool kit revealed an average of 95.85% HQ reads based on Q20 score. Total number of HQ reads on an average per library was ~ 47 million, indicating significant amount of reads to proceed with transcriptome assembly (Table 1, Fig. 1). HQ paired reads from all the libraries were merged and provided as input to Trinity Assembler V2014. Transcriptome assembly performed with parameters as outlined resulted in assembly of 74,966 putative transcripts with an overall size of 78.61 Mb. N50 of the assembled transcriptome was 2.12 kb. Transcript length distribution analysis revealed 51.79% of the transcript length less than 500 bases. This is a typical observation of most of the de novo transcriptome assemblies using Illumina short read deep sequencing approach [13], [14], [15] (Table 2, Fig. 2).

Table 1.

Distribution of reads based on quality score from each library indicating percentage of high quality and low quality reads.

Sample ID	HQ reads	Low quality reads
SRR1238087	47,523,826 (99.69%)	148,556 (0.31%)
SRR1238089	49,202,920 (99.68%)	160,166 (0.32%)
SRR1238090	49,791,084 (99.68%)	160,610 (0.32%)
SRR1239328	46,376,028 (99.76%)	113,622 (0.24%)
SRR1239329	48,642,138 (93.75%)	3,244,266 (6.25%)
SRR1239330	31,475,588 (91.96%)	2,753,084 (8.04%)
SRR1239331	66,399,420 (94.23%)	4,065,716 (5.77%)
SRR1239333	32,605,304 (93.83%)	2,142,264 (6.17%)
SRR1239334	56,840,276 (91.69%)	5,151,868 (8.31%)
SRR1239335	42,943,308 (94.27%)	2,612,200 (5.73%)

Fig. 1.

Histogram representation of high quality and low quality reads from each individual library.

Table 2.

Statistics of merged de novo transcriptome sequence assembly using Illumina paired end reads using Trinity Assembler 2014 release.

Number of transcripts	74,966
Transcriptome size (Mb)	78.61
Mean (bp)	1049
Stdev (bp)	1319
Median (bp)	472
Smallest (bp)	201
Largest (bp)	29,186
N50 length (bp)	2123

Fig. 2.

Histogram representation of abundance of putative assembled transcripts based on their length range.

3.2. Validation of putative assembled transcriptome and quantitation

A merged de novo assembly is expected to provide representative transcriptome of transcripts from individual libraries. This would be evident from mapping the reads from each library to the putative transcriptome to validate the build. In our approach too, we mapped/aligned the reads back to the putative transcripts from the assembly to understand if there is any library specific bias (enrichment or depletion). Copy number of each transcript from each library was also measured to understand any copy number specific bias that could arise due to upstream sample preparation artifacts. We observed on an average of 72% of HQ reads from each library mapped to the putative transcriptome (Table 3, Fig. 3).

Table 3.

Alignment statistics indicative of reads aligned to the assembled transcriptome as a result of standard parameters used in Trinity Assembler.

Sample ID	Aligned reads	Unaligned reads
SRR1238087	34,972,638 (73.36%)	12,551,188 (26.33%)
SRR1238089	36,915,066 (74.78%)	12,448,020 (25.22%)
SRR1238090	37,346,712 (74.77%)	12,604,982 (25.23%)
SRR1239328	33,406,742 (71.86%)	13,082,908 (28.14%)
SRR1239329	32,489,850 (62.62%)	19,396,554 (37.38%)
SRR1239330	23,811,380 (69.57%)	10,417,292 (30.43%)
SRR1239331	53,009,538 (75.23%)	17,455,598 (24.77%)
SRR1239333	25,688,662 (73.93%)	9,058,906 (26.07%)
SRR1239334	45,313,564 (73.1%)	16,678,580 (26.9%)
SRR1239335	34,384,314 (75.48%)	11,171,194 (24.52%)

Fig. 3.

HQ read alignment to putative transcriptome.

3.3. Analysis of transcriptome integrity and refinement

HQ aligned reads were subjected to integrity analysis with reference to coverage and average depth ratio in each library. Coverage was calculated as percentage of the transcript length supported by aligned reads and average depth was calculated as number of bases supporting each nucleotide in a transcript. We found that a total of 37,930 transcripts (50.59%) were covered at ≥ 70% with an average depth of 5 × in one or more of the libraries (Table 4). Further, we subjected the 37,930 transcripts to length and annotation analysis to establish the integrity and refinement. We observed the minimal transcript length of 201 bases and maximum to be 29.18 kb with an average length of 1.19 kb. A total of 15,197 out of 37,930 transcripts got assigned to a protein based on homology at protein level (Fig. 4). Comparative analysis of assembled and ameliorated transcriptome on the basis of length, coverage, depth and annotation showed clear improvement in the assembly process (Fig. 5). Isotig analysis of assembled transcriptome and ameliorated transcriptome also showed significant improvement in the quality of the transcriptome build (Fig. 6).

