Genome-scale datasets are required to train models to predict gene expression using histone PTMs. Therefore, we obtained histone PTM ChIP-seq and RNA-seq data for 13 different human cell types from ENCODE (Schreiber et al., 2020a; The ENCODE Project Consortium, 2012; Appendix 1—table 1). We inspected metagene plots (histone PTMs averaged across genes within gene expression quantiles) describing 6 histone PTMs in each of these 13 different cell types. Based on different overall signal levels across cell types, we concluded that batch effects, likely due to inconsistent sequencing depths, would need to be corrected prior to training models (Figure 2—figure supplement 1).

We corrected these batch effects by adapting S3norm (Xiang et al., 2020; ‘Materials and methods’, Figure 2—figure supplement 2). These corrected histone PTM tracks were then used for the remainder of our analyses along with RNA-seq data for each of the 13 cell types (Figure 2—figure supplement 3).

Importantly, we observed that H3K27ac and H3K4me3 histone PTM signal strengths positively covaried with gene expression quantile (representative cell types shown in Figure 2; all cell types shown in Figure 2—figure supplement 3). Conversely, repressive histone PTMs such as H3K27me3 and H3K9me3 were strongly inversely correlated with gene expression quantiles. Spatial patterns in the metagene plots for H3K36me3 suggested that this mark covaried more strongly with gene expression in the gene body than near the TSS. Taken together, these observations recapitulated the current understanding of these well-studied histone PTMs with respect to their associations to gene expression (Kimura, 2013; Millán-Zambrano et al., 2022; Zhao et al., 2021).

To predict how epigenome editing affects gene expression, we first trained models to predict gene expression from endogenous histone PTMs. We trained several convolutional neural networks (CNNs) and ridge regression models to predict the gene expression of each gene in each of the 13 cell types, using only histone PTM data proximal to the TSS as features (‘Materials and methods’, Figure 2). We observed that Spearman’s rank correlation between the true gene expression and the models’ predicted gene expression on held-out chromosomes improves as the input context size increases; and for all input context sizes, the CNNs outperform ridge regression models (Figure 3A). Therefore, for the remainder of the analyses, we use a context size of 10,000 base pairs.

To assess the models’ ability to generalize to unseen cell types, we trained a set of 10 models for each cell type. In particular, we held out the histone PTMs for a given cell type during training and then tested the models on that held-out cell type.

We observed that the CNNs outperformed ridge regression models on this cross-cell type generalization task across essentially all cell types (Figure 3B). The reduced performance on the adrenal cell type may be driven by a cell-type-specific biological mechanism that leads to a lower correlation of its epigenetic data with other cell types, particularly for H3K36me3 (Figure 3—figure supplement 1).

Although our models accurately predicted endogenous gene expression, this does not guarantee their ability to accurately predict the relationship between local histone PTM variations and gene expression for a particular gene across different cell types. Therefore, we determined Spearman’s rank correlations between the observed expression and the predicted expression for each held-out gene across the different cell types. The distribution of these correlations suggests that overall the CNNs can better rank cell types by gene expression than ridge regression (Figure 3C). In particular, the median cross-cell type correlation is ∼0.53 for CNNs compared to ∼0.39 for ridge regression.

We also benchmarked the predictive performance of our CNN model against existing methods and observed that we outperform all existing methods across multiple cell types (Figure 3—figure supplement 3).

We investigated what features of the data the models used to predict gene expression. For a given gene, we modified the input histone PTMs one-by-one at nucleosome-scale and measured the predicted fold-change in gene expression (Figure 4, Figure 4—figure supplement 1, ‘Materials and methods’).

