TRANSCRIPTION FIDELITY: NEW PARADIGMS IN EPIGENETIC INHERITANCE, GENOME INSTABILITY AND DISEASE

Abstract

RNA transcription errors are transient, yet frequent, events that do have consequences for the cell. However, until recently we lacked the tools to empirically measure and study these errors. Advances in RNA library preparation and next generation sequencing (NGS) have allowed the spectrum of transcription errors to be empirically measured over the entire transcriptome and in nascent transcripts. Combining these powerful methods with forward and reverse genetic strategies has refined our understanding of transcription factors known to enhance RNA accuracy and will enable the discovery of new candidates. Furthermore, these approaches will shed additional light on the complex interplay between transcription fidelity and other DNA transactions, such as replication and repair, and explore a role for transcription errors in cellular evolution and disease.

1. Introduction

DNA→RNA→protein has been the mantra of biology for over half a century [1]; but while this tenet relies on RNA as the messenger moderating the relationship between genotype and phenotype, little is appreciated of the consequences of errors in the RNA message. And yet far more variability exists between cells than can be explained by DNA variants and a spontaneous mutation rate of ~10⁻¹⁰ [2], begging the question: What is our mantra missing? To bridge this gap, biologists have turned to epigenetics. While the word has been co-opted in recent years for chromatin modification and histone markers, in the purest sense, “epigenetic” describes gene expression alterations that cannot be accounted for by the DNA sequence [3,4]. As gene expression is regulated by complex networks, there exists a dizzying number of epigenetic modifiers that exert their effects through chromatin accessibility, phase separation [5], transcription, post-transcriptional modifications, translation and post-translational modifications. Dysregulation or errors in any of these processes can affect gene expression and phenotype. Thereby, on a larger scale epigenetic errors can promote heterogeneity between cells and thereby evolution. Ribosomes make translation errors at rates estimated between 10⁻² and 10⁻⁶ per codon and are a source of non-genetic phenotypic variation (reviewed in [6]). However, transcription as a source of phenotypically relevant errors has been largely overlooked. If we narrow our definition of epigenetics to include specifically those gene expression changes that are heritable, any purported epigenetic mechanism must produce heritable change. One goal of this review is to not only illustrate the importance of transcription errors, but also elucidate their potential to generate heritable change.

Compared to the well-studied mechanisms of DNA mutation, less is known about the biology of RNA errors or their role in health and disease. Broadly defined, transcription errors, also known as “epimutations”, are sequence deviations from the DNA template that occur in nascent transcripts. They are distinct from downstream RNA modifications and editing but can exert similar, and potentially functional, consequences on the heterogeneity of the protein pool. Transcription errors stem from the inherently error-prone activity of DNA-dependent RNA polymerases (RNAP) with error rates reported between 10⁻⁵ and 10⁻⁶ events per base, >5,000-fold higher than DNA replication errors [7]. Transcription factors, DNA damage, certain sequence motifs and the general genomic environment are all believed to modulate the frequency of transcription errors, in addition to external conditions (e.g. oxidative stress [8]). A class of proteins exists to enhance RNAP proofreading - so-called “RNA fidelity factors” - that together share a fascinating link to genomic stability, epigenetics, phenotypic heterogeneity, and cellular evolution through mechanisms under current investigation. This review intends to present what is known about transcription errors, highlight essential, unanswered questions in the field, and predict that transcription fidelity will define a new paradigm in evolution and disease over the coming years.

2. Mechanism of transcription errors

2.1. General Overview

Transcription errors encompass both misincorporations and frameshifts, giving rise to point epimutations and indels, respectively. Frameshifting errors tend to be more destabilizing to the RNA:DNA hybrid in the catalytic center of RNAP and thus are much less frequent (~10⁻⁷ per base), with insertions outnumbering deletions [7]. Error rates vary greatly between species, polymerases, genes and even when comparing nascent to total RNA, the latter highlighting the importance of proofreading (Table 1). In fact, one study of nascent transcripts reported that a staggering 1–3% of elongation complexes captured by this method were misincorporated, far in excess of any transcriptome-wide error rate [9]. In eukaryotes RNAPII commits the fewest errors and RNAPIII the most, but both appear more accurate than their bacterial counterpart [7]. The cause and consequences of these differences remain to be discovered.

