RNA extraction using RNeasy kit (Qiagen, Australia) according to the manufacturer’s instructions. The RNA-Seq experimental process shows the steps for the experimental pipeline. During the QC step, Agilent 2100 Bioanaylzer was used to qualify and quantify the sample library. The library products were ready for sequencing on the Ion Proton platform performed by the Beijing Genomics Institute (BGI-Shenzhen).

The total RNA samples were first treated with DNase I to degrade any possible DNA contamination. Then the mRNA was enriched by using the oligo (dT) magnetic beads. Mixed with the fragmentation buffer, the mRNA was fragmented into short fragments (about 200 bp). Then the first strand of cDNA was synthesized by using random hexamer-primed reverse transcription. Buffer, dNTPs, RNase H and DNA polymerase I were added to synthesize the second strand. The double strand cDNA was purified with magnetic beads. End reparation was then performed. After the previous step, adaptors were ligated to the end of these fragments. Next, ligation products were selected by size and purified on TAE-agarose gel. Finally, the fragments were enriched by PCR amplification, then purified by magnetic beads and dissolved in the appropriate amount of Epstein-Barr solution. During the QC step, Agilent 2100 Bioanaylzer was used to qualify and quantify the sample library. The library products were then ready for sequencing via Ion Proton platform.

The original image data was transferred into sequence data via base calling, which is defined as raw data or raw reads and saved as BAM file. As the raw reads may contain low quality reads or adaptor sequences, preprocessing was necessary before starting further analysis.

As there are some adaptor sequences and/or low quality reads present in the raw reads data filtering was carried out to obtain high quality reads as the clean reads (clean data). Filtering steps were as follows:

1) Remove reads where length was less than threshold;

2) Trim reads adapter, if length of trimmed reads was less than threshold, then remove it;

3) Calculate the average quality of 15 bases from 3’ end until the average quality was larger than 10, then trim the bases that have been counted.

After filtering, the remaining reads are called “clean reads” and used for downstream bioinformatics analysis.

We performed quality control on clean data through drawing base composition, quality and length distribution charts. ATGC base content refers to the proportion of the total bases of the four types of bases. Through detecting the content of ATGC each base to measure the stability of library and sequencing eligibility. The nucleotide distributions at each position were stable under the normal circumstances, non-AT or GC separation. Bases quality reflects the accuracy which can be affected by RNA-Sequencing, reagents or sample quality. If the percentage of the bases with low quality is low, then it indicated the sequencing quality is good.

After data quality statistics, clean reads were mapped to gene reference and/or genome reference set.

We used TMAP to map clean reads to gene reference and/or genome reference. In general, the higher ratio of alignment, indicated the closer the genetic relationship between the samples and the reference faba bean species.

Sequence saturation analysis was used to measure the sequencing data of the sample. The number of detected genes increased with an increase in the number of reads. However, when the number of reads reaches a certain amount, the growth curve of detected genes flattens, which indicates that the number of detected genes tends to saturation.

If the randomness is good, the reads in every position would be evenly distributed. If the randomness is poor, reads preference to specific gene region will directly affect subsequent bioinformatics analysis. We used the distribution of reads on the reference genes to evaluate the randomness. A distribution of reads position on genes shows that the randomness of RNA fragmentation is better than cDNA fragmentation [32]. Therefore, we used the mRNA fragments to construct the library in the experiment.

We provide genome alignment data in BAM format and recommend using IGV (Integrative Genomics Viewer) tool to visualize BAM file in different scales. IGV supports loading of multiple samples to do comparison in the same scale, and can view distribution of reads on the Exon, Intron, UTR, and Intragenic regions, which makes it very convenient and intuitional.

Expression levels of individual Unigenes were quantified using the software package Sailfish [33]. Sailfish quantifies genes expression by k-mer, first, built index files use reference and length of k-mer, then computes maximum likelihood abundance estimates using the Expectation-Maximization (EM) algorithm for its statistical model to determine which transcripts are isoforms of the same gene. Expression level was measured in reads per kilobase per million mapped reads (RPKM) according to Equation (1):

RPKM = 10 6 C N L / 10 3 (1)

where C is the number of reads that are uniquely aligned to a specified gene (A), N is the total number of reads uniquely aligned to all genes and L is the length of the specified gene (A) in bases. The RPKM values can be directly used for comparing the difference of gene expression among samples. If there was more than one transcript for a gene, the longest one was used to calculate its expression level and coverage. Linear correlation of RPKM values was used to assess the robustness of experimental comparisons made between genotypes.

For multiple samples, we can do more deep analysis based on gene expression to do a comprehensive assess on the whole project.

We calculated the correlation value between each two genotypes based on RPKM results. According to the standard that Encode plan recommends, the square of the correlation value should be ≥0.92 (under ideal experiment environment and with reasonable samples).

The distances of expressed genes in samples were calculated by Euclidean method. Meanwhile, the algorithm of Sum of Squares of Deviations was used to calculate the distance between samples so that a cluster tree could be built.

Genes with similar expression patterns usually have same the functional correlation. We performed cluster analysis of gene expression patterns with cluster [34] [35] and java Treeview software [36].

Principal component analysis (PCA) can reduce the complexity of the data, and dig deep into the relation between sample size and variation. The basic principle is that diverse samples have different measurements; PCA identifies the main factors of observed value differences, considering all the factors are combined and sorted according to importance. In PCA graph it can see the distance of the relationship between each sample, including visual effect of clusters groups. The PCA analysis is presented as a figure where each dot represents the genotype on the principal component value.