We observed considerable changes to the predicted fold-change upon modifying different histone PTMs. In particular, our CNN models predict that repressive marks such as H3K27me3 and H3K9me3 proximal to the TSS result in a slight decrease in expression. In contrast, activating histone PTMs such as H3K27ac and H3K4me3 results in an almost twofold increase in predicted gene expression near the TSS. Activating both of these markers exhibits a periodic pattern, likely reflecting nucleosome occupancy. However, activation of H3K4me3 results in a sharp increase in gene expression downstream of the TSS. Additionally, we observed that H3K36me3 is predicted to increase expression, but only if it is deposited in the gene body, and the degree of activation gradually increases as it is deposited further inside of the gene body. The consistency of these observations with established mechanisms, observed previously in the literature, via which these histone PTMs modulate gene expression (Kimura, 2013) lends credence to our gene expression models and shows that these models learn the spatial patterns of histone PTMs.

To test if our gene expression models could accurately predict the outcome of in situ epigenome editing experiments, we first generated dCas9-p300 data in the HEK293T cell line for eight genes (Figure 5). We assayed at least five gRNAs per gene with at least three replicates for each gRNA. We used the HEK293T cell line because it is a widely used testbed for epigenome editing strategies (Hilton et al., 2015; Nuñez et al., 2021; O’Geen et al., 2017; Mahata et al., 2023; Escobar et al., 2022; Wang et al., 2022). Based on Figures 2 and 4, the largest changes in H3K27ac across gene expression quantiles occur within 500 base pairs of the TSS, so we constrained gRNA targeting to this critical window. We filtered gRNAs for predicted specificity (Concordet and Haeussler, 2018) and on-target activity scores (Sanson et al., 2018). Each gRNA was tested individually, and relative mRNA abundance was measured using quantitative PCR (qPCR).

We successfully increased gene expression of all eight genes with fold-change activation using the most effective respective gRNA for each gene ranging from 3-fold to ∼6500-fold relative to a non-targeting control gRNA (Figure 5C). Some of this variation may be explained by differences in endogenous gene expression levels, with the targeting of lowly expressed genes resulting in higher fold-change measurements (Appendix 1—table 2), as observed previously (Wang et al., 2022). Nevertheless, substantial variability was observed in gRNA efficacy for all targeted genes. In particular, two (MYO1G and PRSS12) out of eight genes had the most efficacious gRNA downstream of the TSS. This contrasts with other reports where targeting CRISPR/Cas-based activators upstream of the TSS leads to the highest activation (Mohr et al., 2016; Gilbert et al., 2014).

These data indicate that the rules governing the outcomes for successful dCas9-p300-based epigenome editing – and subsequent increased transcriptional activation – are complex and highlight the fact that locus-specific nuances can be important factors in epigenome editing experiments. For example, two gRNAs targeting within ∼50 base pairs of each other on C2CD4B have a 100-fold difference in measured mRNA (Figure 5C). Further, gRNAs targeting the same position in different genes can have vastly different effects. For instance, several gRNAs targeting ∼250 base pairs upstream of the CYP17A1 TSS result in a high fold-change, while two gRNAs targeting roughly the same position in MYO1G failed to produce substantial activation (Figure 5C).

To test the hypothesis that dCas9-p300 acts through the local deposition of H3K27ac, we modeled this process in silico and used these perturbations as inputs to our models trained on endogenous gene expression.

We modeled the effect of dCas9-p300 on histone PTMs based on evidence from the literature as well as additional experiments we performed. The key assumptions of this model are (1) there exists steric hindrance of dCas9 by nucleosomes (Makasheva et al., 2021; Horlbeck et al., 2016; Isaac et al., 2016; Radzisheuskaya et al., 2016); (2) dCas9-p300 acts locally, altering H3K27ac levels near the gRNA target locus (Gemberling et al., 2021; Dominguez et al., 2022) (we adopted this simplifying assumption since off-target effects are unpredictable and underexplored; Dominguez et al., 2022; Gemberling et al., 2021; Weinert et al., 2018); and (3) dCas9-p300 can deposit H3K27ac at nucleosomes, as defined by MNase activity (see ‘Materials and methods’; Segelle et al., 2022; Zhou et al., 2016). Our resulting in silico perturbation model had a number of free parameters that we briefly describe below. Wherever possible, we used values for these parameters obtained from the literature or tested a range of plausible values. For a more complete description of the model, see ‘Materials and methods’.

