To establish comprehensive methylation landscapes across evolutionarily relevant cell types, we differentiated human, chimpanzee, and their fused tetraploid hybrid iPSCs into four cell types (see Methods)^25,27. We included an additional cell type, dental pulp stem cells, which are primary cells isolated from human and chimpanzee donors²⁵. We then generated whole-genome bisulfite sequencing (WGBS) data from human and chimpanzee parental and hybrid cranial neural crest cells (CNCC) and dopaminergic neurons, as well as parental dental pulp stem cells (DPSC), complemented by reduced representation bisulfite sequencing (RRBS) data from induced pluripotent stem cells (iPSCs), skeletal myocytes (SKM) and hepatocytes (HEP) (Figure 1A). Principal component analysis and hierarchical clustering of methylation patterns showed that samples clustered primarily by cell type, then by species, and finally by parental versus hybrid system (Figure 1B and 1C and Figure 1—figure supplement 1). CNCC samples displayed a unique clustering pattern where parental vs. hybrid samples clustered separately, though examination of the dendrogram indicated this was due to a single human parental sample being a slight outlier (Figure 1B). Known cell type-specific gene expression markers validated the correct identity of each differentiated cell type (Figure 1D and Figure 1—figure supplement 2)^36–50. Overall, these results are consistent with our expectation that cell type is the predominant determinant of DNA methylation patterns, while species explains a much smaller proportion⁷. In other words, a human hepatocyte is much more similar to a chimpanzee hepatocyte than it is to a human dopaminergic neuron.

To understand the molecular mechanisms underlying species-specific methylation patterns, we first established a framework for distinguishing between cis- and trans-regulatory control of DNA methylation. By comparing methylation levels between human and chimpanzee alleles within the same cellular environment (hybrid cells), we can isolate cis effects from trans effects and determine which mechanism predominates in establishing species-specific methylation patterns (Figure 2A).

To quantify the relative contributions of cis- versus trans-acting factors in establishing human-chimpanzee methylation differences, we applied a beta-binomial model framework adapted from Hallgrímsdóttir et al. (2024) to classify differentially methylated regions based on allele-specific patterns in hybrid versus parental samples. To determine whether a CpG site reflects a difference in methylation due to cis or trans regulation, two hypothesis tests are used: one to evaluate the null hypothesis that the difference in methylation is due solely to trans, and the other solely to cis (see Methods). Each site is categorized into one of five mutually exclusive regulation groups– pure cis-driven, pure trans-driven, cis + trans, and cis x trans, and conserved (Figure 2A). We investigated genome-wide methylation divergence by fitting the model to methylation levels at each CpG site. In all cell types examined, most CpG sites (83-93%) had conserved patterns of methylation. Among divergent sites, cis-regulation was over twice as common as trans-regulation, while a smaller fraction showed both cis and trans divergence (Figure 2B, 2C, Figure 2—figure supplement 1 and Supplemental File 1). The predominance of cis- over trans-regulation was maintained after FDR correction and remained stable across a range of FDR cutoffs (FDR < 0.25, 0.20, 0.10, 0.05, 0.01; Figure 2—figure supplement 1-2 and Supplemental Files 1-2). This genome-wide predominance of cis-regulation indicates that local sequence differences between human and chimpanzee genomes are the primary drivers of methylation divergence, consistent with previous hypotheses of gene regulation that cis-regulatory changes may be less disruptive than broad trans-acting modifications^1–4,6.

The regional distribution of sites from different regulation groups reveals a consistent pattern across cell types: conserved CpG sites are most likely to be near other conserved sites, while divergent CpG sites are likely to be near conserved sites as well as sites within the same regulation group (Figure 2D). This suggests that CpG sites within the same regulation group tend to form coherent clusters in the genome, and divergent CpG clusters typically have conserved sites scattered within. When conducting regional analyses, methylation data is usually pooled from multiple sites within a pre-defined genomic region^6–10. However, this regional heterogeneity can affect classification as well as estimates of how much each regulatory mechanism contributes to divergence between species (Figure 2—figure supplement 2-3, Supplemental File 2, see Methods). Therefore, pooling data across regions should be approached carefully, as subsets of CpG sites within the same regulatory element can be controlled by distinct mechanisms. These findings underscore the importance of studying methylation divergence at high resolution.

