Facts

  • Transcriptional memory is defined as a phenomenon where transient gene activation by an environmental signal results in a heritable primed state that is maintained in the absence of active transcription.

  • Transcriptional memory phenomena occur in a wide range of organisms such as yeast, Drosophila and mammals and are triggered by a variety of signals, including stress, nutrients and pathogens.

  • Transcriptional memory lasts for days to weeks but is fully reversible.

  • Key molecular features associated with primed genes are transiently poised RNA polymerase II, maintenance of histone H3 lysine 4 methylation and interaction with nuclear pore proteins.

Open Questions

  • How do different mechanisms and levels of control integrate with each other?

  • How do the carriers of transcriptional memory engage in positive feedback and are mitotically inherited?

  • What is the adaptive value of transcriptional memory?

Introduction

Transcriptional memory is a cellular phenomenon in which an environmental stimulus triggers a heritable primed state, resulting in a secondary response to the stimulus that is distinct from naive cells. The memory of previous signals is inherited during cell divisions, even in the absence of ongoing target gene transcription. While memory can last for days to weeks of continued proliferation it is ultimately reversible [1,2,3]. This phenomenon is observed across broad domains of life, including prokaryotes [4] and in many eukaryotes - the focus of this review. In budding yeast, transcriptional memory is best defined at the inositol starvation-induced INO1 gene and the galactose-inducible GAL1 gene lasting for several cell generations [5, 6]. In Drosophila embryos, transcriptional memory is observed upon ecdysone-induction of target genes [7]. Plants are generally sessile organisms, exposed to a vast array of environmental challenges and have evolved elaborate mechanisms to “remember” their surroundings. Long-term and transgenerational epigenetic memory phenomena are widespread in plants [8] but more relevant to this review, plants also display transiently memorized responses to environmental cues such as temperature shock, high salt, dehydration, and infection, which leads to a modified response to recurrent stress [9,10,11,12]. In human cells, transcriptional memory is observed in response to inflammatory cytokines such as TNF-α and interferons in epithelial cells [13], fibroblasts [14], cancer cells [3, 15, 16] and, importantly, macrophages [2, 17]. The fact that innate immune effectors such as macrophages can be primed raises the possibility that such memory contributes to trained immunity, a form of memory in the mammalian innate immune system [18, 19].

Here, we will review the basic phenomena of transcriptional memory and dissect to what extent molecular mechanisms have been uncovered to explain memory. In the final section, we will discuss the key outstanding question including a discussion on the physiological and adaptive function of memory.

What is the physical basis for transcriptional memory?

In principle, any heritable differential effect in expression following priming can be categorized as a form of transcription memory. However, transcriptional memory typically refers to priming of an active state leading to enhanced future expression (Fig. 1a) and is the subject of this review. It should be noted that an initial stimulus can also result in a decreased expression upon re-exposure to the stimulus, or genes can become refractory to reinduction (Fig. 1b). The latter, referred to as tolerance has been described principally for human or mouse cells that are challenged with lipopolysaccharide (LPS) [20, 21] leading to a subsequent refractory period and decreased in anti-inflammatory cytokine production [22, 23]. The molecular mechanisms underlying tolerance include receptor desensitization and incorporation of repressive histone marks but are not the focus of this review and are discussed in more detail elsewhere [23, 24].

Fig. 1: Definition of transcriptional memory.
figure 1

a Transcriptional memory is defined as a heritable alteration in the transcriptional response following initial priming by an external stimulus. Most commonly, this involves an enhanced expression following initial priming, which is the primary subject of this review. b Tolerance is also a form of memory where, following priming, cells become refractory to the response (see refs. [23, 24]). These forms of transcriptional memory contrast with cellular differentiation c in which a transcriptional change is perpetuated, e.g., by the alternation of a transcription factor network.

