Avalilação PREreview de Induction of hysteretic chromatin states by CRISPRa reveals locus-specific encoding of transcriptional memory

In this work by Bedford et al, the authors studied the transcriptional memory (or hysteresis) phenomenon, whereby gene responses are exacerbated in the subsequent activations following a first challenge. Specifically, they aimed to evaluate the contribution of the event of transcription itself to the installation of transcriptional memory. To do this, they used a model of pro-inflammatory gene activation by exposing mesenchymal stromal cells to a pulse of IFNγ and then re-exposing the cells after several days. RNA-seq was used to determine the set of genes that displayed enhanced response in the re-exposure compared with the first IFNγ pulse (memory genes). Afterwards, they performed CRISPRa (VP64-dCas9-VP64) at two loci that showed strong memory behavior in their IFNγ experiments: IDO1 and CXCL11, and observed that activation by CRISPRa alone can install a transcriptional memory independently of IFNγ. Finally, by using DAF-seq and Cut&Run of histone marks H3K4me3, H3K36me2/3 and H3K27me3, the authors found that genes showing a memory behavior display a persistent decrease in H3K27me3 that is not observable in the other histone marks examined, nor in chromatin accessibility in either IDO1 or CXCL11 promoters, concluding that H3K27me3 hysteresis appears central to the establishment of transcriptional memory. The authors temper this conclusion by noting that a subset of non-memory ISGs follows the same H3K27me3 trajectory without showing transcriptional potentiation (data not shown in this pre-print) and propose that additional cis-regulatory features must therefore contribute to this phenomenon.

Overall, the article is well-written, its main claims are supported by the data presented and provides new evidence supporting the determinants of the establishment of a transcriptional memory being gene-level based, meaning that some genes may be intrinsically prone to show memory behavior compared to others in specific cell types and contexts. The central advance, in my view, is the demonstration that a transient, locus-targeted activation is sufficient to install memory in the complete absence of cytokine signalling, which separates the contribution of local activation from the global transcriptomic, proteomic and metabolic changes that accompany IFNγ exposure and which has not, to my knowledge, been shown this cleanly in primary cells. Below, I wrote suggestions with the aim of strengthening the manuscript's conclusions and data presentation so that readers can fully appreciate the relevance of this work.

Major points:

a) It is well established that Polycomb and CBP/P300 antagonize each other and deposit mutually exclusive marks H3K27me3 and H3K27ac, respectively. Considering this, and the fact that VP16, the monomer from which VP64 is derived, binds CBP/p300 through its acidic domain (doi: 10.1073/pnas.97.1.430) and recruits CBP in vivo (doi: 10.1074/jbc.M608451200 — cited in this manuscript as ref. 31) I suggest three things. First, discussing this explicitly in the introduction/discussion, since the antagonism is well known and readers working on Polycomb will raise it, so I consider this essential to contextualize and understand the whole molecular mechanism studied in the manuscript. Second, considering that recruitment of a coactivator complex is mechanistically distinct from the act of transcription, I suggest framing the perturbation as targeted, signalling-independent transcriptional activation at a defined locus, rather than as transcription in isolation. Third, if feasible, adding H3K27ac to the CUT&RUN panel in both the IFNγ and CRISPRa arms would significantly enhance the mechanistic picture of transcriptional hysteresis; adding EZH2 and CBP/p300 occupancy by ChIP or CUT&RUN would be more definitive but is clearly a larger undertaking, and I would not regard it as essential for this manuscript.

b) As written, the Methods indicate a single deaminated reaction per condition per locus was performed for DAF-seq, while the per-fiber MSP comparisons are tested with Mann-Whitney across fibers. Because fibers from one library are not independent biological observations, these P values address molecular heterogeneity rather than reproducibility across experiments. Adding one or two independent biological DAF-seq replicates per condition, with the comparison made across replicate-level summaries, would let the accessibility conclusions carry the same evidentiary weight as the RNA-seq and CUT&RUN-qPCR analyses (both n = 3).

c) I would also suggest extending DAF-seq to the CRISPRa conditions. At present the accessibility and footprint results are available only for the IFNγ paradigm, so the inference that the CRISPRa-stimulated state shares the same accessibility-independent, factor-reacquisition mechanism is an extrapolation. Given that CRISPRa is the manuscript's key conceptual advance, this is an experiment I would like to see.

d) In addition to the previous point, the authors performed amplicon-based chromatin accessibility measurements in the promoter-proximal regions of IDO1, CXCL11 and STAT1, and based on their analysis, conclude that chromatin accessibility changes do not reflect the memory behavior shown for IDO1 and CXCL11 in transcription and H3K27me3 measurements. However, considering that both IFNγ and CRISPRa can alter the landscape of regulatory regions such as enhancers that are proximal in the tridimensional space of the nucleus but distal in linear genomic distance to target genes (doi: 10.1038/s41594-025-01522-8 , 10.1093/nar/gkac582 ), I suggest extending the measurements of chromatin accessibility by performing ATAC-seq. Otherwise, acknowledging this limitation in the discussion may help clarify the conclusions drawn.

Minor points:

a) Considering the forementioned balance between Polycomb and CBP/P300, it would be interesting to evaluate the effect of inhibition of CBP/P300 (e.g. by using A-485) and observe if this is able to prevent the loss of H3K27me3 and thus the establishment of memory.

b) Figure 1F merges two k-means clusters (n = 201 and n = 51) into a single Memory-Like class. The smaller sub-block, where IL6 and PLOD2 are annotated, appears to retain elevated expression in the rested condition, which is a mechanistically distinct behavior from classical memory (full resolution followed by potentiation). Reporting these clusters separately, and testing whether this sub-group is enriched among the genes that separate the conditions along PC2 in Fig. 1G would strengthen the link between the response classes and the transcriptional drift the authors describe.

Disclosure statement: I reviewed this preprint as part of the HHMI Transparent and Accountable Peer Review training program. This report was written with the support of my research mentor.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.