To identify differentially methylated regions (DMRs) exhibiting consistent inter-species methylation directionality and minimal classification heterogeneity of regulation groups, we implemented a changepoint detection method with a tolerance threshold for regulatory homogeneity (see Methods). This method systematically identifies genomic loci where co-localized CpG sites demonstrate uniform regulatory behavior punctuated by a pre-defined fraction of heterogeneous sites: non-conserved classifications showing consistent directional methylation bias, and conserved sites maintaining equal methylation levels between species. This method prioritizes DMRs that are controlled by the same mechanism and display coordinated methylation patterns (Supplemental File 3), thereby enriching for loci with interpretable regulatory mechanisms. Notably, DMRs identified through this method are enriched in 5’ UTRs (just downstream of each transcription start site [TSS]) and CTCF-bound regions, supporting their functional relevance in gene regulation (Figure 2E). The quantification of cis- vs. trans-DMRs yielded the same conclusion that cis-regulation outweighs trans-regulation at establishing divergent methylation patterns (Figure 2F).

To identify the molecular mechanisms driving methylation divergence, we analyzed transcription factors (TFs) associated with trans-DMRs and sequence variants associated with cis-DMRs.

TFs represent a major class of trans-acting factors capable of establishing and maintaining methylation patterns across the genome, often by recruiting methyltransferase or demethylase enzymes. Interspecies variation in TF expression levels, binding activity, or target site accessibility can propagate to downstream methylation differences at their binding sites. To identify TFs potentially driving species-specific methylation divergence through such trans-acting mechanisms, we performed TF binding site motif enrichment analysis on trans-DMRs (Figure 3A, Supplemental File 4 and 5)⁵². This analysis identified several candidate factors, including a pioneer factor FOXA2 (Benjamini-Hochberg corrected FDR = 3.7 × 10⁻⁵ in DA Hu<Ch trans-DMRs) and a chromatin modifier FOXM1 (FDR = 0.002, 4.8 × 10⁻⁶, and 0.021 in CNCC, DA, and HEP Hu<Ch trans-DMRs, respectively), both known to actively reshape DNA methylation landscapes^53–60.

If species-specific levels of a TF are responsible for the enrichment of its predicted binding sites in trans-DMRs, then this enrichment may be strongest in cell types with greater species-specific differential expression of the TF. Consistent with this expectation, higher relative expression of FOXM1 and FOXA2 was associated with lower methylation at their predicted binding sites in that same species and cell type. For example, FOXM1 was over 4-fold more highly expressed in human HEPs than in chimpanzee, and its predicted binding sites were highly enriched specifically for trans-DMRs with lower methylation in human HEPs; in contrast, three other cell types with more similar human vs. chimpanzee expression of FOXM1 showed weaker enrichment and little species-specific asymmetry in trans-DMR methylation directionality (Figure 3B). We observed a similar pattern for FOXA2 (Figure 3—figure supplement 1).

Other TFs showed different patterns of association between species-specific expression and trans-DMR methylation. For example, FOXP1 and OTX2 were most differentially expressed in HEPs, consistent with their predicted binding sites also being most enriched in trans-DMRs in this same cell type (Figure 3—figure supplement 1). However, unlike the pioneer factors described above, binding sites for FOXP1 and OTX2 showed similar patterns of enrichment across cell types for both Hu>Ch and Ch>Hu methylated trans-DMRs. This lack of directionality may suggest that these transcription factors are not restricted to either demethylation or methylation, perhaps having a more indirect or context-specific role in regulating the epigenetic landscape^61–63, though further work is required to test this hypothesis. Altogether, our trans-DMR analysis nominated many TFs that may contribute to methylation divergence, including several whose trans-DMR binding site motif enrichment is associated with the TF’s differential expression.