Full size image

Generally, the molecular principles of the epigenetic maintenance of a repressed state have been well studied [25,26,27,28]. However, explaining the heritable maintenance of an active state has proven challenging. This difficulty arises from the absence of a well-defined molecular mechanism for the maintenance of active histone marks and the lack of proper experimental models that uncouple the active state while avoiding confounding effects of transcription. It is important to note that transcriptional memory is distinct from cellular differentiation which is also heritable but involves a rewiring of transcriptional programs resulting in sustained expression of target genes (Fig. 1c).

Here, we will focus on priming effects that lead to the inheritance of an active state which, in turn, results in enhanced expression during future encounters with the signal. Two broad concepts can be considered. First, the carriers of memory which can be broadly categorized as “trans-acting” or “cis-acting” factors. Second, the nature of memory which either, involves promotion of an active gene state or alternatively, alleviation of a repressive state during priming.

Possible trans-acting mechanisms

Passive versus active mechanisms

A memory of previous environmental cues may be maintained by soluble factors that act systemically in cells as “trans-acting” factors. The most trivial mechanism is the “spill-over” of stable soluble factors. For instance, cytoplasmic pools of a trans-activating factors may be passively passed on to daughter cells. This may be leading to sustained expression of target genes (Fig. 2a) which has been suggested to underpin at least part of galactose memory in yeast [29]. However, such a passive mechanism does not account for long-term mitotically stable memory in which initial pools would be significantly diluted.

Fig. 2: Possible trans-acting factors that contribute to transcriptional memory.
figure 2

Box: General phenomenon of transcriptional priming, resulting in enhanced re-expression of primed genes. Trans-acting factors that may contribute to memory are: (a) cytoplasmic inheritance “spill-over” of transcription factors that mediate gene activation in daughter cells, e.g., ref. [29], (b) continued expression of co-factors that collaborate with the external inducer to activate memory genes, e.g., refs. [37, 38], (c) changes in the dynamics of upstream signaling cascades involved in priming, (d) changes in post-transcriptional stability of transcripts leading to elevated expression [42], or (e) prion-like propagation of gene expression regulators [44] (see text for details).

Full size image

Role of trans-acting transcription factors in transcriptional memory

A more elaborate mechanism involves a positive feedback loop of sustained expression of a trans-acting factor (e.g., a transcription factor), allowing for a more rapid transcription re-activation of its target genes. If sufficient for target gene expression, it would essentially constitute transcriptional rewiring of cells which is the basis of cellular differentiation [30,31,32]. However, such a mechanism may still contribute to transcriptional memory if the maintenance of transcription factor expression by itself is not sufficient to activate target genes, requiring an external inducer (Fig. 2b). For instance, in yeast, Tup1 promotes the incorporation of the histone variant H2A.Z and H3K4me2 modification at the GAL1 promoter, required for memory to galactose exposure [6]. Similarly, Sfl1 and Hms2 are required for INO1 memory in yeast. Both bind to a cis element required for memory called the Memory Recruitment Sequence and their loss disrupts transcriptional memory [33, 34]. HSFA2 has been shown to have a role in heat stress memory in plants. While the initial response to stress was not significantly impacted in hsfa2 mutants, reinduction of target genes was reduced [12]. Further, transcription factor MYC2, which is induced upon dehydration is required for dehydration stress memory [35]. In mammals, the immediate early transcription factor AP1 FOS/JUN has been shown to mediate long-term inflammation, in part by retention of JUN on chromatin in the memory state [36]. Similarly, the polarization of macrophages depends on both cytokines and the lineage-specific transcription factors PU.1 and C/EBP [37, 38].

Furthermore, establishment of IFNγ-induced transcriptional memory is STAT1-dependent but this factor is dispensable for the maintenance of the primed state [39]. Importantly, STAT1 induces many genes upon IFNγ induction that do not display memory, indicating it is not the sole driver of memory in this system. More generally, there is a growing body of evidence indicating that transcription factors such as FOX1A and GATA1 can remain bound to their target sites during mitosis. This phenomenon, called bookmarking may contribute to transcriptional memory [40]. It should be noted that, while transcription factors are contributors to memory, they are likely not sufficient and unlikely to be the sole carriers of memory. Instead, they may cooperate with local chromatin-based mechanisms to execute memory (see below).