Table 1:

Cumulative transcription error rate per transcribed nucleotide

Organism	Error rate	Method
C. elegans [15]	5.3 × 10−6	Rep-seq
Buchnera aphidicola [18,102]	7.7 × 10−5	CirSeq
E. coli MG1655 [18,102]	9.9 × 10−5	CirSeq
S. cerevisiae [8]	5.6 × 10−5	ARC-seq
S. cerevisiae [7]	4.9 × 10−6	modCirSeq

The cumulative wild-type transcription error rate per transcribed nucleotide for each organism in vivo includes all base substitution and indel events. All methods sequence multiple repeats of an RNA fragment to generate a consensus sequence revealing true epimutations and to negate errors inherent to reverse transcription, PCR and Illumina sequencing. Buchnera aphidicola is a low G+C content (26%) bacterial endosymbiot whose genome is enriched with long tracts of adenosine and thymine; E. coli has a G+C content of 50.8%. The E. coli error rate is the average transcription error rate over all conditions tested (tryptic soy broth complex media and M9 glucose minimal media each sampled at midlog and stationary phase), since there were no significant differences observed in overall transcription error rates between any pair of individual conditions tested. The base substitution errors rates between E. coli and Buchnera, 8.23 × 10⁻⁵ and 4.67 × 10⁻⁵, respectively, were found to be significantly different [18]. Moreover, a 10-fold increase in transcription insertions was found in Buchnera compared to E. coli that can be ascribed to the incidence of homopolymeric runs in the two genomes [102]. The two S. cerevisiae error rates were determined from log-phase yeast. For comparison with the cumulative transcription error rate for E. coli MG1655, the cumulative in vivo mutation rate for MG1655 is 2.2 × 10⁻¹⁰ mutations per nucleotide per generation as determined by whole-genome sequencing [103]. (Rep-seq, replicated sequencing; CirSeq, circle sequencing; ARC-seq, accurate RNA consensus sequencing; modCirSeq, modified circle sequencing).

2.2. Spontaneous Errors

RNAPs are error-prone and regularly make spontaneous errors, which are defined as those errors arising solely due to intrinsic properties of the polymerase. An important factor regulating spontaneous errors is the thermodynamic stability of the RNA:DNA hybrid within the transcription bubble. Screens for RNAP fidelity mutants along with biochemical approaches have revealed key domains in the Rpb1 and Rpb9 subunits in yeast [10,11] or in the β and β’ subunits in E. coli [12–14] that modulate spontaneous errors.

There exists no clear epimutational signature for the detection of spontaneous transcription errors. However, globally G>A errors are most common prior to proofreading, after which C>U transitions predominate in RNAP I and II transcribed genes [7,9,15–18]. Cytosine deamination likely contributes to C>U bias as this pattern is absent from nascent transcripts [9]. Curiously, in RNAPIII transcribed genes G>A errors remain common in the presence of general fidelity factors, supporting the premise that RNAPIII is naturally more error-prone or resists proofreading.

Once thought to be purely stochastic, new insights have emerged from patterns in which spontaneous errors occur. For example, in addition to their role in transcription termination, poly-T tracts trigger RNAP backtracking with the potential to induce indels through template slippage, and poly-A tracts have a similar effect [19]. The anti-termination λ N factor inhibits transcription slippage at poly-T sequences by stabilizing the RNA:DNA hybrid at the 5’ end, and thus preventing loss of register between transcript and template [20].

Less characterized is the overlap between polymerase pausing and spontaneous transcription errors. While errors are known to trigger pausing, the details of where and when remain obscure in vivo. Using new NGS technologies (e.g. NET-seq and RNET-seq), routine RNAPII pausing in gene bodies has been observed [21]. Errors are enriched at paused/backtracked transcription complexes, as expected [16]. But interestingly, pause sites are not randomly distributed and instead correlate with specific sequence motifs, suggesting that errors also may not occur randomly [9,21]. Whether purported pause site motifs directly induce RNA errors is unknown. Perhaps stalled or backtracked RNAP complexes can themselves trigger errors through a low-fidelity conformation, harkening back to the chicken-or-egg conundrum: Which came first – the stall or the error? Importantly, this logic does not address other gene regulatory mechanisms for RNAP pausing within gene bodies, which likely factor in. Ultimately, these patterns, while interesting leads, have no power to finely predict where an error will occur and cannot supplant stochasticity as the chief mechanism for spontaneous transcription errors.