To identify the molecular mechanisms driving cis-regulatory methylation differences, we then examined whether local sequence variants could explain cis-acting methylation divergence. CpG dinucleotides are the primary targets of DNA methylation in mammalian genomes, and their presence, density, and spacing are known to influence local methylation patterns^64–66. Therefore, single nucleotide variants (SNVs) that either create or disrupt CpG dinucleotides represent a plausible mechanism by which sequence divergence could drive species-specific methylation differences. To investigate this hypothesis, we systematically analyzed SNVs and their associated methylation differences in neighboring CpGs within windows of 25, 50, and 100 base pairs upstream and downstream of each SNV. Our analysis revealed that SNVs affecting CpG dinucleotides—but not other SNVs—are associated with altered methylation patterns of nearby CpG sites up to 50 base pairs from the variant site (Figure 3C). Notably, the effect diminished with increasing distance from the SNV, with the strongest methylation changes observed within 25 base pairs and progressively weaker effects at 50 base pairs. Beyond 50 base pairs, the magnitude of influence of CpG-disrupting variants on neighboring methylation patterns became negligible, suggesting a limited spatial range for these local sequence effects. This distance-dependent relationship indicates that sequence variants affecting CpG density and spacing might influence the local recruitment or activity of DNA methyltransferases or demethylases.

This model predicts that while cis-DMRs should often be close to CpG-disrupting SNVs, trans-DMRs—which arise from regulatory mechanisms independent of nearby sequence variation—should not be. To test this, we compared the genomic distance between each DMR and its nearest CpG-disrupting SNV. Consistent with our model, we found consistently shorter distances for cis-DMRs than trans-DMRs across all cell types examined (Mann-Whitney p < 0.001 for all comparisons)⁶⁷. This difference suggests that cis-DMRs tend to occur near sequence variants and are likely directly influenced by local changes in CpG content, while trans-DMRs show no preferential association with CpG-disrupting variants. This genomic organization is consistent with trans-DMRs being regulated by diffusible factors such as transcription factors rather than by local sequence variants, supporting a model in which cis-and trans-acting mechanisms operate through fundamentally different molecular pathways to generate species-specific methylation patterns.

Given the well-established repressive effect of promoter methylation on gene expression, we assessed whether methylation differences contribute to expression divergence by analyzing genome-wide correlations between allele-specific promoter methylation and expression levels across different genomic contexts (Figure 4A)^68,69. We first examined whether cis-acting variants drive promoter methylation consistently across cell types or in a context-dependent manner. We found that while some promoters are allele-specifically methylated across all cell types, many more showed cell type-specific cis-methylation patterns, suggesting that promoter methylation regulation is often context-dependent (Figure 4B and 4C). However, while DMRs consistently detected across cell types provide strong evidence of shared regulation, DMRs detected in only one cell type may reflect either genuine cell type-specificity or insufficient power to detect them elsewhere. Our estimate of cell type-specific regulation therefore represents an upper bound on the degree of cell type-specific methylation regulation.

Several specific examples illustrate different patterns of methylation-expression coordination (Supplemental Files 5 and 6; see Methods). The promoter of CTSF is a cell-type agnostic cis-DMR with consistent Hu>Ch methylation across cell types corresponding to coordinated Ch>Hu expression (Figure 4D and Figure 4—figure supplement 1). In contrast, a putative enhancer in LGALS8 represents a case where methylation divergence is confined to specific cellular contexts. The candidate enhancer overlaps with a cis-DMR with Hu>Ch methylation and LGALS8 exhibits allele-specific Ch>Hu expression exclusively in CNCCs, while having no divergence in methylation or expression in other cell types (Figure 4E and Figure 4—figure supplement 2). A third type of pattern is exemplified by the HOXA9 promoter, with cell type-specific hypomethylation corresponding to cell type-specific expression only in SKM (Figure 4F). These examples illustrate how different regulatory mechanisms can produce distinct patterns of methylation-expression coordination and how epigenetic changes may contribute to both species-specific and cell type-specific regulatory evolution.

