Author response:
The following is the authors’ response to the original reviews.
Reviewer #1 (Public Review):
Batra, Cabrera, Spence et al. present a model which integrates histone posttranslational modification (PTM) data across cell models to predict gene expression with the goal of using this model to better understand epigenetic editing. This gene expression prediction model approach is useful if a) it predicts gene expression in specific cell lines b) it predicts expression values rather than a rank or bin, c) it helps us to better understand the biology of gene expression, or d) it helps us to understand epigenome editing activity. Problematically for points a) and b) it is easier to directly measure gene expression than to measure multiple PTMs and so the real usefulness of this approach mostly relates to c) and d).
We thank the reviewer for their comment and we agree that directly measuring gene expression (e.g., by performing RNA-seq) is easier than performing multiple PTMs in a new cell line. We designed our approach keeping in mind that the primary use case is to understand how epigenome editing would affect gene expression.
Other approaches have been published that use histone PTM to predict expression (e.g. 27587684, 36588793). Is this model better in some way? No comparisons are made. The paper does not seem to have substantial novel insights into understanding the biology of gene expression. The approach of using this model to predict epigenetic editor activity on transcription is interesting and to my knowledge novel but I doubt given the variability of the predictions (Figures 6 and S7&8) that many people will be interested in using this in a practical sense. As the authors point out, the interpretation of the epigenetic editing data is convoluted by things like sgRNA activity scoring and to fully understand the results likely would require histone PTM profiling and maybe dCas9 ChIP-seq for each sgRNA which would be a substantial amount of work.
We thank the reviewer for this insightful comment. We have included citations for a series of papers (PMIDs: 27587684, 30147283, 36588793) that performed gene expression prediction using histone PTM data. However, each of these methods perform classification of gene expression as opposed to predicting the actual gene expression value via regression. Additionally, the referenced studies all work with Roadmap Epigenomics read depth data as opposed to p-values obtained from the ENCODE pipelines, making it difficult to make direct comparisons.
We outline in the Discussion section that by creating a comprehensive dataset of epigenome editing outcomes, which include quantification of histone PTMs before and after in situ perturbations, will improve our understanding of the effects of dCas9-p300 on gene expression and assist in the design of gRNAs for achieving fine-tuned control over gene expression levels.
Furthermore from the model evaluation of H3K9me3 it seems the model is not performing well for epigenetic or transcriptional editing- e.g. we know for the best studied transcriptional editor which is CRISPRi (dCas9-KRAB) that recruitment to a locus is associated with robust gene repression across the genome and is associated with H3K9me3 deposition by recruitment of KAP1/HP1/SETDB1 (PMID: 35688146, 31980609, 27980086, 26501517). However, it seems from Figures 2&4 that the model wouldn't be able to evaluate or predict this.
We thank the reviewer for their comment. We have included a supplementary figure, Figure 4 – figure supplement 1, that quantifies how sensitive the trained gene expression model is to perturbations in H3K9me3. Indeed our data suggests that the model predictions are sensitive to perturbations in H3K9me3. For instance, there is a clear decrease and a gradual increase as the position where the perturbation is performed moves from upstream to downstream of the TSS. Additionally, the magnitude of the predicted fold-change is a function of how much the H3K9me3 is perturbed and hence the magnitude of change would be even higher if the perturbation magnitude is increased. However, this precise magnitude is hard to estimate In the absence of experimental perturbation data for H3K9me3.
The model seems to predict gene expression for endogenous genes quite well although the authors sometimes use expression and sometimes use rank (e.g. Figure 6) - being clearer with how the model predicts expression rather than using rank or fold change would be very useful.
We thank the reviewer for this important suggestion. We have added text in the revised manuscript to clarify that the model predicts gene expression values, which can be interpreted as rank or fold change, depending on the use case.
One concern overall with this approach is that dCas9-p300 has been observed to induce sgRNA-independent off-target H3K27Ac (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8349887/ see Figure S5D) which could convolute interpretation of this type of experiment for the model.
This is an excellent point and indeed, we and others have observed that dCas9-p300 can result in off-target H3K27ac levels (both increased and suppressed) across the genome. However, p300 is one of the few known proteins that can catalyze H3K27ac in the human genome, and H3K27ac remains a proxy for active genomic regulatory elements. Nevertheless, dCas9-p300 off target activity could certainly convolute our approach. We have included language to address this caveat in our discussion. Interestingly, even though dCas9-p300 (and other epigenome editing enzymes) can lead to off-target chromatin modifications, these effects often occur without coincident disruptions to the transcriptome. This suggests that many chromatin modifications, while “supportive” or “instructive” of/for transcription, may be insufficient (either alone or in the context of dCas9-based fusions) for transcriptional effects.
Figure 2
It seems this figure presents known rather than novel findings from the authors' description. Please comment on whether there are any new findings in this figure. Please comment on differences in patterns of repressive and activating histone PTMs between cell lines (e.g. H1-Esc H3K27me3 green 25-50% is more enriched than red 0-25%).
Thank you for pointing out this issue. We have revised the text in both the Results and Discussion sections to better articulate that the goal of this figure is to validate the hypothesis that there are consistent patterns of histone PTMs with respect to gene expression across different human cell types.