Priming of pre-transcriptional or post transcriptional processes

At the most upstream level, memory of a signal could potentially be the consequence of alterations in the signaling machinery that is involved in detecting the initial signal (Fig. 2c). This could include changes in receptor concentration and hyperactivation of signal transducers. However, there is little evidence to support such models. In IFN-β transcriptional memory, the cytokine receptor and STAT1 activation rate are not altered in naive and primed cells [2]. Similarly, the rate of IFNγ-induced STAT1 activation and import are similar in naive or primed cells [39], arguing against a role in memory. Additionally, following priming by TNFα, the transcriptomes of naive and primed cells are similar except for genes displaying memory, indicating that priming did not change the general physiology of cells [41].

Another conceivable mechanism of transcriptional memory involves a change in mRNA or protein stability of target genes upon re-exposure to the stimulus, which would also lead to an increase in overall gene output (Fig. 2d). In support of this, changes in mRNA stability contribute to memory at the yeast GAL genes [42]. Even further downstream, one can envision heritable changes in the biophysical properties of effector proteins or formation of prion-like protein assemblies that result in a change in transcriptional output during re-exposure (Fig. 2e) [43]. For instance, a yeast chromatin modifier complex has been shown to have prion properties that can change transcriptional state of silent domains [44]. While an attractive mechanism, this has not yet been tied to transcriptional memory of an inducing signal.

Cis-acting mediators of memory

In addition to “trans-acting” factors, transcriptional memory may be regulated locally through “cis-acting” mechanisms. A contribution from local factors is highly likely as often many genes are activated by an external signal but only a subset is memorized, indicative of cis-acting factors specific to those genes.

Retention of RNA polymerase II at primed genes

The most direct form of cis regulation is the maintenance of RNA polymerase on memorized genes (Fig. 3a). Priming of the human HLA-DRA gene upon IFNγ-induction correlates with the maintenance of poised RNA polymerase II at the promoter, even after transcription ceases [16, 45]. Similarly, in Arabidopsis thaliana, RNA polymerase is retained at memorized genes following heat stress [10, 35]. However, other reports did not detect RNA polymerase poising following priming by IFNβ and IFNγ at different genes [2, 3]. Possibly, more sensitive methods are needed to firmly establish whether interferon memory is associated with poised polymerases.

Fig. 3: Possible cis-acting factors that contribute to transcriptional memory.
figure 3

Box: General phenomenon of transcriptional priming, resulting in enhanced re-expression of primed genes. cis-acting factors that may contribute to memory are: (a) retention of non-productive transcription machinery [2, 6, 10, 16, 33, 35], (b), altered chromatin structure [2, 3, 6, 12, 15, 16, 35], (c, d) loss of DNA methylation or repressive histone modifications, e.g., refs. [41, 66, 68], (e, f) changes in long-range chromatin interactions or nuclear position [1, 5, 6, 33, 47, 69] (see text for details).

Full size image

A single cell study of GAL1 priming revealed that the enhanced expression upon reinduction is mainly due to a shorter delay in the response to galactose rather than a change of the expression rates, indicating that transcription initiation commences faster, consistent with polymerase II poising [46]. On the other hand, ongoing transcription itself does not appear to be required to maintain memory, at least for IFNγ-primed genes in human cells [3]. Further, artificial activation of IFNγ memory genes is not sufficient to induce memory [39]. Similarly, blocking transcription during ecdysone induction still allows priming of some genes [47]. Finally, in the INO1 paradigm, while RNA polymerase is promoter-poised, its depletion or inhibition at target genes did not impair memory of prior activation [34]. Thus, while poised polymerase may be present in several memory systems, this is likely correlative and mediated by an upstream factor that is mediating memory.