2.3. Induced Errors

DNA damage is a potent source of transcription errors, but not all damage is created equal. Bulky lesions such as DNA adducts inhibit elongation and require repair to avoid premature transcription termination [22]. Translesion RNA synthesis can occur across deaminated (uracil), oxidized (8-oxoguanine) and methylated (O⁶-methylguanine) bases as well as abasic sites to ensure full-length transcripts, but at the cost of accuracy [23,24]. In one example, transcription across UV-induced pyrimidine dimers induced misincorporation events and multiple nucleotide deletions [25]. This process by which DNA damage generates RNA errors with potential phenotypic consequences is referred to as transcriptional mutagenesis [26,27].

Many essential questions regarding the mechanisms of transcription errors remain. Future studies should aim to address the relationship between RNA errors and gene expression, DNA replication, GC content, poly-A:T tracts, telomeric/centromeric location, CpG islands, methylation state, chromatin remodelers, R-loops, common fragile sites and evolvability.

3. Detection of transcription errors

Perhaps the earliest error rate for RNAP was reported in vitro in 1975 using holoenzyme purified from E. coli [28]; however, studying these errors in vivo remained an intractable problem for decades as the field awaited new technologies. Transcription fidelity research was limited to cleverly engineered reporter systems and complementation screening in single-cell model organisms like E. coli and yeast. Some systems leveraged gene networks with positive feedback loops, like the lac operon in E. coli, to capture transient transcription errors through epigenetic phenotypic switching [29]. Another approach incorporated a Cre/Lox gal reporter system that rendered E. coli Gal⁺ following a site-specific transcription error [30]. While these approaches are efficient methods to screen for transcription fidelity modifiers, epigenetic heterogeneity and heritability, they provide only indirect evidence for transcription errors and limited information on the type of errors involved.

Following the advent of NGS platforms, transcription errors remained technically challenging to sequence as individual events are rare, transient, and RNA requires reverse transcription (RT) into cDNA – itself an error-prone process – for analysis. Early attempts did not control for RT errors [17], which was solved by barcoding individual RNA fragments prior to multiple rounds of cDNA synthesis and termed replicated sequencing (Rep-seq) [15]. By aligning families of cDNAs sharing the same barcode, true RNA errors present in all copies could be resolved from artifact. However, independent rounds of reverse transcription led to inconsistent cDNA synthesis and low yields [15]. These technical hurdles were addressed with the development of circle sequencing (CirSeq), a technique that generates tandem cDNA repeats from a single mRNA fragment to overcome RT and DNA amplification artifact [31,32]. (Reviewed in [33].)

The CirSeq library preparation protocol was recently modified (modCirSeq) to replace the metal ion-based RNA fragmentation strategy, which artificially increased the detected error rate in previous attempts, with an enzymatic reaction [7]. Independently, CirSeq was also adapted for easy discrimination between PCR duplicates, RNA fragments and independent samples via barcoding and named Accurate RNA Consensus sequencing (ARC-seq) (Fig. 1) [8]. Both library preparation methods were successfully applied in yeast, yielding the most accurate transcription error rates reported in eukaryotes to-date [7,8]. For wild-type S. cerevisiae in log phase, ARC-seq provided an RNA epimutation rate of 5.6 × 10⁻⁵, and modCirSeq provided a rate of 4.9 × 10⁻⁶ (Table 1). It will be interesting to see how this 10-fold difference in the cumulative RNA error rates for yeast is resolved; differences in RNA fragmentation (zinc-based vs enzymatic approaches) and different custom-built bioinformatic pipelines are possible contributing factors. In sum, these exciting new technologies have rendered a prohibitively challenging field accessible to any lab with access to NGS. We predict ARC-seq and modCirSeq will be leveraged to compare transcription error rates between species, individuals, cell types, genomic loci, and states of health and disease, yielding discoveries that could revolutionize the field of epigenetics.

Fig. 1.

