Transcript length bias in RNA-seq data confounds systems biology

“In each analysis we only used genes found on both platforms and then performed a pathway analysis using the DAVID software [7,8].”

“In each analysis we only used genes found on both platforms and then performed a pathway analysis using the DAVID software [7,8].”

“array algorithm [9] but this suggestion requires further exploration and analysis.”

“They then used a method based on the proportion of counts from each library proposed by Audic and Claverie [11] to determine the significant of differential expression.”

“For the data from Marioni et al (2008) genes were matched between the RNA-seq platform and the microarray platform.”

“The first data set compares sequencing from the Illumina Genome Analyzer with Illumina microarrays and looks for differential expression between a human embryonic kidney and B cell line [4].”

“…a predictable consequence of the sampling process and cannot be corrected by dividing by length of the transcript (e.g. the statistical methods in Cloonan et al (2008) or Sultan et al. Differential expression as a function of transcript lengthFigu 1 Differential expression as a function of…”

“By contrast, intensity measurements from microarrays are proportional only to the expression level of the transcript plus any features intrinsic to the probe itself such as GC content [2,3].”

“However, it has been well established that gene length biases are prominent in the analysis of gene expression [15].”

“It should be noted that these estimates apply to the average gene expression level, but as dynamic ranges in RNA-seq data are large, the probability that highly expressed genes will be detected as differentially expressed is greater than that for low-count genes [192].”

“Second, commonly used RNA-seq strategies can result in transcript-length bias because of the multiple fragmentation and RNA or cDNA size-selection steps they us”

“Under the assumptions presented in [11], with the unweighted t-statistics and using the same cutoff across genes, power increases with gene length for a given level of DE.”

“The nature of the short read procedure makes it inevitable that longer transcripts will be preferentially detected over shorter ones, and this has been shown to have implications in the biological interpretation of the data (Oshlack and Wakefield 2009; Young et al. 2010).”

“Even in cases where a strong positive association between gene counts and length is observed, scaling counts by gene length with RPKM is not sufficient for removing this bias [16, 20].”

“Note: Supplementary information is available on the Nature Methods website. aCKnowLedGments expression depends on the sequencing depth of the sample, the expression of the gene, and even the length of the gene 64,68 .”

“To accurately estimate gene expression, read counts must be normalized to correct for systematic variability, such as library fragment size, sequence composition bias, and read depth (Oshlack and Wake fi eld 2009; Roberts et al. 2011b).”

“Another well-known selection bias in RNA-Seq is due to gene length [3,14].”

“Hence there is a clear length bias when measuring gene expression by RNA-seq [20].”

“Technology biases, such as the transcript length (12), GC content (13), PCR artifacts, uneven transcript read coverage, contamination by off-target transcripts, or large differences in transcript distributions (14), are factors that interfere in the linear relationship between transcript abundance and the number of ma…”

“For example, gene length bias is an issue in RNA-seq data, in which longer genes have higher counts (at the same expression level) [66].”

“For example, RNA fragmentation and size selection steps of 200-base-pair fragments in current RNA-seq protocols will likely result in under-representation of the shortest transcripts, as has already been note”

“Because of this length bias, longer transcripts may have more significant differences in differential gene expression studies and in some cases will lead to false-positive differential expression calls (Oshlack and Wakefield, 2009; Bullard et al., 2010).”

“Regarding this point in [92] it is illustrated the inherent bias in transcript length that affect RNA-Seq experiments.”

“To convert raw read counts into informative measures of gene expression, normalization is needed to account for factors that affect the number of reads mapped to a gene, like length [5], GC-content [6] and sequencing depth [7].”

“3c), because this bias stems from oversampling longer genes in approaches that cover the entire mRNA (Oshlack & Wakefield 2009).”

“Normalization by gene length removes information about the original count value and decreases power to detect many genes (Oshlack and Wakefield, 2009).”

“This bias is substantially reduced, again by including the normalization constant as a term; however, some slight read length bias is still apparent (Figure 3f), as previously described [27].”

“However, it should be noted that the Poisson test suffers from gene length bias in a similar way to RNA-seq data (37).”

“In addition, technical variance, sequencing bias and mapping bias have been reported [9-11].”

“…2011 & 2011 EMBO and Macmillan Publishers Limited The accuracy of RNA-seq is biased toward longer and more highly expressed genes, e.g., 5% of all genes account for 50% of all reads in our data as well as in other data sets (Mortazavi et al, 2008; Oshlack and Wakefield, 2009; Bullard et al, 2010).”

“72,77,78 Since the RPKM method handles all the RNA-seq reads almost equally, for example, without concern for isoforms, it has been criticized.”