Retention of active chromatin modifications at primed genes

The comparatively slow dynamics of nucleosomes and their modifications along with pervasive feedback mechanisms make changes in chromatin structure an attractive carrier of transcriptional memory [48] (Fig. 3b). Indeed, primed genes have been found to retain specific histone modifications and histone variants established during priming. Histone H3 lysine 4 dimethylation (H3K4me2), is broadly associated with active gene transcription and is correlated with transcriptional memory in several studies. During yeast galactose and inositol priming, plant heat stress, and human IFNγ induction, H3K4me2 is maintained at the promoter of primed genes [2, 3, 6, 12, 15, 16]. Yeast mutants expressing H3K4R and H3K4A are unable to display INO1 memory, which indicates a functional importance of H3K4 methylation in yeast transcriptional memory [33].

Interestingly, the establishment of H3K4me2 is mechanistically distinct from its maintenance. H3K4me2 is deposited by Set1/COMPASS, an eight subunit methyltransferase complex [49], responsible for H3K4me2 and H3K4me3 deposition. A non-canonical version of COMPASS complex that lacks Spp1, selectively deposits H3K4me2 at INO1 during priming, in a manner dependent on RNA polymerase-associated factors, the transcription factor Sfl1 and Nup100 as well as SET3C, part of a histone deacetylase complex [33]. The mark then remains stable and heritable for four cell divisions in the absence of transcription, even after Sfl1 degradation but relies on continued SET3C presence, which may itself be, in effect, a carrier of memory [34]. Trimethylation of H3K4 is generally enriched at promoters of actively transcribed genes and is correlated with retention of a primed state in several cases of plant transcriptional memory. This includes accumulation of H3K4me3 at dehydration memory genes as well as heat stress memory genes in Arabidopsis [10, 12]. In mouse monocytes, H3K4me3 enrichment is correlated with stronger transcription of the genes primed with β-glucan [50].

However, it is unlikely that either H3K4 di- or trimethylation is solely or even primarily responsible for memory phenomena. For instance, at human GBP genes priming by IFNγ results only in short-term retention of H3K4me2 and H3K4me3 is lost as soon as transcription ceases [2, 3]. In case of retention, such as in Arabidopsis [10], these marks tend to associate with low level transcription or poised polymerase. Thus, determining whether such marks can be drivers of memory or simply reflect ongoing transcription requires careful examination of polymerase levels and nascent transcription in the primed state.

Finally, H3K4me1 is a potentially interesting candidate to propagate memory as it is associated with enhancers, including transcriptionally silent latent enhancers. For instance, LPS stimulation of macrophages results in the transient activation of specific enhancers. While most active chromatin marks associate with these enhancers only transiently, H3K4me1 is maintained, albeit only for a short period of 24 h [17]. Nevertheless, in the context of LPS-induced memory of NK cells, in vivo priming can last 9 weeks, resulting in enhanced IFNγ expression upon restimulation [51]. Interestingly, H3K4me1 is retained at the IFNγ enhancers throughout this period, and the chemical inhibition of this mark disrupts memory. However, as this is an in vivo study, it is not possible to ascertain whether the H3K4me1 mark is maintained by autonomous feedback or by continued transactivation by other cells or factors in primed animals. More recently, another study found that prolonged treatment of cancer cells with IFNγ leads to a STAT1-dependent activation of enhancers. Removal of STAT1, thereby abrogating IFNγ-signaling, resulted in the loss of active enhancer marks such as H3K27Ac but H3K4me1 and chromatin accessibility was maintained in an in vivo tumor transplantation model [52]. These studies suggest that, from the myriad of chromatin changes at enhancers, H3K4me1 may directly contribute to the primed state (Fig. 4a).

Fig. 4: Common elements implicated in transcriptional memory.
figure 4

a Transcription activation of genes prone to priming is associated with establishment and functional retention of H3 lysine 4 methylation. Maintenance of methylation (particularly H3K4me1) may involve feedback between the inherited modification and SET/MLL methyltransferases that create the mark. b Stimulus-activated transcription factors (e.g., phosphorylation of STAT1 in IFNγ induction) recruit active Mediator to promoters, driving transcription. During priming, the Cdk8 module-containing form of Mediator is retained possibly priming the promoter for rapid reactivation. Transcription factors such as STAT1 are also modified by Cdk8, possibly aiding in gene reactivation [61]. c Priming results in recruitment of nuclear pore proteins (NUPs) to promoters (and nuclear pore complex tethering in yeast). NUPs remain chromatin associated which is functionally required for memory following priming (see text for details).