In Figure 2, which illustrates the raw histone marks data, the non-monotonic behavior of H3K27me3 in H1-hESC cells is indicative of a real biological phenomenon. This interpretation is supported by the relatively low Pearson correlation for the H3K27me3 mark observed in these cells, as documented in Figure 1b of another study: https://www.biorxiv.org/content/10.1101/2024.03.29.587323v1.
Figure 3&4
There are a number of approaches including DeepChrome and TransferChrome that predict endogenous gene expression from histone PTMs. I appreciate that the authors have not used the histone PTM data to predict gene expression levels of an "average cell" but rather that they are predicting expression within specific cell types or for unseen cell types. But from what is presented it isn't clear that the author's model is better or enabling beyond other approaches. The authors should show their model is better than other approaches or make clear why this is a significant advance that will be enabling for the field. For example is it that in this approach they are actually predicting expression levels whereas previous approaches have only predicted expressed or not expressed or a rank order or bin-based ranking?
We thank the reviewer for this comment. We have added text to clarify the difference between our approach and existing approaches. There are two key differences between our model and other approaches. First, the gene expression model that we have trained here predicts gene expression values instead of gene expression levels as either high or low. Second, we have trained our models on ENCODE p-value data instead of read depths obtained from the Roadmap Epigenomics Consortium.
Figure 5
From the methods, it seems gene activation is measured by qpcr in hek293 transfected with individual sgRNAs and dCas9-p300. The cells aren't selected or sorted before qPCR so how are we sure that some of the variability isn't due to transfection efficiency associated with variable DNA quality or with variable transfection efficiency?
This is a good question. All DNA preps were generated using high-quality reagents and consistent protocols. In addition, the only variable that changed with respect to transfection efficiency was the gRNA-encoding vector used in qPCR assays. We have added new data which demonstrates that transfection efficiency is shared across experiments (Figure 5 – figure supplement 1). We have also added additional experimental data as well as computational analysis analyzing a new dCas9-p300 based Perturb-seq dataset to the manuscript (Figure 6 – figure supplement 1), which use lentiviral transduction and RNA-seq as readouts and thus, are buffered against the variances mentioned by the Reviewer.
Figure 6
The use of rank in 6D and 6E is confusing. In 6D a higher rank is associated with higher expression while in 6E a higher rank seems to mean a lower fold change e.g. CYP17A1 has a low predicted fold-change rank and qPCR fold-change rank but in Figure 5 a very high qPCR fold change. Labeling this more clearly or explaining it in the text further would be useful.
We thank the reviewer for their suggestion. We have made relevant changes to the caption of Figure 6 to clarify this.
Reviewer #2 (Public Review):
Summary:
The authors build a gene expression model based on histone post-translational modifications and find that H3K27ac is correlated with gene expression. They proceed to perturb H3K27ac at 8 gene promoters, and measure gene expression changes to test their model.
Strengths:
The combination of multiple methods to model expression, along with utilizing 6 histone datasets in 13 cell types allowed the authors to build a model that correlates between 0.7-0.79 with gene expression. This group also utilized a tool they are experts in, dCas9-p300 fusions to perturb H3K27ac and monitor gene expression to test their model. Ranked correlations showed some support for the predictions after the perturbation of H3K27ac.
Weaknesses:
The perturbation of only 8 genes, and the only readout being qPCR-based gene expression, as opposed to including H3K27ac, weakened their validation of the computational model. Likewise, the use of six genes that were not expressed being most activated by dCas9-p300 might weaken the correlations vs. looking at a broad range of different gene expressions as the original model was trained on.
We thank the reviewer for their comments. We have added additional experimental data as well as computational analysis analyzing a new dCas9-p300 based Perturb-seq dataset to the manuscript. We observe that the models we have developed are able to predict the fold-change rank across genes reasonably well (Figure 6 – figure supplement 1), similar to what we observe in Figure 6E.
Reviewer #1 (Recommendations For The Authors):
The authors should comment on how their model is different from or better than other models that use histone PTM data to predict gene expression.
We thank the reviewer for this insightful suggestion. We have included citations for a series of papers (PMIDs: 27587684, 30147283, 36588793) that performed gene expression prediction using histone PTM data. However, each of these methods perform classification of gene expression as opposed to predicting the actual gene expression value via regression. Additionally, the referenced studies all work with Roadmap Epigenomics read depth data as opposed to p-values obtained from the ENCODE pipelines, making it difficult to make direct comparisons.
The authors need to make clear whether their model will apply to other common epigenetic or transcriptional editors such as CRISPRi/H3K9me3 which is widely used.
In this study, we focus on the histone changes induced by p300. However, future studies may use the framework described in our manuscript and apply it to other transcriptional editors as well.
The authors need to be clearer about where they are predicting expression and where they are using rank. Ideally, show both.
We thank the reviewer for this important suggestion. We have added text in the revised manuscript to clarify that the model predicts gene expression values, which can be interpreted as rank or fold change, depending on the use case.
The authors should ideally show a case where they use the model to make a prediction of genes that can and can not be activated by dCas9-p300 or other epigenetic editors and then prove this with experiments.
Thank you for the excellent suggestion. While it is indeed relevant, exploring this would extend beyond the scope of our current study. We consider it a valuable topic for future research.
Reviewer #2 (Recommendations For The Authors):
The y-axis in 5C needs to be labeled. The authors state it is "relative mRNA" but these numbers correlated with fold changes shown in Table S2.
We have clarified the definition of the Y-axis in the caption for Figure 5C.
