Large-scale analysis of the integration of enhancer-enhancer signals by promoters

Reviewer #1 (Public Review):

This manuscript by Martinez-Ara et al investigates how combinations of cis-regulatory elements combine to influence gene expression. Using a clever iteration on massively parallel reporter assays (MPRAs), the authors measure the combinatorial effects of pairs of enhancers on specific promoters. Specifically, they assayed the activity of 59x59 different enhancer-enhancer (E-E) combinations on 8 different promoters in mouse embryonic stem cells. The main claims of the paper are that E-E pairs combine nearly additively, and that supra-additive E-E pairs are rare and often promoter-dependent. The data in this study generally support these claims.

This paper makes a good contribution to the ongoing discussions about the selectivity of gene regulatory elements. Recent works, such as those by Martinez-Ara et al. and Burgman et al., have indicated limited selectivity between E-P pairs on plasmid-based assays; this paper adds another layer to that by suggesting a similar lack of selectivity between E-E pairs.

An interesting result in this manuscript is the observation that weak promoters allow more supra-additive E-E interactions than strong promoters (Figure 4b). This nonlinear promoter response to enhancers aligns with the model previously proposed in Hong et al. (from my own group), which posited that core promoter activities are nonlinearly scaled by the genomic environment, and that (similar to the trend observed in Figure 5b) the steepness of the scaling is negatively correlated with promoter strength.

My only suggestion for the authors is that they include more plots showing how much the intrinsic strengths of the promoters and enhancers they are working with explain the trends in their data.

Specific Suggestions
Supplementary Figure 4 is presented as evidence for selectivity between single enhancers and promoters. Could the authors inspect the relationship between enhancer/promoter strength and this selectivity? Generating plots similar to Figure 4B and Figure 5B, but for single enhancers, should show if the ability of an enhancer to boost a promoter is inversely correlated to that promoter's intrinsic strength. Also, in Supplementary Figure 4, coloring each point by promoter type would clarify if certain promoters (the weak ones) consistently show higher boost indices across all enhancers. If they do not, the authors may want to speculate how single enhancers can show selectivity for promoters while the effect of adding a second enhancer to an existing E-P has little selectivity. An alternate explanation, based solely on the strength of the elements, would be that when the expression of a gene is low the addition of enhancer(s) has large effects, but when the expression of a gene is high (closer to saturation) the addition of enhancer(s) have small effects.

Can anything more be said about the enhancers in E-E-P combinations that exhibit supra-additivity? Specifically, it would be interesting to know if certain enhancers, e.g. strong enhancers or enhancers with certain motifs, are more likely to show supra-additivity with a given promoter.

Reviewer #2 (Public Review):

Summary
This work investigates how multiple regulatory elements combine to regulate gene expression. The authors use an episomal reporter assay which measures the transcriptional output of the reporter under the regulation of an enhancer-enhancer-promoter triple. The authors test all combinations of 8 promoters and 59 enhancers in this assay. The main finding is that enhancer pairs generally combine additively on reporter output. The authors also find that the extent to which enhancers increase reporter output is inversely related to the intrinsic strength of the promoter.

This manuscript presents a compact experiment that investigates an important open question in gene regulation. The results and data will be of interest to researchers studying enhancers. Given that my expertise is in modeling and computation, I will take the experimental results at face value and focus my review on the interpretation of the results and the computational methodology. I find the result of additivity between enhancers to be well supported. The findings on differential responsiveness between promoters are very interesting but the interpretation of such responses as 'non-linear' or 'following a power-law' may be misleading. More broadly, I think a more rigorous description of the mathematical methodology would increase the clarity and accessibility of this manuscript. A major unanswered question is whether the findings in this study apply to enhancers in their native genomic context. Regardless, investigating such questions in an episomal reporter assay is valuable.

Main comments
Applicability to native genomic context: The applicability of the results in this paper to enhancers in their native genomic context is unclear. As the authors state in the discussion section, the reporter gene is not integrated into the genome, the spacing between enhancers does not match their native context etc. It is thus unclear whether this experimental design is able to detect the non-additivity between enhancers which is known to be present in the genome. This could be investigated by testing the enhancer-enhancer-promoter tuples for which non-additivity has been observed in the genome (references are given in the introduction) in this assay.

Interpretation of promoter responses as non-linear and following a power-law: In Fig 5, the authors demonstrate that enhancer-enhancer pairs boost reporter output more for weak promoters as opposed to strong promoters. I agree the data supports this finding, but I find the interpretation of such data as promoters scaling enhancers according to a power-law (as stated in the abstract) to be misleading. As mentioned on line 297, it is not possible to define an intrinsic measure of enhancer strength, thus the authors assign the base of the power-law to be the average boost index of the enhancer-enhancer pair across the 8 promoters. But this measure incorporates some aspect of a promoter and is not solely a property of enhancers. It would also be useful to know whether the results in Fig 5 apply to only enhancer-enhancer-promoter triples or also to enhancer-promoter pairs.

Enhancer-promoter selectivity: As a follow-up to a previous study (Martinez-Ara et al, Molecular Cell 2022) the authors mention that the data in this study also shows that enhancers show selectivity for certain promoters. The authors mention that both studies use the same statistical methodology and the data in this study is consistent with the data from the 2022 paper. However, I think the statistical methodology in both studies needs further exposition. This section of the review is thus meant to ensure that I understand the author's methodology, to guide the reader in understanding how the authors define 'selectivity', and to probe certain assumptions underlying the methodology.

My understanding of the approach is as follows: The authors consider an enhancer to be not selective if its 'boost index' is the same across a set of promoters. 'Boost index' is defined to be the ratio of the reporter output with the enhancer and promoter divided by the reporter output with just the promoter. Conceptually, I think that considering the boost index is a reasonable way to quantify selectivity.

The authors use a frequentist approach to classify each enhancer as selective or not selective. The null hypothesis is that the boost index of the enhancer is equal across a set of promoters. This can be visualized in Fig. 2C where the null hypothesis is that the mean of each vertical distribution is equal. Note that in Figure S4 of this paper (and in Figure 4B of their 2022 paper) the within-group variance is not plotted. Statistical significance is assessed using a Welch F-test. This is a parametric test that assumes that the observations within each vertical distribution in Fig 2C are normally distributed (this test does allow for heteroskedasticity - which means that the variance may differ within each vertical distribution). Does the normality assumption hold? This analysis should be reported. If this assumption does not hold, is the Welch test well calibrated?