Full size image

Contribution of other chromatin marks and cis-acting non-histone proteins

Upon IFNβ and IFNγ priming in mouse fibroblast, macrophages and cancer cells, histone H3K36 trimethylation as well as the H3.3 histone variant are retained on memory genes following priming but only for up to 2 days [2, 3]. Therefore, whether this retention is independent of residual transcription is not yet clear. More interestingly, in yeast, the histone H2A variant H2A.Z is incorporated at the promoter of primed genes during memory (INO1 and GAL1) in a transcription factor-dependent manner. Loss of H2A.Z leads to loss of nuclear peripheral localization, RNA polymerase II binding and memory [1, 5, 6]. However, depletion of H2A.Z has also been reported to have indirect consequences suggesting memory defects may be the result of downstream effects [53]. In Table 1, we list an overview of histone modifications and their roles in the maintenance of a primed transcriptional state.

Table 1 Summary of factors that correlate with transcriptional memory.
Full size table

Further, in heat stress memory in Arabidopsis, chromatin regulator FORGETTER1 (FGT1) binds to the promoter of heat-inducible memory genes and is required for heat stress memory. FGT1 interacts with SWI/SNF chromatin remodelers BRAHMA (BRM) and ISWI leading to low nucleosome occupancy and subsequently increasing accessibility at promoters of memory genes. Loss of BRM and ISWI results in heat stress memory disruption [54]. While these factors are required for memory, whether they are carriers of memory, or whether their presence reflects heritable chromatin structure is not yet clear.

Finally, Mediator is a multi-protein complex and critical regulator of transcription [55, 56] that includes a Cdk8 kinase module of Mediator (CKM) [57, 58]. It has been proposed that the Cdk8 module of Mediator (CKM) prevents Mediator association with the preinitiation complex (PIC) [58]. Kinase activation releases the Cdk8 module allowing Mediator to bind the PIC, leading to rapid activation of genes [59]. In the context of transcriptional memory, Cdk8 (or its homolog SSN3) remains bound to the yeast INO1 promoter, and IFNγ-primed genes in human cells [33]. At INO1, the Mediator subunit MED13 and SSN8 are also maintained during primed state, suggesting the entire Cdk8 module is maintained [33]. Further, when Cdk8/Mediator complex is present, it lacks canonical Cdk7 [59] and consequently, serine 5 phosphorylation of the RNA polymerase is prevented, blocking productive initiation [16, 33]. This has led to the proposal that maintenance of Cdk8 facilitates assembly of the RNA polymerase in a non-productive form in the primed state but prevents initiation until reinduction [45] (Fig. 4b). Cdk8 may have further relevance in the context of IFNγ-priming. IFNγ activates STAT1 via JAK and TYK kinases at the plasma membrane [60]. It is further stimulated on chromatin by Cdk8 phosphorylation of a serine residue 727, which is required for full STAT1 activity [61]. Interestingly, this phosphorylated form of STAT1 is maintained for multiple days in IFNγ-primed cells [39]. It is, therefore, possible that Cdk8 plays additional roles in memory, beyond regulation of PIC assembly (Fig. 4b).

Combined, these findings suggests that, in addition to histone modification signatures, non-histone proteins may engage in continuous feedback to maintain a dynamic primed state in the absence of ongoing transcription. Whether Mediator directly carries memory or engages in feedback with local retention of chromatin modifications is an important question to answer.

Memorizing what is not there: priming by removal of repressive chromatin

Conceptually, memory of prior gene activation can be achieved not only by promoting a poised active state, as outlined above, but equally by removal of a repressive state that, if heritable, promotes future activation (Fig. 3c).