The first component of our perturbation model is steric hindrance of dCas9-p300 by nucleosomes (Figure 6A). Intuitively, if DNA is tightly wound around a nucleosome, the gRNA would be less likely to bind successfully. Mathematically, we modeled this as an inverse relationship between the amount of H3K27ac deposited and the MNase activity at the gRNA target locus.

It is widely assumed that dCas9-p300 activates genes through the local deposition of H3K27ac (Klann et al., 2017; Dominguez et al., 2022). To model this, we increased local levels of H3K27ac relative to endogenous levels according to a Gaussian kernel centered at the gRNA target locus (Figure 6B). This adds acetylation primarily within a distance controlled by the standard deviation (σ\begin{document}$\sigma$\end{document}) of the kernel. We performed CUT&RUN experiments (see Appendix 1) that suggest that this distance is at least 1000 base pairs (Figure 6—figure supplement 1). Since we also do not know the degree to which dCas9-p300 alters H3K27ac levels, we modeled this as another free parameter, λ, which we varied over a range of plausible values (‘Materials and methods’).

Finally, we assumed that dCas9-p300 does not affect the positioning of nucleosomes and hence can only add H3K27ac at positions currently occupied by histones (Zhou et al., 2016). As such, we expect H3K27ac levels to only increase at loci where there is MNase activity. In particular, we modulated the Gaussian kernel described above by performing point-wise multiplication with MNase activity (Figure 6C).

Since nucleosome positioning plays a crucial role in our perturbation model, we generated, to our knowledge, the first MNase-seq data for the HEK293T cell line (see Appendix 1).

To get a baseline of how well our perturbation model might be able to predict the effect of dCas9-p300 on gene expression, we considered the 13 distinct cell types as being analogous to natural perturbations of local histone PTMs. Across the eight genes discussed above, which were excluded from the training set, we observed a Spearman’s rank correlation of ∼0.8 between the endogenous expression and that predicted by our expression model (Figure 6D). This correlation was in line with the correlation observed across the endogenous transcriptome (Figure 3A and B). We further observed that our expression models were able to accurately rank gene expression across cell types within individual genes (Figure 6—figure supplement 2).

We then computed fold-changes between the expression predicted using endogenous histone PTMs and the expression predicted using in silico perturbations of these histone PTMs. We observed that our models were effective in ranking relative fold-changes across genes in response to dCas9-p300, achieving a Spearman’s rank correlation of ∼0.8 between these predicted fold-changes and the experimentally determined mRNA fold-changes induced by dCas9-p300 (Figure 6E). However, the performance in ranking fold-changes within individual genes was less accurate (Figure 6—figure supplement 3) when compared to the prediction of cell-type-specific gene expression from native epigenetic signatures (Figure 6—figure supplement 2).

We extended this analysis to a Perturb-seq dataset, consisting of gRNAs targeting proximal to the TSS of five genes in the K562 cell line to further assess the model’s ability to estimate gene expression changes. Consistent with the performance observed in Figure 6E, the model demonstrated robustness in predicting gene expression fold-changes across these 27 gRNAs targeting these five genes. Notably, these predictions achieved a Spearman’s rank correlation of ∼ 0.47 with the experimentally determined mRNA fold-changes measured by Perturb-seq, as shown in Figure 6—figure supplement 4 (see Appendix 1). These results reinforce the model’s effectiveness in capturing the nuanced effects of epigenome editing across different genes and cell types.

We obtained −log10⁡(p-value)\begin{document}$-\log_{10}({p\text{-value}})$\end{document} ChIP-seq tracks created by running the MACS2 peak-caller (Feng et al., 2012) on read count data, from the ENCODE Imputation Challenge (Schreiber et al., 2020a). For three tracks where data were not available, we downloaded Avocado (Schreiber et al., 2020b) imputations from the ENCODE data portal (The ENCODE Project Consortium, 2012). We binned each epigenetic track at 25 base pair resolution and pre-processed them with an additional log\begin{document}$\log$\end{document} operation before inputting them into the models for training.