Previous studies have reported weak but significant correlations between DNA methylation and gene expression divergence in primates^{7,12,14,16,70}. Consistent with these findings, our genome-wide estimates of human–chimpanzee transcriptome divergence produced similar weak but significant correlations with methylation divergence (Figure 4—figure supplement 3 and Supplemental File 6). Our comparison of CpG island (CGI) vs. non-CGI promoters also yielded trends consistent with past studies showing that CGI promoter methylation correlates more strongly with gene expression repression than does non-CGI promoter methylation^71–73 (Figure 4—figure supplement 4).

However, such genome-wide correlations may mask stronger relationships that occur within specific regulatory contexts where methylation and expression are more directly coupled. We hypothesized that our hybrid system could help identify genes exhibiting stronger methylation–expression coupling by stratifying them based on regulatory mechanisms. We reasoned that genes whose methylation and expression are both influenced by cis-acting variants— potentially driven by the same underlying genetic differences— should display tighter coupling than genes regulated through independent mechanisms.

To test this, we evaluated the extent to which methylation and expression show a negative relationship (i.e., higher methylation associated with lower expression) consistent with a repressive effect of methylation. In both CNCC and DA hybrid cells, we found that genes with cis-regulated expression were more likely to exhibit repressive patterns than those with trans-regulated expression (Figure 4—figure supplement 5). This is consistent with DNA methylation acting as an upstream cis-regulatory mechanism: genes with stronger cis-regulation are expected to show tighter methylation-expression coupling due to the greater contribution of local regulatory factors, including methylation itself. Further restricting the analysis to genes with concordant regulatory mechanisms revealed that genes with both methylation and expression being cis-regulated (“cis–cis” genes) exhibited significantly stronger negative methylation–expression correlations than “trans–trans” genes (Figure 4G, see Supplemental File 6 for detailed statistics). This supports the hypothesis that methylation–expression coupling is strongest when both are driven by local genetic variation, perhaps through shared causal variants.

To identify pathways that have evolved under lineage-specific selection, we used a two-step binomial sign test approach that tests for non-random directional bias in methylation-expression patterns in hybrids^{1,24,25,27,74–78}. The use of hybrid cell lines is critical for this analysis as it ensures that different promoters represent independent observations, controlling for trans-acting factors and environmental variation that could otherwise confound the detection of cis-regulatory effects. This test operates on the logic that neutral evolution should act like flipping a fair coin: if human-biased and chimpanzee-biased changes occur at equal background frequencies, then methylation and expression changes of genes within a pathway would accumulate randomly, showing no preference for human-biased versus chimpanzee-biased directionality. In contrast, lineage-specific selection would bias the coin flip, potentially causing functionally related gene sets to consistently favor one direction over the other rather than splitting evenly by chance.

To maximize our ability to find phenotypically impactful divergence, we adapted the standard sign test to account for directionality of both methylation and gene expression. In the first step, we applied a standard binomial sign test to all promoters showing ASM, counting the number of genes showing human-biased versus chimpanzee-biased methylation changes within each gene set. The background probability for this test was defined as the overall fraction of promoter ASM in each cell type that was human-biased. This initial analysis identified gene sets showing significant enrichment for directional methylation changes, establishing candidates for lineage-specific regulatory divergence. In the second step, we restricted our attention to genes showing a negative association between ASM and ASE. This subset represents cases where methylation changes are most likely to have a simple and direct repressive effect on transcription, filtering out more complex regulatory scenarios where methylation and expression changes may be uncoupled or subject to competing regulatory influences.

Gene sets showing significant deviation from neutral expectation in the ASM/ASE concordance test (nominal binomial p-value < 0.05 and FDR < 0.25 after multiple test correction; see Methods), with supporting evidence of directional bias in the methylation-only sign test (nominal binomial p-value < 0.05), are interpreted as likely targets of lineage-specific selection on promoter methylation that could directly impact gene expression, establishing a potential mechanistic connection to phenotype (Figure 5, Supplemental File 7). Compared to a standard sign test that focuses solely on directionally biased gene expression, this more stringent two-step filtering approach may identify higher-confidence candidates where regulatory divergence is reflected at both the methylation and expression levels⁷⁴.