Accurate RNA Consensus sequencing (ARC-seq). (1) Barcoded RNA adaptors are ligated to a pool of fragmented RNA molecules to uniquely identify individual RNA fragments. Ligated RNAs are then circularized and subjected to rolling-circle reverse transcription generating a cDNA containing tandem multiple copies of each RNA molecule (each copy is denoted by green, red and blue lines in the cDNA). (2) AscI primers are annealed to the cDNA multimer that is restricted into monomers that are copies of the original RNA fragment. (3) Each monomer is tagged with a unique index to identify the different monomer copies and then amplified and Illumina sequenced. Red circles denote the epimutation event; black and green circles are monomer-specific reverse transcription errors; blue circles are amplification/NGS errors. (4) By aligning the cDNA tags and RNA barcodes a consensus sequence is generated from the cDNA copies and the original epimutation is revealed [8]. Whole-transcriptome data is obtained using a primer against the RNA adaptor during rolling-circle reverse transcription; specific locus data is obtained using a transcript-specific primer at this step. Rep-seq and CirSeq have been recently reviewed [33].

While tandem repeats perform well at differentiating RNA errors from sequencing artifact, these rolling circle-based technologies do have limitations. Firstly, they cannot distinguish transient RNA errors from rare de novo DNA mutation. In sequencing DNA from pooled cells, discovery of a rare variant is unlikely, rendering DNA validation of an RNA error exceedingly difficult. However, mathematically, the odds of encountering a genuine RNA error (~10⁻⁵/10⁻⁶ per base) are many logs greater than encountering a rare de novo DNA variant. Another caveat of modCirSeq/ARC-seq is the inability to differentiate errors from RNA editing, as occurs when APOBECs and ADARs catalyze nucleotide changes within RNA [34]. However, many edits follow recognizable patterns that can be excluded from ARC-seq analysis; for example, >99% of edits occur in Alu repeats and the majority are A-to-I, which appear as A-to-G transitions by RNA-seq [34]. To work around issues of post-transcriptional modifications, new sequencing technologies that select for nascent transcripts may be utilized.

Native elongating transcript sequencing (NET-seq) pairs RNAPII immunoprecipitation, allowing for 3’ end capture of nascent RNAs, with NGS [21]. With NET-seq, genome-wide patterns for RNAP pause sites have been described and linked to misincorporation events that are remarkably common, occurring in 1–3% of elongation complexes – much higher than the rate of errors in the total RNA pool [9,21]. NET-seq can be optimized to differentiate paused from backtracked RNAPs with the addition of an RNase footprinting step (RNET-seq) as each conformation “protects” nascent RNAs of different sizes [16,35]. However, the sensitivity of RNET-seq relies on sequencing small RNAs 14–16 bp in length, which tend to map poorly on analysis, limiting its application. Ultimately, the significance of paused versus backtracked RNAP in epimutation and transcription conflicts is not well understood. New scalable in vivo methods of detection are needed to distinguish the various inactive states of RNAP to facilitate further study.

4. Proofreading of transcription errors

Elongation occurs through a translocating transcription bubble containing a critical RNA:DNA hybrid of 8–9 bp in length on the antisense DNA strand [36]. Importantly, processive elongation requires an active complex and does not occur when RNAP is paused, backtracked or arrested – all alternative, but distinct, inactive conformations of RNAP [37,38]. Spontaneous errors during elongation such as base mispairings, misalignments and indels destabilize the RNA:DNA hybrid [36]. This disruption impedes elongation and energetically favors inactive conformations such as backtracked RNAP in which the complex slides backwards along the DNA, causing the 3’ end of the nascent transcript to disengage from the catalytic site and thread through the secondary channel [39]. Backtracked RNAP alone is stable with limited proofreading of misincorporated nucleotides through low-level intrinsic cleavage of the 3’ end of nascent transcripts [40]. However, addition of transcription fidelity factors GreA/B (E. coli) or TFIIS (eukaryotes) greatly increases the rate of RNA cleavage, allowing the 3’ end to efficiently re-engage the catalytic site and elongation to continue [41–44]. Besides Gre and TFIIS, Gre-homologue DksA also enhances transcription accuracy; dksA deletion confers a 7-fold increase in errors as measured by a lac reporter system in E. coli [45,46]. These transcription fidelity factors are highly conserved across species, demonstrating their functional importance.