We downloaded polyA-plus RNA-seq gene expression transcripts per million (TPM) values for each of the 13 cell types in Appendix 1—table 1, from the ENCODE data portal (The ENCODE Project Consortium, 2012) and preprocessed them with a log\begin{document}$\log$\end{document} operation.

We assigned IMR-90 to be a reference cell type, for each of the six histone PTMs and kept its p-values unchanged. We then performed a transformation for each of the remaining cell types adapted from the core technique developed by S3norm (Xiang et al., 2020), in order to normalize each histone PTM track in each of these remaining cell types, with respect to the corresponding histone PTM track in IMR-90.

First, we computed peaks in both the reference as well as the target cell type. Peaks were defined as the 25 base pair bins corresponding to FDR-adjusted p-values <0.05 (Benjamini and Hochberg, 1995). For histone PTM tracks that were obtained from Avocado imputations (due to lack of availability of experimental data), peaks were defined to be the 1000 bins containing the smallest Avocado imputed p\begin{document}$p$\end{document}-values (based on suggestions from the authors of Avocado Schreiber et al., 2020b). All the remaining bins were defined to be background, for both the reference as well as the target cell types.

We then computed the list of peaks that were common to both the reference and the target cell types. These were termed common peaks. Similarly, we defined common background as the list of bins that were assigned to be background in both the reference as well as the target cell types.

The S3norm method was designed to work with count data, which is always ≥ 1. However, the histone PTM tracks, which are represented as −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document}, are not guaranteed to always be ≥ 1; hence, we transformed all the histone PTM tracks by adding 1 to the −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document}, in both the reference as well as the target cell types.

Additionally, since the histone PTM tracks obtained from imputations performed by Avocado were not guaranteed to be distributed similar to experimental −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document}, we scaled all the histone PTM tracks (both experimental as well as Avocado imputations) by dividing them by the minimum observed value in common peaks and common background, in order to bring experimental data and Avocado imputations onto a similar footing. In particular, before applying the S3norm normalization, we transformed −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document} in common peaks and common background for both the reference as well as the target cell type as follows:(1)TransformedCommonPeaksi,reference=1+CommonPeaksi,referencemini(CommonPeaksi,reference)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{TransformedCommonPeaks}_{i,\text{reference}} &= \frac{1 + \text{CommonPeaks}_{i,\text{reference}}}{\min_i(\text{CommonPeaks}_{i,\text{reference}})}$$\end{document}(2)TransformedCommonPeaksi,target=1+CommonPeaksi,targetmini(CommonPeaksi,target)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{TransformedCommonPeaks}_{i,\text{target}} &= \frac{1 + \text{CommonPeaks}_{i,\text{target}}}{\min_i(\text{CommonPeaks}_{i,\text{target}})}$$\end{document}(3)TransformedCommonBackgroundi,reference=max(1+CommonBackgroundi,referencemini(CommonBackgroundi,reference),0)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{TransformedCommonBackground}_{i,\text{reference}} &= \text{max}(\frac{1 + \text{CommonBackground}_{i,\text{reference}}}{\min_i(\text{CommonBackground}_{i,\text{reference}})}, 0)$$\end{document}(4)TransformedCommonBackgroundi,target=max(1+CommonBackgroundi,targetmini(CommonBackgroundi,target),0)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{TransformedCommonBackground}_{i,\text{target}} &= \text{max}(\frac{1 + \text{CommonBackground}_{i,\text{target}}}{\min_i(\text{CommonBackground}_{i,\text{target}})}, 0)$$\end{document}