Table 4.

Matrix representation of depth vs coverage of individual libraries with highlight on transcripts with ≥ 70% coverage and ≥ 5 × depth.

Fig. 4.

a — Distribution of assembled transcripts with reference to length and their annotation with NRDB representing noise of smaller transcripts.

b — Distribution of ameliorated transcripts with reference to length and their annotation with NRDB shows nice correlation by reducing noise of smaller length transcripts.

Fig. 5.

Species distribution of assembled transcriptome and ameliorated transcriptome [Top 10].

Fig. 6.

Isotig analysis with respect to assembled transcriptome and ameliorated transcriptome.

3.4. Expression profiling, Gene Ontology and Pathway enrichment of ameliorated transcriptome

Normalization of assembled transcriptome and ameliorated transcriptome was done as per the discussed method. Box plot representation of both the transcriptome showed significant difference in the ameliorated global expression profile in comparison to assembled transcriptome (Fig. 7). Ameliorated transcriptome was subjected to Gene Ontology and Pathway analysis as discussed in the methods to identify key enriched gene ontology categories and pathways. Top 10 GO categories were found to represent essential biological processes (Fig. 8). Complete assembled and annotated transcriptome along with transcript length, read count, depth and coverage is provided along with the transcriptome sequence (Supplementary File 1, Supplementary File 2).

Fig. 7.

Global expression profiling of ameliorated transcriptome (post-amelioration) in comparison to complete expression profile (pre-amelioration) represented in Box Plot.

Fig. 8.

Top 10 GO categories were found to represent essential biological processes.

4. Discussion

Next Generation Sequencing based gene expression studies enable faster and cheaper data generation. Illumina is the widely used sequencing platform for whole transcriptome studies. Since the advent of novel sequencing methodologies, de novo transcriptome sequencing is the method of choice for conducting spatial, temporal and condition specific gene expression profiling in both non-model and model organisms.

With hundreds of de novo transcriptomes published with majority using Illumina sequencing platform, the integrity and resolution of the assembled transcriptome remain un-addressed. The choice of the platform, assembler and sample size and study design largely determines the sensitivity and specificity of the assembled transcriptome. The most important step in de novo RNA-seq analysis is the assembly of the sequencer generated short reads into full-length transcripts. Among the well-known, publicly available software's for de novo transcriptome sequence assembly are: Trinity, Velvet-Oases, SOAPdenovo-trans assembler and the Trans-ABySS. Trans-ABySS, SOAPdenovo-trans and Velvet-Oases are extensions of the pre-developed genome assembler programs. The Illumina based Trinity Assembler is the most widely used tool that was developed primarily for de novo RNA-seq data assembly.

In case of a multiple sample study involving analysis of the differentially expressed genes it is recommended to combine all the reads from independent samples and obtain a merged assembly [16]. Thus obtained merged assembly has a representation of all the transcripts in the given set of samples allowing for a true differential expression analysis. Although Trinity is among the most efficient tools for reconstructing transcripts in the absence of a reference sequence yet number of limitations has been encountered with this assembler. The first among them is the lack of reproducibility. Second are the high rate of false positives in the assembly ranging from 20 to 30% and the presence of large number of fusion transcripts as well as partial transcripts. Third, the number of obtained transcripts is too high compared to the expected number in the particular organism in the study. As a result the number of annotated transcripts is observed to be very hugely different in every experiment in the range of 40–90%. The high degree of variability in the results is evidenced by a low validation score.

In this study we attempted to benchmark various parameters when taken into consideration, could result in enhancing the sensitivity and specificity of the assembled transcriptome, considering Illumina as the sequencing platform of choice and Trinity as the assembler of choice.

The following are the supplementary data related to this article.

Supplementary File 1

Spreadsheet with list of 74,966 transcripts with information about transcript length, read count, coverage, average depth and annotation.

Supplementary File 2

FASTA file of assembled transcriptome (74,966 transcripts).