“However, we note that it has been shown that counts, rather than RPKM values, are preferable for the differential analysis of RNA-seq data (Oshlack and Wakefield, 2009).”

“Of the identified junctions, a larger number of exon skipping events are shared between human and chimpanzee (950) than between either human (808) or chimpanzee (783) and rhesus macaque, as might be expected given the known phylogeny of these species.”

“Using standard PA methods to analyze RNA-seq data may give biased results due to over detection of differential expression for long and highly expressed transcripts (Oshlack and Wakefield, 2009).”

“The RNA-Seq method to determine differentially expressed genes does have an inherent bias towards longer transcripts [16,17].”

“However, for the reason that in RNA-seq experiments the expected read counts of a given transcript is proportional to its abundance multiplied by its length [21,71,72], x, y represented by the mapped read counts of a transcript can not reflect the real expression level of transcripts.”

“In DE analysis, as methods that correct counts for length bias can introduce additional biases [9, 44, 56, 61], normalizations that apply to DE test statistics while leaving gene counts unchanged have been proposed [9, 64].”

“contrary to what had been seen previously [19,43].”

“This nature makes the comparison across samples more efficient for longer transcripts than shorter ones [40, 53].”

“Of these three algorithms, the Fisher exact test is the only one that fails to account for gene length bias; a phenomenon where longer genes are overrepresented in GSEA results, as they contain more CpGs than shorter genes [79].”

“The most commonly used normalization method that includes gene length correction is TPM (Transcripts Per kilobase Million) [9], as other methods like RPKM [1]/ FPKM [10] (Reads/Fragments Per Kilobase per Million reads, respectively, proved to be inadequate and biased [5, 6, 11, 12].”

“required by other high-throughput measurement systems, such as microarrays or sequencing (Churchill, 2002; Marshall, 2004; Frantz, 2005; Bammler et al, 2005; Tang et al, 2007; Balázsi and Oltvai, 2007; Oshlack and Wakefield, 2009).”

“It has been shown that these abundance measures are prone to biases correlated with the nucleotide composition [14,17] and length of the transcript [1,18].”

“Estimating isoform expression levels based solely on unambiguous reads, as suggested, e.g., in [2], results in splicing-depen-dent biases similar to the transcript-length bias noted in [11], further complicating the design of unbiased differential expression tests based on RNA-Seq data.”

“It is known that RNA-sequencing data suffer from transcript length bias—that is, longer transcripts tend to result in a greater number of reads, resulting in an increased tendency to call such transcripts differentially expressed [Oshlack and Wakefield (2009)].”

“, and quantification accuracy for shorter transcripts is lowe”

“However, while the number of RNAseq reads available already runs into the hundreds of millions, concerns have been raised about the power of this technique to find transcripts with very low expression levels (Oshlack and Wakefield 2009).”

“To alleviate the bias in the detection of differential expression influenced by the transcript length (Oshlack and Wakefield, 2009; Young et al, 2010), we added the expression ratio cutoff (two-fold).”

“Genes with unknown gene length were excluded (n = 680), because there might be gene length effects [25] and we need them later in the simulation study.”

“found transcript length biases in RNA-Seq data [9].”

“As in all RNASeq experiments, the variable length of transcripts may produce biases in the data [23]; therefore, we analyzed GO enrichment via the Goseq R package, which was specifically designed to limit bias in the data [24].”

“a completely unbiased measure of the expression level due to different transcript length and/or overrepresentation of 50- or 30-ends of the transcripts [13, 40].”

“Several studies have begun to investigate on the parameters like sequencing depth, precision, GC bias, length bias, lane effects, and processing artifacts [16,29,48,75-77].”

“Although RPKM/FPKM is a popular normalization strategy, it has been demonstrated to bias subsequent calls for differential expression in favor of longer transcripts [11].”

“Figure 3 highlights that more than 80% of reads are attributed to the 10% most expressed genes, acknowledging that transcript length also plays a role [22] .”

“While systemic deviations due to length bias, lane effects, and processing artefacts are increasingly being investigated (Bullard et al., 2010; Marioni et al., 2008; Oshlack and Wakefield, 2009), reproducibility has in general received much less attention.”

“Recent work on RNA-Seq has revealed a potential detection bias toward differential expression of genes with long transcripts (36).”

“Simply dividing read counts by exon length does not correct for this [26], and the use of AC would then be inappropriate since the data would no longer be Poisson-distributed.”

“RNA-seq analysis requires specialized computational tools that can account for the shortcomings of sequencing technologies, including the generation of sequencing errors (Le et al., 2013), length biases (Oshlack and Wake fi eld, 2009), and fragmentation (Tuerk et al., 2017).”