The normalization procedure of S3norm then wishes to find two positive parameters, α\begin{document}$\alpha$\end{document} and β\begin{document}$\beta$\end{document} that are to be learned from the data such that both the following equations are satisfied:(5)mean(TransformedCommonPeaksreference)=mean(α×TransformedCommonPeaksβtarget)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{mean}(\text{TransformedCommonPeaks}_{\text{reference}}) &= \text{mean}(\alpha \times {\text{TransformedCommonPeaks}^\beta}_{\text{target}})$$\end{document}(6)mean(TransformedCommonBackgroundreference)=mean(α×TransformedCommonBackgroundtargetβ)\begin{document}$$\displaystyle \newcommand{\fref}[1]{Figure~\ref{#1}}\newcommand{\sref}[1]{Section~\ref{#1}} \text{mean}(\text{TransformedCommonBackground}_{\text{reference}}) &= \text{mean}(\alpha \times \text{TransformedCommonBackground}^\beta_{\text{target}})$$\end{document}

Specifically, α is a scale factor that shifts the transformed −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document} of the target data set in log scale, and β is a power transformation parameter that rotates the transformed −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document} of the target data set in log scale (Figure 2—figure supplement 2). There is one and only one set of values for α and β that can simultaneously satisfy both the above equations for common peaks and the common background (Xiang et al., 2020).

The values of α and β were estimated by the Powell minimization method implemented in scipy (Fletcher and Powell, 1963; Virtanen et al., 2020). The resulting normalized −log10⁡(p-values)\begin{document}$-\log_{10}({p\text{-values}})$\end{document} were used for all downstream analyses (Figure 2—figure supplement 3).

We trained CNN and ridge regression models, each, to predict gene expression using histone PTM tracks. Input features for each gene were centered at its TSS. We used an input context size of 10,000 base pairs for all analyses subsequent to Figure 3. For all analyses, we obtained predictions from our models by averaging predictions ensembled across 100 computational replicates.

To train CNN models, the normalized histone PTM tracks for each gene were processed with successive convolutional blocks. Each convolutional block consisted of a batch-normalization layer, rectified linear units (ReLU), a convolutional layer consisting of 32 convolutional kernels, each of width 5, followed by a dropout with 0.1 probability. Finally, a pooling layer was applied to gradually reduce the dimension of the features. After being processed with five such convolutional blocks, the output was flattened and passed through a fully connected layer consisting of 16 neurons and a ReLU activation. This was ultimately processed with a fully connected layer with a single output and a linear activation (since this was a regression task). The models were trained with a mean squared error loss using the Adam optimizer with a learning rate of 0.001 for the first 50 epochs and 0.0005 for the remaining 50 epochs. Training CNN models took about 1.5 hours on 1 NVIDIA A100 Tensor Core GPU.

To see how different features affected predicted levels of expression, we systematically perturbed each input feature and determined how much the perturbation affected predicted expression levels. To be concrete, we denoted the epigenetic feature at position i\begin{document}$i$\end{document} of gene g\begin{document}$g$\end{document} in cell type CT\begin{document}$CT$\end{document} as EiCT,g\begin{document}$E^{CT,g}_{i}$\end{document}. We then defined a perturbation function that added a scalar value of λ0=2500\begin{document}$\lambda_{0}=2500$\end{document} to the epigenetic features within 3 bins of a focal position, say, j\begin{document}$j$\end{document}:Fj⁢(E1C⁢T,g,…,EWC⁢T,g):=(E1C⁢T,g,…,Ej-3C⁢T,g+λ0,…,Ej+3C⁢T,g+λ0,…,EWC⁢T,g),\begin{document}$$\displaystyle F_{j}\left(E^{CT,g}_{1},\ldots,E^{CT,g}_{W}\right):=\left(E^{CT,g}_{1},\ldots,E^{CT,g}_{j-3}+\lambda_{0},\ldots,E^{CT,g}_{j+3}+\lambda_{0},\ldots,E^{CT,g}_{W}\right),$$\end{document}

recalling that W\begin{document}$W$\end{document} is the number of bins of 25 base pairs considered by our models, which is set to 401, corresponding to a 10,000 base pair input context length, for all analyses subsequent to Figure 3. These perturbations corresponded to ∼150 base pairs, which is roughly the length of DNA wrapped around a nucleosome.