DNA methylation

DNA methylation is generally associated with repression of transcription initiation and is mitotically stable due to self-templating feedback mechanisms [62, 63]. Its stability is exploited in a unique example of transcriptional memory at TNF-α-induced genes, where the maintenance of an active state is caused by removal of repressive DNA methylation (Fig. 3d). TNF-α activates NF-κB, and genes that show priming tend to be heavily methylated at CpG dense regions carrying NF-κB binding sites. TNF-α mediated NF-κB targeting results in DNA demethylation and the initial density of DNA methylation correlates with the strength of memory [41]. The enhancers of CALCB, a TNF-α-primed gene, are demethylated upon prolonged TNF-α treatment resulting in a more than 100-fold sensitivity to TNFα upon reinduction [41]. In this case, memory appears to be a consequence of the slow kinetics of remethylation of these genes.

In other systems, demethylation of CpG sites at target genes has been shown to be induced under acute exercise in human muscles and retained after the exercise, correlating with faster induction upon subsequent exercise [64, 65]. In the plant immunity paradigm of systemic acquired resistance (SAR), a form of transcriptional memory, exposure to the pathogen P. syringae results in priming of immunity-related, salicylic acid-induced genes, which also appears to be functionally dependent on the loss of non-CpG DNA methylation that is inherited intergenerationally [66]. This type of intergenerational memory may be unique to plants or lower animals, as DNA methylation patterns in the mammalian germline are reset every generation [67].

Repressive histone modifications

Histone modifications associated with gene silencing have also been implicated in regulation of transcriptional memory. Using an elegant synthetic transcription factor induction system, it has been shown that a short pulse of C/EBPα causes commitment of B cells toward macrophage lineage [68]. This induces transcriptional memory of selected target genes such as inflammatory response genes that are substantially increased upon C/EBPα restimulation or LPS treatment even 6 days following priming. Memory does not appear to be controlled by active histone marks (H3K4me1/2/3, H3K27ac), relevant transcription factors (PU.1 and C/EBPα), or polymerase II. Instead, target genes show sustained demethylation of H3K27, suggesting memory is a consequence of delayed remethylation of this residue. Indeed, transient inhibition of H3K27me3 demethylases in primed cells impairs their transcriptional memory. Moreover, inhibition of EZH1/2 methyltransferases causes partial loss of H3K27me3 that is sufficient to trigger transcriptional memory even in the absence of a C/EBPα pulse. These findings suggest that maintenance of H3K27 hypomethylation is causal in transcriptional memory in this system [68].

Meso-scale mechanisms of transcriptional memory: nuclear positioning and long-range interactions

Beyond local changes in gene promoter composition and architecture, memory may also be maintained by changes in long-range chromatin contacts and nuclear position. This is best characterized in yeast, in which specific transcription factors recruit promoters of primed genes to nuclear pore complexes (NPCs) at the nuclear periphery [5, 6, 33] (Fig. 3e). NPC tethering and memory are critically dependent on the nuclear pore protein Nup100 [5, 16]. This feature is conserved in humans where Nup98 is required for memory, at least in part by maintaining H3K4me2 and RNA polymerase II [16]. Furthermore, genes that show salt stress memory in yeast also associate with the nuclear pore via the Nup42p component, which is required for the maintenance of the primed state [69]. Similarly, the yeast GAL1 promoter also carries sequence elements that are required for tethering the gene to the nuclear periphery during memory. Although, in this case, relocalization appears to be a consequence of memory rather than a requirement, as disruption of the NPC interaction does not affect the rate of GAL1 reactivation [6]. It has been postulated that the interaction of primed genes with the nuclear pore complex could result in stronger gene expression by facilitating post-transcriptional processing such as mRNA processing and export [70, 71]. Remarkably, the association of primed genes with nuclear pore proteins has been reported in other transcriptional memory models including ecdysone-inducible memory in Drosophila and IFNγ memory in human cells, suggesting that it is a core, conserved mechanism [16, 47]. However, intriguingly, these interactions appear to be relevant within the nucleoplasm, with no evidence of association to the nuclear pore complex. Instead, in the ecdysone transcriptional memory paradigm in Drosophila, priming causes formation of enhancer-promoter loops at transcriptionally activated target genes (Fig. 3f). Importantly, these long-range contacts are maintained even after removal of ecdysone and depend on the nuclear pore protein, Nup98 [7]. Depletion of Nup98 prevents formation of enhancer-promoter loops and specifically erases transcriptional memory without affecting ecdysone-induced transcriptional activation. This suggests that interaction with NPC proteins is a broadly conserved aspect for maintenance of transcriptional memory, associated with NPC tethering in yeast and chromatin looping in flies (Fig. 4c).