“Clusters of similar DE hit lists were defined by first performing hierarchical clustering of Euclidean distances with average linkage, then using the cutreeDynamicmethod (22) with the following parameter settings: deepsplit= 2, minClusterSize = 10, and pamRespectsDendro = FALSE.”

“The application of GSA to lists of genes found to be differentially expressed between groups of samples using RNA-seq is known to be affected by such a bias (Oshlack and Wakefield, 2009).”

“Standard RNA-Seq has been applied to mutation identification and transcriptional profiling.”

“For example, an unexplained bias towards increased variance in RNA-Seq quantification of smaller transcripts was recently reported (Oshlack &Wakefield 2009).”

“Oshlack and colleagues [85] have reevaluated some published data, which demonstrated that the differences in statistical power rely on the chances of a read to hit a certain gene.”

“As noted previously in the literature [11, 12, 13, 14] (illustrated in Fig.”

“However, it has recently been demonstrated that transcript length affects differential expression analysis [47].”

“One class of errors concerns biases in the abundance of read sequences due to RNA priming preferences (12,13), fragment size selection (14,15) and GC-content (16).”

“Gene length bias has been shown to be an important bias in RNA-Seq, as longer genes will be sequenced deeper than shorter genes [31].”

“In particular, these variations include between-sample differences such as library size (sequencing depth) [5] or within-sample differences, for example, in gene length [6], guanine-cytosine (GC) content [7, 8], or unwanted variation introduced by batch effect [9].”

“It is becoming clear now that differences in template preparation protocols create biases in the detection of gene expression and several bioinformatic applications have been developed to compensate for this bias [16,17].”

“However, this would exclude a certain proportion of the data [20].”

“Normalization by gene length is important for comparing different genes within a sample because longer genes tend to have more reads to be sequenced [6].”

“…lengths and sequencing depths, non-uniformity of read distributions along transcripts and strong sample-specific GC-content effect (Mortazavi et al., 2008; Oshlack and Wakefield, 2009; Zenoni et al., 2010; Risso et al., 2011; Roberts et al., 2011; Tarazona et al., 2011; Hansen et al., 2012).”

“As gene length biases both gene counts and their downstream analysis [26], it is not very surprising that TPM usually outperforms CPM, as CPM only corrects for library size”

“Additionally, mRNA-Seq is known to have gene-length bias toward genes of larger lengths, and the within-library normalization to normalize individual genes within the same libarary was proposed (Oshlack and Wakefield 2009).”

“Variation introduced by different gene lengths (195), as is frequently encountered in sequencings of longer RNAs, has a negligible effect.”

“[28]] and that accuracy depends on proper mapping and coverage, and that GC”

“Although deep-sequencing experiments have become the major source supporting microRNA annotations, technical variations, transcript length bias and mapping bias have been reported with this approach for transcriptome analysis of coding RNAs (mRNAs) [17,55,56].”

“We also found little bias in favour of detecting differential expression for longer genes or reference sequences, as has been reported in some studies [31].”

“…of bias in gene expression estimates based on next-generation sequencing data are not yet fully understood, it is becoming clear that base composition, library preparation and transcript length all have an impact on RNA-Seq data (Marioni et al. 2008; Gilad et al. 2009; Oshlack & Wakefield 2009).”

“…somewhat shorter RNA fragment size in the TGIRT-seq libraries, which results in more frequent sampling across RNA molecules (Mortazavi et al. 2008; Oshlack and Wakefield 2009), and the high processivity and strand displacement of the TGIRT enzyme operating at 60°C, which enables it to reverse…”

“RNAseq), which generate longer tags, can overcome this limitation but they introduce new problems, the most important of which is multiple tags per transcript and a bias towards highly expressed long transcripts [55].”

“To accurately estimate gene expression, read counts must be normalized to correct for systematic variability, such as library fragment size and read depth [97, 140, 158].”

“Because longer transcripts give more statistical power for detecting differential expression between samples [12], Gene Ontology (GO) analysis was performed using the goseq package that adjusts transcript length bias in deep sequencing data [13].”

“In addition, RNA-Seq has two notable sources of bias that an analyst may need to address: transcripts with longer lengths [4] and higher GC content [5] have their relative abundances over-estimated.”

“When cDNAs are fragmented, long genes result in more fragments for the same number of transcripts, resulting in higher counts and more power to detect differential expression (Oshlack & Wakefield, 2009).”

“…Pollock, Glendinning, Wisedchanwet, & Watson, 2018; Schori, Appel, Kitko, & Showalter, 2013), cycle count (Kelly et al., 2019; Silverman et al., 2019), GC content (Laursen, Dalgaard, & Bahl, 2017; Risso, Schwartz, Sherlock, & Dudoit, 2011), and length of the amplicon (Oshlack & Wakefield, 2009).”