To produce Figure 4, we applied the above perturbation functions, F1,…,FW\begin{document}$F_{1},\ldots,F_{W}$\end{document} to a histone PTM track of interest, and then measured the fold-change in predicted expression. To account for differences in the endogenous histone PTM tracks between genes, we averaged these fold-changes across 500 randomly chosen genes.

Our model of how dCas9-p300 perturbs local histone PTMs has three separate components. We describe each of these components in turn, and then present the full model below. Throughout, we write j\begin{document}$j$\end{document} for the position that the gRNA targets.

First, we modeled steric hindrance of dCas9 due to nucleosomes. We used MNase-seq signal strength as a proxy for nucleosome occupancy. Letting mj\begin{document}$m_{j}$\end{document} be the MNase-seq read coverage at the gRNA binding site, we modeled steric hindrance by scaling the acetylation activity of dCas9-p300 by a factor of exp⁡(−5×mj)\begin{document}$\exp\left(-5\times m_{j}\right)$\end{document}.

Second, we assumed that dCas9-p300 primarily alters the levels of H3K27ac only locally. As such, we modeled the acetylation activity of dCas9-p300 at a particular locus as a Gaussian kernel centered at the gRNA. Concretely, the acetylation activity at position i\begin{document}$i$\end{document} is multiplied by a factor of exp⁡(−(i−j)2/2σ2)\begin{document}$\exp\left(-(i-j)^{2}/2\sigma^{2}\right)$\end{document}, where σ2\begin{document}$\sigma^{2}$\end{document} is a parameter of the model.

Finally, we assumed that dCas9-p300 can only acetylate histones where they currently are – it cannot move histones or increase H3K27ac levels outside of histones. To model this mathematically, we multiplied the acetylation activity at site i\begin{document}$i$\end{document} by the MNase read coverage, mi\begin{document}$m_{i}$\end{document}. Therefore, if the MNase read coverage is 0 (i.e., there is no evidence of histones at that locus), then the amount of H3K27ac added to that position is also 0.

Putting this all together, for a guide targeting at position j\begin{document}$j$\end{document}, the effect on H3K27ac levels at position i\begin{document}$i$\end{document} is proportional toexp⁡(-5⁢mj)×exp⁡[-(i-j)22⁢σ2]×mi\begin{document}$$\displaystyle \exp\left(-5\,m_{j}\right)\times\exp\left[\frac{-(i-j)^{2}}{2\sigma^{2}}\right]\times m_{i}$$\end{document}

The constant of proportionality (i.e., how strong we expect dCas9-p300 to be overall) is treated as another free parameter, which we denote by λ.

ENCODE has epigenetic data for the HEK293 cell line, but we performed our dCas9-p300 perturbations in the HEK293T cell line. As such, we used the HEK293 histone PTM as well as RNA-seq data as a stand-in for the HEK293T histone PTM and RNA-seq levels. This substitution is justified as gene expression levels for HEK293 and HEK293T are highly concordant (Figure 3—figure supplement 2). Indeed, the Spearman’s rank correlation between expression levels for HEK293 and two independent measurements of expression levels in HEK293T are 0.86 and 0.88, which are comparable to the correlation between the two independent experiments in HEK293T (ρ=0.92\begin{document}$\rho=0.92$\end{document}). That is, the correlation across experiments within HEK293T cells is only slightly higher than the correlation between HEK293 and HEK293T, suggesting that cross-cell type differences between HEK293 and HEK293T are on the same order as the inherent experimental and biological noise within a single cell type.

The details of dCas9-p300 epigenome editing, qPCR, CUT&RUN, and MNase-seq experiments are provided in the supplementary material.