Among gene sets from Kyoto Encyclopedia of Genes and Genomes (KEGG) and Human Phenotype Ontology (HPO)^79,80, we identified several with clear relevance to human-specific traits. In iPSC hybrids, multiple gene sets including “poor speech,” “highly arched eyebrow,” and “growth delay” all showed Hu>Ch methylation and Ch>Hu expression (Figure 5, Figure 6A and Figure 6B). In the SKM hybrids, genes related to “inability to walk” showed significant Hu>Ch methylation and Hu<Ch expression (Figure 5). In DA hybrids, genes involved in “intellectual disability” show significant Hu<Ch methylation and Hu>Ch expression (Figure 5 and Figure 6C). In HEP hybrids, we identified the “Hepatitis C” gene set with significant Hu>Ch methylation and Hu<Ch expression (Figure 5 and Figure 6D). In primary DPSC samples, we identified significant bias toward Hu<Ch methylation and Hu>Ch expression in genes related to tooth eruption timing, consistent with known differences in dental development between humans and chimpanzees, where human permanent molars emerge approximately 6 years apart compared to 3-4 years in chimpanzees (Figure 5—figure supplement 1)^81–83. However, the DPSC sign test results should not be interpreted as evidence of lineage-specific selection due to the use of parental rather than hybrid cell lines, which may introduce dependencies among observations due to trans-acting effects, potentially inflating statistical significance.

We then analyzed individual genes within each candidate gene set to assess whether expression bias in associated phenotypes aligns with known human-chimpanzee trait differences, focusing on genes showing repressive methylation. Multiple gene sets demonstrated directional bias in gene expression consistent with the direction of phenotypic divergence found between humans and chimpanzees. For instance, in the “highly arched eyebrow” gene set—where most genes show human down-regulation in iPSC hybrids—COLEC11 (which has 8-fold lower expression from the human allele) stands out as an outlier. COLEC11 acts as a guidance cue in directing neural crest cell migration during early development, and in humans its loss-of-function (LOF) causes craniofacial dysmorphism including highly arched eyebrows and underdeveloped brow ridge, both of which are traits unique to humans (Figure 5 and Figure 6A)⁸⁴. In the “growth delay” pathway, most genes with repressive methylation (whose LOF in humans and other species leads to delayed growth)^62–65 show lower expression from their human alleles, consistent with humans’ substantially delayed growth pattern compared to chimpanzees (Figure 6B)^85,86. Many of the genes in DA whose LOF causes “intellectual disability” may have contributed to the evolution of human cognitive abilities. For example, the high-magnitude genes with Hu>Ch ASE and Hu<Ch ASM such as GRIK2, TUBB3, EGF and CC2D2A are essential for excitatory synaptic plasticity and long-range axon wiring (Figure 6C)^87–91. In addition, most genes whose inhibition or LOF increases susceptibility or severity of Hepatitis C (e.g. MX1 and OAS3) show lower expression from human alleles in HEP, consistent with human’s heightened vulnerability to Hepatitis C viral infections (Figure 6D)^92–94.

Interestingly, we also discovered gene sets where the direction of most gene expression differences predicted the opposite of a known human/chimpanzee phenotypic difference. For example, in the “poor speech” gene set, where loss-of-function is associated with impaired speech in humans, we counterintuitively observed that most genes were down-regulated in humans compared to chimpanzees. However, the two genes with the largest fold-changes, TUBB3 and GRIK2, showed higher expression from the human allele (Figure 6E). Notably, these two genes are also among the large-magnitude genes in the predominantly human up-regulated “intellectual disability” pathway, which does align with species-specific phenotype (Figure 6C). This suggests the speculative possibility of compensatory evolution in which large, pleiotropic effects of a few genes are counterbalanced by smaller changes in many genes (Figure 6C and Figure 6E; also see Discussion). Our previous study found that genes affecting voice box show particularly extensive down-regulation in humans, consistent with our observation in the “poor speech” pathway²⁵.