In the human context, genes with strong IFNγ memory tend to reside in genomic clusters such as the MHC II genes as well as GBP genes. Strikingly, it has been shown that transcriptional memory of these genes was locally restricted by cohesion as systemic degron-induced cohesin loss or deletion of a local cohesin binding site resulted in enhanced priming of GBP genes [3] (Fig. 3f). This indicates that long-range interactions also play an important role in human transcriptional memory, although the role of Nup interactions, in this case, is not yet clear.

A further insight into the control of transcriptional memory of gene clusters came from the discovery of a set of immune gene–priming long-noncoding RNAs (lncRNAs) in a human endothelial culture system activated by TNF-α [72]. One of these lncRNAs, termed UMLILO has been shown to be involved in recruiting MLL1 (H3K4me3 methyltransferase) to a cluster of nearby TNF-α activated promoters in cis, facilitating their expression. Interestingly, UMLILO is upregulated by β-glucan which was previously shown to prime e.g., monocytes for future hyper expression of cytokines upon e.g., LPS exposure [50]. This implicates UMLILO in inducing transcriptional memory for future activation of β-glucan primed genes, although this was not directly tested.

Molecular dynamics of memory

While many factors have been implicated in memory and some are locally maintained on primed chromatin, which factors physically “carry” memory through mitotic cell divisions is still very poorly understood. Likely, not a single factor is solely responsible for maintaining the primed state and other factors could pass on the “baton” of memory as in a “relay race” along the cell cycle. Histone variants and their modifications display a range of lifetimes [73]. For instance, while H2A.Z is required for yeast transcriptional memory, it turns over faster than even canonical H2A implying that, if H2A.Z is a carrier of the memory, it must be continuously and actively re-incorporated. Other modifications last for hours or even days, such as silent marks including H3K27me3 that turns over largely through passive turnover [74, 75]. Thus, in the case of repressive chromatin marks, slow turnover and slow re-establishment post-stimulus could be an important ‘passive’ mechanism promoting transcriptional memory. Active chromatin marks typically turn over faster that the duration of the memory they are associated with [73]. Nevertheless, they may serve as a platform for more dynamic modifiers and scaffolds to maintain the primed state. Thus, in effect, the “baton” of memory is held by the feedback between a histone modification and its modifier. Many chromatin modifiers are recruited by proteins that carry reader domains for the catalysed mark, creating a potential feedback loop, including the SET/MLL proteins that create H3K4 methylation [76] (Fig. 4a). Another question is how transcriptional memory is ultimately reversible. Research has typically focussed on how primed states are propagated but ultimately priming must be reset. Is this a regulated process or do self-propagating mechanism simply loose out against the attrition of turnover? These are key questions for future work.

The adaptive function of transcriptional memory

Transcriptional memory is a widespread phenomenon present across biological kingdoms. Therefore, it is likely physiologically relevant as it enables a stronger response and faster adaptation to a future encounter with changing environmental and physiological conditions.

Transcriptional memory in yeast allows for rapid responses to sudden changes in nutrient availability, possibly enhancing the yeast’s survival in fluctuating environments. For instance, primed yeast show faster adaptation and growth upon glucose starvation [29]. Moreover, in oxidative and salt stress, the cells that lose memory (e.g., through mutation of nuclear pore components) exhibit increased susceptibility to salt and oxidative stress [69].