Transcription errors can also occur secondary to DNA damage such as pyrimidine dimers and DNA adducts [25,47]. In this case, the faulty template may cause RNAP to pause, backtrack or arrest, serving as a damage signal to repair mechanisms [48]. Anti-backtracking factors can restart elongation and promote error-prone translesion synthesis (transcriptional mutagenesis) [23], but some lesions, especially bulky ones, cannot be bypassed and require repair to avoid premature transcription termination [49]. Complex damage-sensing and repair pathways have evolved with transcription – collectively called transcription coupled repair (TCR) – that have been thoroughly reviewed elsewhere [22,50].

Regardless of where or why a stalled RNAP complex occurs, a dynamic equilibrium is thought to ensue between pro- and anti-backtracking forces [51]. Promoting backtracking, the repair helicase UvrD works to pull the polymerase backwards [52] while secondary channel interactors DksA and ppGpp may destabilize active RNAPs [53–55], together disrupting elongation. Conversely, Mfd helicase pushes RNAP forward [56,57] while GreA/B oppose backtracking through their cleavage function at the secondary channel where they compete with DksA [58]. Given their opposing functions, these backtracking factors likely work through independent pathways to promote fidelity. A growing body of literature suggests that anti-backtracking factors evolved to mitigate spontaneous errors and promote elongation to enhance processivity [9,16] but at sites of DNA damage these factors impede repair [22,59]. On the other hand, pro-backtracking factors appear important for DNA repair as a ppGpp null mutation renders cells sensitive to genotoxic agents [53] and uvrD is induced by the SOS response to DNA damage in E. coli [60]. How transcription fidelity factors coordinate at different types and sources of inactive RNAP complexes will likely be a hot topic for future research into the mechanisms of transcription:DNA repair conflicts.

4.1. GreA/B

In E. coli, GreA and GreB proteins interact with stalled RNAPs through the secondary channel to induce RNA cleavage and relieve backtracking, with cleavage fragments reportedly ranging in size from 2 to 18 nucleotides [61]. At misincorporated complexes, particularly G>A errors [9,17,62], Gre-mediated cleavage removes the offending base that triggered the stall and repositions the new 3’ RNA end at the active site, thus Gre proteins play dual roles as proofreading and pro-elongation factors. Loss of GreA alone confers no significant phenotype under normal growth conditions, however, when combined with a greB null mutation strains become temperature and starvation sensitive with pronounced replication blockage [63]. Confirming their in vivo role in fidelity, Gre knockouts have increased rates of transcription errors (2.4 to 100-fold) as measured by reporter screens [30,64] and NGS [9,17,62]. These studies depict a role for GreA/B in maintaining high-quality, full-length transcripts, especially under harsh conditions with strong selective pressures for survival.

4.2. TFIIS

Originally described as an activator of RNAPII, TFIIS was discovered to promote transcription fidelity in vitro [42] six years after the TFIIS analogue GreA was identified in E. coli [65]. Most recently, the fidelity function of TFIIS has been validated in yeast by modCirSeq as loss of TFIIS increases transcription errors 10-fold (~4 × 10⁻⁵ per bp) with no confounding effect on the DNA mutation rate [7]. Errors in ΔTFIIS strains showed strong bias for G>A transitions, but rates of other misincorporations, insertions and deletions are affected as well [7]. TFIIS is not essential in yeast [66] or C. elegans [67], whereas in mice TFIIS null mutants are embryonic lethal [68]. In humans, TFIIS has three homologues – TCEA1, 2, and 3 – which differ in tissue distribution but share functionality [69].

Functionally, TFIIS behaves like GreA, inducing nascent transcript cleavage to relieve RNAP backtracking and restart elongation across the eukaryotic domain [42,43]. However, many studies support a role for TFIIS in gene regulation and promoter-proximal pausing as well [67,70,71]. Using a dominant-negative TFIIS mutant that inhibits RNA cleavage and stabilizes backtracked complexes, it has been shown that rescue of backtracked RNAPII by TFIIS is a major stimulus of elongation under normal conditions [72]. Additionally, loss of TFIIS in yeast effected a global 5–18 bp shift in pauses sites observed by NET-seq [21]. Important questions remain regarding which function of TFIIS – pro-elongation or proofreading – contributes most to the phenotypes observed in eukaryotes.