Collectively, these results provide the first evidence for lineage-specific selection on human-specific DNA methylation. Our use of hybrid cell lines was critical to this discovery, as it allowed us to isolate cis-acting regulatory changes from confounding trans-acting and environmental effects. This approach reveals that polygenic selection has acted on entire sets of functionally related genes whose expression is modulated through DNA methylation. Importantly, by connecting lineage-specific methylation divergence to expression changes within functionally coherent pathways, we can generate plausible hypotheses linking regulatory evolution to human-specific phenotypes. These include craniofacial morphology, dental development, viral infection susceptibility, and neurodevelopmental characteristics underlying speech and cognition (Figures 5 and 6, Figure 5—figure supplement 1).

We utilized two previously characterized human-chimpanzee hybrid iPSC lines generated through in vitro cell fusion maintained on matrigel in mTeSR1 or mTeSR Plus medium (Stem Cell Technologies cat #85850 or cat #100–0276)²⁴. The Columbia Stem Cell Core Facility differentiated the iPSCs into multiple cell types including dopaminergic neurons (DA), skeletal myocytes (SKM), and hepatocytes (HEP) using published protocols^{27,46,96–104}. Cranial neural crest cells were differentiated as described²⁵. Primary dental pulp stem cells (DPSC) samples were obtained from humans and chimpanzees as described²⁵.

All samples were cryopreserved in liquid nitrogen before RNA extraction¹⁰⁵. Cells were gently thawed and then washed with PBS and cell pellets were collected via centrifugation at 1000 RPM for 5 min. RNA extraction was performed using the RNeasy Mini Kit (Qiagen, 74104) and DNase digestion was performed using RNase-Free DNase Set (Qiagen, 79254) following standard protocols. RNA quality was assessed using the Agilent Bioanalyzer RNA Pico assay. Only samples with an RNA integrity number (RIN) greater than or equal to 7 were used to prepare cDNA libraries. Libraries were prepared using TruSeq Stranded mRNA kits (Illumina, 20020594) and sequenced on Illumina HiSeq 4000 to generate paired-end reads. An allele-specific expression pipeline adapted from our previous work was implemented, using dual reference genome alignment to hg38 and panTro6 to eliminate mapping bias^24,25. Reads were processed with SeqPrep, mapped using STAR, deduplicated with Picard, and corrected for allelic bias using Hornet. DESeq2 was used for differential expression analysis with FDR correction, defining significant allele-specific expression as FDR < 0.05 with consistent results across both reference genomes^{24,25,27,106,107}. For details, see “Generation of allele-specific count tables” and “Identifying genes with ASE” in the Methods section in Wang et al. (2024)²⁷.

DNA was extracted and processed, and whole-genome bisulfite sequencing (WGBS) and reduced representation bisulfite sequencing (RRBS) was performed at Admera Health Biopharma Services using Illumina Novaseq 6000. To minimize reference mapping bias in cross-species comparisons, reference genomes were first masked at previously identified human-chimpanzee SNV positions using bedtools maskfasta, reducing the potential for allelic mapping preferences^27,108. Reference genome preparation was performed using Bismark’s bismark_genome_preparation with Bowtie2 as the aligner, which converts unmethylated cytosines in silico to enable bisulfite-aware alignment^109,110. All samples were aligned to both masked human (hg38) and chimpanzee (pantro6) reference genomes to assess mapping quality and species-specific biases. Reads were quality-trimmed and adapter sequences removed using Cutadapt and Trim Galore with default parameters for RRBS libraries, including the removal of artificially filled-in cytosines from the 3’ end of reads^111,112. Trimmed reads were aligned using Bismark, and methylation levels were calculated as the ratio of unconverted cytosines to total coverage at each CpG site¹⁰⁹. For hybrid samples, SNPsplit was used to assign reads to either human or chimpanzee allele¹¹³. Strand information was collapsed for symmetric CpG sites and genomic annotations were lifted between genomic assemblies hg38 and pantro6 using UCSC LiftOver with reciprocal mapping to ensure orthologous regions^114,115.

Fractional methylation of individual CpG sites was estimated using bismark_methylation_extractor from Bismark¹⁰⁹. CpG sites shared across samples were identified using bedtools intersect¹⁰⁸. PCA was performed using PCASamples from methylKit across all samples and samples within each cell type¹¹⁶. Euclidean distance matrices were computed across all samples for hierarchical clustering analysis, with visualization performed using the pheatmap package¹¹⁷. Clustering patterns were assessed to validate cell type identity and characterize species-specific methylation signatures.