In plants, transcriptional memory serves as an adaptive strategy in response to recurring environmental stress, particularly in the case of temperature fluctuation. Primed Arabidopsis plants exhibit the ability to retain water upon reexposure to dehydration stress, resulting in improved fitness during drought conditions [77]. Notably, Arabidopsis plants that have been primed with salt display enhanced survival capabilities under both salt and drought stress, making them cross-resistant to these challenging conditions.

In mammals, IFN-induced transcriptional memory enables cells to mount a stronger response against infections. MEF cells primed with IFNβ exhibit enhanced survival against encephalomyocarditis virus (EMCV) when compared to naive cells [2]. The observation that transcriptional memory in human cells is triggered by various innate immune stimuli raises an intriguing possibility that it may underpin a long-standing physiological phenomenon of memory effects in the mammalian innate immune system. This memory is distinct from classical adaptive immune memory that relies on somatic mutations. The central tenet is that activation of an innate immune response can lead to a form of epigenetic memory that results in enhanced resistance to reinfection, even to different pathogens, a state that can be long-lasting but is reversible. Literature on this phenomenon spans decades [78,79,80], but was recently brought together under a common denominator as “trained immunity” [18, 19].

Does transcriptional memory explain trained immunity?

The phenomenon of trained immunity is of special interest as many innate immune and inflammatory agents such as β-glucan, TNF-α, LPS and interferons have not only been shown to induce transcriptional memory [21, 50, 81] but also to trigger trained immunity. Examples include exposure of humans or mice to BCG (Bacillus Calmette–Guérin) vaccine that induces a primed response resulting in enhancement of inflammatory cytokine production upon a secondary infection [82]. Also, beyond infection, epithelial stem cells that reside in skin tissue remember inflammatory signals even 6 months after the initial trigger and accelerate wound healing by sustained open chromatin at specific genes [36]. Transcriptional memory and trained immunity are conceptually similar, as in both phenomena the primed state is inherited in somatic tissues suggesting that aspects of transcriptional memory contribute to innate immune memory. Possible molecular mechanisms that may help explain innate immune priming, include: chromatin accessibility and modification [83], transcription of long non-coding RNAs (lncRNAs) [72], DNA methylation [84] and reprogramming of cellular metabolism [23, 85].

However, beyond the local, gene-specific memory of transcription, priming in trained immunity likely involves additional mechanisms, many of which are not cell autonomous. First, signals such as LPS, TNF-α or other cytokines, in addition to transiently activating genes, also rewire the transcription factor network in immune cells, including lineage-specific transcription factors e.g., in macrophage polarization [86]. In effect, these cells become to some extent differentiated, involving continued activation of novel enhancers and continued expression of genes. Thus, while important for trained immunity, such mechanisms are likely distinct from those that govern transcriptional memory - a phenomenon that is more transient, maintains similar cell identity and allows clearer uncoupling between active transcription and the inheritance of primed state. Another challenge in understanding the molecular basis of trained immunity is to establish where memory resides at the organismal level. For instance, most peripheral mature macrophages are short-lived and thus unlikely are carriers of innate immune memory [85]. The discovery that stem cells may carry memory may help explain why trained immunity appears to be more robust and lasts for up to several months [13, 87]. Transcriptional memory only lasts for weeks in culture but can be carried by rapidly proliferating cells over multiple cell division cycles [2, 3]. Perhaps in slowly dividing stem cells, the same number of cell divisions stretch for months, allowing transcriptional memory to act at time scales of trained immunity. Finally, memory in trained immunity is not only retained in cells that receive priming signals such as interferons but also in cells that produce these signals, thereby potentially creating additional feedback. NK cells are a key example of this; during activation, these cells clonally expand and display enhanced interferon production upon reactivation, leading to more cytokine production and further clonal expansion, creating more cells participating in this production [85, 88]. This illustrates that in vivo, multiple mechanisms likely synergize to create long-term innate immune memory.

In sum, while mechanisms underlying trained immunity are likely multi-layered, transcriptional memory is likely a part of the equation. It can offer important insights into the core mechanisms behind trained immunity and allow to exploit them in clinical applications such as vaccination, as well for prevention and treatment of chronic inflammation.