5. Consequences of transcription errors

Transcription fidelity is hypothesized to affect cells through two primary mechanisms: genetic and epigenetic. The relative contributions of each to the phenotypes observed in fidelity mutants in vivo remains elusive.

5.1. Genetic consequences of transcription fidelity

There is extensive evidence that conflicts between transcription and DNA replication and repair are under negative selection and therefore detrimental to cells. From bacteria to humans, highly expressed genes have evolved to be co-oriented with the direction of replication to avoid the mutagenic effects of head-on conflicts between RNAP and the replisome [73,74]. The timing of replication in eukaryotes is tightly regulated through the actions of cyclins, checkpoints and transcription factors to minimize local transcription in replicating regions [75]. In mitochondria transcription and replication appear to be mutually exclusive processes [76]. Additionally, transcriptional repression has been widely observed following DNA double-strand breaks (DSBs) or UV-induced DNA damage [77,78]. (For reviews of transcription:replication conflicts see [79–81].)

As fidelity factors regulate both RNAP elongation and TCR of polymerase-stalling lesions, they influence the global timing of transcription and thus affect the likelihood of genome-destabilizing conflicts with the replisome. Significant evidence supports roles for GreA/B and TFIIS, among others, in DNA damage and repair independent of their effects on transcript quality [82]. Increased RNAP backtracking secondary to loss of GreA may even promote DNA damage [83]. However, recent evidence suggests that Gre proteins antagonize DNA repair. A greA null mutant was shown to enhance survival by 10-fold following restriction enzyme-induced DSBs and by 500-fold after phleomycin treatment [59]. Additionally, loss of GreA/B rescues ppGpp and uvrD null mutants from DNA damage susceptibility, highlighting the opposing forces of pro- and anti-backtracking factors in genome stability [52,53]. Increased survival in ΔgreA E. coli is dependent on the homologous recombination enzymes RecA and RecBCD [59]. and on the Holliday junction resolvase RuvABC [84], implicating DNA repair in the mechanism of rescue. Interestingly, expression of a backtracking-prone RNAP mutant confers the same rescue as the greA null [59,84]. Taken together, these studies depict GreA as a negative regulator of TCR and break repair at bulky DNA lesions. Perhaps at minor roadblocks the anti-backtracking activity of GreA instead promotes translesion synthesis, enabling production of full-length transcripts despite damage. Such duality in GreA function at different types of DNA damage has not been investigated but may be of future interest.

While Gre factors are the subject of an extensive body of research, their human counterparts remain relatively understudied, likely due to both inadequate tools and the lethal phenotype in knockouts [85,86]. In contrast to GreA, the role of TFIIS in DNA repair remains elusive. TFIIS is recruited by the TCR complex and restores transcription following UV damage, but TFIIS knockdown has no effect on UV sensitivity or the p53 damage response in human cells [87,88]. The advent of new sequencing and gene knockdown technologies will enable further study of this essential transcription factor, promising to yield novel insights in gene regulation, aging and human disease.

The profound implication that RNA errors may pose a threat comparable to DNA damage in cells suggests an evolutionary tradeoff between transcription fidelity and DNA repair. Given that transcript quality regulates cell state through proteins and noncoding RNAs, perhaps the repercussions of erroneous transcripts on survival can at times outweigh the deleterious effects of DNA damage. For instance, the DNA damage response (DDR) itself is a transcriptional program. To enable global repair and maximize the chances of survival following damage, DDR genes must be transcribed even at the cost of impeding repair at local lesions.

Given the role of DNA damage in disease, the competing priorities of repair and transcription lend credence to a new paradigm that transcription errors might also play a significant role in disease.

5.2. Non-genetic consequences of transcription fidelity

At ~10⁻⁵/10⁻⁶ per base, the rare and transient nature of transcription errors has led them to be overlooked by many. To fully appreciate their consequences, the combined impact of errors in a cell must be abstracted, but first we will consider the effects of a single, individual error.