Screening for DEGs was based on the Poisson distribution method described by Audic & Claverie [37] and corrected P-values using the Bonferroni method [38]. Since DEGs analysis generates a large multiplicity of problems in which thousands of hypotheses (is gene x differentially expressed between the two groups) are tested simultaneously, correction for false positive (type I errors) and false negative (type II) errors was performed using the False Discovery Rate (FDR) method [39]. We used FDR ≤ 0.001 and the absolute value of Log₂Ratio ≥ 1 as the threshold to judge the significance of gene expression difference.

DEGs screening is aimed to find differentially expressed genes between (within) samples and perform further function analysis on them. We provide three optional methods for screening analysis of differentially expressed genes, respectively: Based on the analysis method of the Poisson distribution; Based on Noiseq package method; Based on EBSeq package method.

According to the analysis method of the Poisson distribution which screening DEGs between two genotypes. Referring to the significance of digital gene expression profiles, we have developed a strict algorithm to identify differentially expressed genes between two samples. Denote the number of unambiguous clean tags (which means reads in RNA-Seq) from gene A as x, given every gene’s expression occupies only a small part of the library, x yields to the Poisson distribution with Equation (2):

p ( x ) = e − λ λ x x ! ( λ is the real transcripts of the gene) (2)

The total clean tag number of sample 1 is N₁, and total clean tag number of sample 2 is N₂; gene A holds x tags in sample 1 and y tags in sample 2. The probability of gene A being expressed equally between two samples can be calculated with Equation (3):

2 ∑ i = 0 i = y p ( i | x )

or   2 × ( 1 − ∑ i = 0 i = y p ( i | x ) )     ( if   ∑ i = 0 i = y p ( i | x ) > 0.5 )

p ( y | x ) = ( N 2 N 1 ) y ( x + y ) ! x ! y ! ( 1 + N 2 N 1 ) ( x + y + 1 ) (3)

We do correction on P-value corresponds to differential gene expression test using Bonferroni method. Since DEG analysis generates a large multiplicity problems in which thousands of hypotheses (is gene x differentially expressed between the two groups) are tested simultaneously, correction for false positive (type I errors) and false negative (type II) errors are performed using FDR method. Assume that we have picked out R differentially expressed genes in which S genes really show differential expression and the other V genes are false positive. If we decide that the error ratio “Q = V/R” must stay below a cutoff (e.g. 5%), we should preset the FDR to a number no larger than 0.05. We use “FDR ≤ 0.001 [40] and the absolute value of Log₂Ratio ≥ 1” as the threshold to judge the significance of gene expression difference. More stringent criteria with smaller FDR and bigger fold-change value can be used to identify DEGs.

Based on Noiseq package method (screening DEGs among two groups which contain biological replicates) Noiseq method [41] can screen differentially expressed genes between two groups, showing a good performance when comparing it to other differential expression methods, like Fisher’s Exact Test (FET), edgeR, DESeq and baySeq. Noiseq maintains good True Positive and False Positive rates when increasing sequencing depth, while most other methods show poor performance. In addition, Noiseq models the noise distribution from the actual data, so it can better adapt to the size of the data set, and is more effective in controlling the rate of false discoveries.

First, Noiseq uses the sample’s gene expression in each group to calculate log₂ (foldchange) M and absolute different value D of all pair conditions to build noise distribution model with Equation (4):

M i = log 2 ( x 1 i x i 1 )     and     D i = | x 1 i − x 12 i | (4)

Second, for gene A, Noiseq computes its average expression “Control-avg” in control group and average expression “Treat-avg” in treatment group. Then the foldchange (M_A = log₂((Controlavg)/(Treat-avg))) and absolute different value D(D_A = |Control-avg-Treat-avg|) will be derived. If M_A and D_A diverge from noise distribution model markedly, gene A will be defined as a DEG. There is a probability value to assess how M_A and D_A both diverge from noise distribution model with Equation (5):

P A = P ( M A ≥ { M } & & D A ≥ { D } ) (5)

Finally, we screen differentially expressed genes according to the following criteria: Fold change ≥ 2 and diverge probability ≥ 0.8.

Cluster analyses of DEGs with similar expression patterns usually have the same functional correlation. We perform cluster analysis of DEGs with cluster and java Treeview software [36]. Expression differences are shown in different colors. Red means up regulation and green means down regulation.

Genes usually interact with each other to play roles in certain biological functions. Pathway-based analysis helps to further understand genes biological functions. KEGG (Kyoto Encyclopedia of Genes and Genomes) [38] pathways, the major public pathway-related database, is used to perform pathway enrichment analysis of DEGs. This analysis identifies significantly enriched metabolic pathways or signal transduction pathways in DEGs comparing with the whole genome background. The calculating formula as where N is the number of all genes with KEGG annotation, n is the number of DEGs in N, M is the number of all genes annotated to specific pathways, and m is the number of DEGs in M. The calculated p-value goes through Bonferroni Correction [40] method, taking corrected p-value < 0.05 as a threshold. Discovered the most significant pathways of faba bean seed RNA-Seq, allows us to see their detailed pathway information in the KEGG database. The pathways showed with their up-regulated genes which are marked with red borders and down-regulated genes with green borders, Non-change genes are marked with black borders. A scatter plot for KEGG enrichment results was generated to display the top 20 enriched pathways (See Figure 1 and Supplementary Figures S1-S14). Rich Factor is the ratio of DEGs numbers annotated in this pathway to all gene numbers. Greater Rich Factor means greater intensiveness. Q-value is corrected p-value ranging from 0~1, and a lower value means greater intensiveness.