Methylation data from human (H) and chimpanzee (C) parental and hybrid samples were analyzed using a beta-binomial generalized linear modeling framework to account for overdispersion inherent to bisulfite sequencing count data. The data included measurements for parental H-genome (ParH), parental C-genome (ParC), hybrid H-allele (HybH), and hybrid C-allele (HybC) for each genomic locus. For each genomic locus (site or annotated region), methylated counts were modeled using two complementary beta-binomial GLM models to separately evaluate cis and trans regulatory effects:

Model I: Cis effect test with methylation data on hybrid alleles

Model II: Trans effect test using System x Allele interaction term with methylation data in both hybrid and parental alleles

Where π is the probability of methylation, System distinguishes parental from hybrid samples, Allele distinguishes between human and chimpanzee methylation, and the interaction term captures any signals involving the change of allelic differences across generations. Statistical significance was assessed through likelihood ratio tests comparing nested models, testing for system effects, allele effects (cis-regulation), and system-by-allele interactions (trans-regulation). Genomic loci were classified into five categories: conserved, trans-only, cis-only, cis x trans interaction, and cis + trans effects:

Conserved: No significant allelic differences in either generation (neither parental nor hybrid allelic tests significant)

Regional analyses combined CpG-level methylation across defined genomic intervals (for example, promoters). Regional methylation was quantified as pooled fractional methylation, calculated as the total number of methylated cytosines divided by the total coverage across all CpG sites within each region. Nominal p-values < 0.05 from both models were used to provide a schematic overview of site-level CpG methylation in scatterplots and bar plots, preserving the full dynamic range of the data and providing input features for downstream de novo DMR identification. For reporting significant candidate loci under each regulation category as well as formal inference of region-level regulation, Benjamini–Hochberg false discovery rate (FDR) was calculated to account for multiple testing and to evaluate the robustness of inferred relative contribution of regulatory mechanisms across significance thresholds⁶⁰. The same classification and visualization were employed using FDR-corrected p-values on loci and regions to confirm that the relative contributions of regulatory classifications were not driven by significance cutoffs alone (Figure 2—figure supplement 1 and 2). To assess the robustness of cis-over-trans contributions to methylation with respect to the significance cutoffs, the proportion of significant loci and regions assigned to each regulatory category was estimated across a range of FDR cutoffs (FDR < 0.05, 0.1, 0.2, 0.25), confirming that the observed cis dominance over other divergent regulation categories (including trans) was robust to the choice of cutoff (Figure 2—figure supplement 1 and 2).

To identify clusters of CpG sites in the same regulation category, we employed the Pruned Exact Linear Time (PELT) algorithm implemented in the changepoint R package to detect shifts in mean methylation scores within each chromosome segment while accounting for a manually specified penalty against over-fitting^118–120. We adapted this method to identify regions in the genome where neighboring CpG sites are consistently classified by the previously mentioned beta-binomial model to fall under the same regulatory classification. We first segmented the entire genome into large regions where a break is assigned whenever neighboring CpG sites are more than 5000 base pairs apart. Then, we transform all CpG sites on a segment into binary sequences, where the focal regulation category is 1 and alternative categories are 0. To identify trans-DMRs, we set focal pure trans regulation to 1, and all other regulation categories to 0. To identify cis-DMRs, we set focal pure cis regulation to 1, and all other regulation categories to 0. To identify trans-DMR specifically for HOMER enrichment inputs, we used a relaxed gradient scoring approach. We set focal pure trans regulation to 1, conserved to 0.75, additive (cis + trans) and interactive (cis x trans) regulation to 0.5 and pure cis regulation category to 0. To determine the optimal penalty parameter, we applied the CROPS (Changepoints for a Range Of Penalties) method, testing penalty values ranging from 0.01 to 2.0. For each penalty range giving the same number of segments, we recorded the number of detected changepoints and corresponding model fit statistics. The optimal penalty for each segment was determined using elbow point detection via the R package kneedle, which identifies the point of maximum distance from line connecting endpoints in the penalty-versus-changepoints curve¹²¹. We then applied a homogeneity filter requiring segments to have a minimum target proportion of 0.6, where the target proportion is the fraction of CpG sites within the segment that (i) belong to the target regulatory class under the beta-binomial cis/trans classification and (ii) share a consistent species bias in methylation (all human-biased or all chimpanzee-biased), to ensure biological relevance of identified regions. For segments where fewer than three distinct penalty values yielded different numbers of changepoints, we selected the penalty that maximized the number of segments passing the homogeneity threshold. Segments without detected changepoints were evaluated directly against the homogeneity criterion to assess their potential as single-segment DMRs.