As with DNA mutation, most RNA errors are silent, but errors can still impact cells in many ways. First, errors can affect critical steps in post-transcriptional processing like splicing [89]. At the protein level, misincorporation errors within key catalytic, regulatory or binding domains can completely abrogate normal protein function, with misalignment errors conferring even more dramatic effects. Additionally, a single transcript is translated multiple times amplifying a single error. For low copy number transcripts and long-lived proteins, the effects of errors are particularly amplified and may persist long after the error occurred. What’s more, transcription errors can exert dominant effects on the protein pool by titrating away binding partners from WT copies or poisoning multimeric complexes, disrupting entire functional networks, which has been directly observed in vivo. Using an engineered TP53 construct with a site-specific 0⁶-methylguanine that could induce a dominant negative epimutation through transcriptional mutagenesis, a staggering 15% of cells displayed impaired cell cycle arrest and apoptosis, illustrating the power of a single error to derail normal functions with potential long-term consequences [90]. However, unlike transcriptional mutagenesis where DNA lesions cause predictable changes [23], spontaneous transcription errors are more random and – to our knowledge – recurrence of a specific error is rare. So, the question stands – How can a rare, transient, one-time transcription error impact cell fate?

There is evidence in bacteria, yeast and human cells for the regulation of critical gene networks, from metabolism to cell cycle control, by bistable transcription switches [91,92]. Switches operate as gene expression in one state – ON or OFF – with a defined threshold to switch states greater than the threshold to maintain the current state. If an inducer, e.g. lactose in the E. coli lac operon, exceeds the threshold the gene is switched ON, and once ON positive feedback loops stabilize the induced gene until a counteracting repressor exceeds threshold and returns the cell to the OFF state. Despite their simplicity, bistable switches are reminiscent of Waddington’s famed epigenetic landscape of development and comfortably fit contemporary models of cellular plasticity. Switches can be designed with increasing complexity, but the essential principle remains that genes in bi- or multi-stable switches require only fleeting signals to stably and heritably alter cellular phenotype (Fig. 2) [64]. In bistable models of gene regulation, transient transcription errors have the potential to dramatically alter cell fate through stochastic phenotypic switching. In support of this model, loss of fidelity factors GreA/B increases the rate of heritable phenotype switching as measured from a lac reporter [29].

Fig. 2.

Positive feedback and all-or-none phenomena. (A) Left panel: The lac operon comprises an autocatalytic positive feedback loop (the presence of LacY permease will produce more permease) allowing a heritable epigenetic switch at a maintenance concentration of inducer. LacY is presented in green because the lacA gene in our system is replaced by gfp, and the lac operon is now lacZYA::gfp, so when permease is made GFP will also be made and the cell will fluoresce green. Stochastic events that lead to a transient depletion of repressor (in red) within a cell will result in a burst of LacY permease expression and the presence of permease will activate the positive feedback loop, so that the new induced state will be heritably maintained, mimicking mutation [29,33]. Right panel: E. coli microcolony growing in a microfluidics chamber. Lac operon Off cells were seeded and grown in minimal media plus maintenance concentration of inducer. A stochastic phenotypic switch from Lac OFF to ON occurred in a single cell and the phenotype was heritably maintained forming an ON lineage of cells. The ON lineage was torn apart due to the growth dynamics of the microcolony. The brightfield image viewed in the blue channel was merged with the GFP fluorescence image viewed in the yellow channel. (B) Left panel: A bistable RB-EF2 switch underlies the restriction point, the critical event when a mammalian cell commits to proliferation and becomes independent of growth stimulation [92]. In OFF, quiescent, cells EF2 (in green) is repressed by RB (in red). With sufficient growth stimulation by Myc, Rb repression is removed via CycD/Cdk_4,6 activation; Myc also induces EF2 which activates CycE/Cdk₂ to further block Rb. Since EF2 activates itself, two positive feedback loops are involved to shut OFF Rb and turn ON EF2 resulting in a pathway that converts graded serum inputs into an all-or-none-response. Right panel: Cellular heterogeneity in a tissue. The variation in blue represents underlying sub-threshold expression levels of a gene network in individual cells; full-blown yellow represents the activation of a network and subsequent heritable epigenetic phenotypic switch in a cell lineage. The cellular pattern is based on an M.C. Escher Alhambra tessellation sketch.

Transcription occurs on enormous scales and is constantly fueling new errors, rendering a rare event frequent enough to arise many times over the lifetime of a cell. The odds of a phenotypically relevant error occurring are further inflated in long-lived, quiescent cells such as neurons, mirroring a cumulative risk model akin to other disease models for neurodegeneration and cancer. Together, RNA errors contribute to molecular noise [29] that can impact proteostasis and induce heterogeneity between cells [7]. Non-genetic heterogeneity plays a central role in normal development, disease and evolution, but the precise contribution of transcription errors to these essential processes is still unknown [93]. With such broad connections across biological phenomena, transcription errors are now more relevant than ever, and further study promises to inform on the most fundamental processes within living systems, including epigenetic regulation, genome maintenance, disease and evolution.

6. Perspectives in evolution and disease

There are many possible mechanisms by which transcription errors may contribute to disease, especially diseases driven by non-genetic heterogeneity and genomic instability (Fig. 3). For example, it was recently demonstrated that stochastic expression of the DNA repair protein Ada generates mutation rate heterogeneity in a monoclonal population of E. coli [94]. Transcription errors could potentiate stochastic Ada activity and thereby drive bacterial evolution and antibiotic resistance. In another example, RNA errors were found to increase with age, induce proteotoxic stress, and shorten the lifespan of yeast, linking transcription infidelity to aging and neurodegeneration [95]. In fact, neurobiologists have studied the effects of translation errors (“molecular misreading”) on protein heterogeneity, proteotoxic stress and disease for years with little recognition of parallel mechanisms for transcription errors [96,97]. Despite these interesting associations, few disease models have been rigorously tested and most hypotheses rely primarily on indirect evidence. However, no disease has yielded more mechanistic clues regarding the role of transcription errors than cancer [98].

Fig. 3.

Transient and permanent phenotypic consequences of an RNA transcription error. Green circle is RNAP; black lines are DNA; blue line is nascent RNA transcript; red explosion is an RNA transcription error.

From its origins to later stage events like drug resistance and metastasis, transcription fidelity may propel cancer development. Promoting oncogenesis, transcription errors in p53 or the Rb/E2F bistable switch can transiently dysregulate DNA repair or induce cell cycle progression, propagating mutations before they can be fixed [90,92]. Moreover, RNA errors could contribute to observed fluctuations in p53 signaling, which has been shown to facilitate cell cycle escape [99]. Following malignant transformation, transcription errors may promote phenotypic heterogeneity among cancer cells and fuel tumor evolution.

However, low-fidelity transcription likely exerts negative consequences in cancer cells as well. For example, erroneous transcripts generate molecular noise that adversely affects cellular fitness [95]. The fitness cost of transcription errors may explain why TFIIS knockdown inhibits proliferation and induces apoptosis in a breast cancer cell line [85]. Given their elevated levels of replication and transcription, cancers could also be particularly sensitive to the genetic consequences of transcription errors and RNAP backtracking as increased traffic on the DNA likely promotes genome-destabilizing collisions [100,101]. In summary, transcription errors are highly relevant to the dysregulated gene networks, molecular noise, and genomic instability observed in cancers and thus warrant further study into their contributions to disease.

7. Conclusion

The transcription fidelity field, yet in its infancy, is flush with creative ideas and primed to emerge on the research scene as a new paradigm in disease and evolution. Clearly, much remains to be learned about the basic biology of transcription errors, including their role in phenotypic heterogeneity, adaptability and cellular evolution. By applying the cutting-edge concepts and tools outlined in this review, researchers can begin to tackle critical questions in the field and explore mechanisms for transcription fidelity in antibiotic resistance, aging, neurodegeneration, cancer and beyond.

Abbreviations:

ARC-seq

accurate RNA consensus sequencing

CirSeq

circle sequencing

DDR

DNA damage response

DSB

DNA double-strand break

modCirSeq

modified circle sequencing

NET-seq

native elongating transcript sequencing

NGS

next-generation sequencing

Rep-seq

replicated sequencing

RNAP

RNA polymerase

RNET-seq

NET-seq with RNase footprinting

RT

reverse transcription

TCR

transcription-coupled repair

TFIIS

transcription factor S-II