To refine segment boundaries and remove spurious edge effects before merging overlapping DMRs, we implemented a post-processing trimming step that removes segment edges containing sites with regulatory patterns inconsistent with the dominant signal within each segment. This ensures that identified DMRs represent coherent regions of coordinated methylation changes rather than artifacts of the segmentation process.

Examples of cell-type specific DMRs (HOXA9), species-specific DMRs (CTSF), and cell type-specific species-specific DMRs (LGALS8) were identified in CNCC, DA, iPSC, SKM and HEP hybrid samples using the DSS-single (Dispersion Shrinkage for Sequencing data)^122–125. For cell type-specific DMR identification, we used the formula ∼species + cell type + species:cell type to test for cell type-specific methylation differences, while species-specific DMRs analysis employed ∼species to identify consistent species differences across cell types. Downstream identification of cell-type specific species-specific DMRs identifies the regions where the interaction term species:cell type is significant. The DSS analysis incorporated smoothing across neighboring CpG sites with a smoothing span of 500 bp and required a minimum of 5 CpG sites per region. Statistical significance was determined using Wald tests with FDR < 0.05, and effect size filtering retained regions with methylation difference ≥20%. Integration with gene expression data was performed through Wilcoxon rank-sum tests to identify cell type-specific gene expression patterns associated with cell-type specific DMRs, while permutations (n=1000) were used to assess the significance of coordinated methylation-expression changes in species-specific patterns, which accounts for any background correlation between neighboring genomic features.

Human transcript coordinates and gene body annotations were obtained from the hg38.knownGene.gtf.gz file via UCSC Table Browser¹¹⁴, with transcript start positions designated as transcription start sites (TSS). CpG methylation data aggregated within ENCODE-defined promoter regions were analyzed using a beta-binomial classification framework to determine regulatory categories, while corresponding gene expression data underwent parallel hypothesis testing using a negative-binomial model (ENCODE Project Consortium, et al., 2020). Genes with the same classification for both methylation and expression—either both cis or both trans—were then analyzed separately. For genes within each regulatory category, the 2 kb surrounding each TSS was split into 10 bins of 200 bp where individual CpG sites were pooled and averaged as fractional methylation values and correlated with expression TPM values. Custom scripts used for binning and correlation are adapted from beta_profile_region.py from CpGtools (Wei et al., 2021).

To identify coordinated methylation-expression changes potentially under selection, we used a binomial sign test framework that accounts for background probabilities of concordant changes^{24,25,27,31,128,129}. Promoter regions were obtained from the ENCODE Project Consortium¹²⁶. Canonical transcripts were downloaded from the UCSC Table Browser hg38 assembly. For genes with multiple annotated promoters, we selected the promoter that is the closest to 5’ UTR of the gene, to ensure each gene is assigned one corresponding promoter¹¹⁴. Promoter regions were mapped to canonical transcripts, and functional enrichment using pathways from the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Human Phenotype Ontology (HPO) was assessed using directional sign tests that leverage the independence of cis-regulatory regions to detect non-random patterns indicative of lineage-specific selection rather than neutral drift^{1,24,25,27,75–79}. For the methylation-only sign test, gene sets with more than 200 genes or fewer than 20 genes with corresponding promoter methylation data were excluded from the analysis. Lineage-specific selection on a gene set was inferred if there was a statistically significant directionality bias in chimpanzee vs. human promoter ASM. For sign test on genes with putatively repressive methylation—that is, genes where increased methylation within 1 kb of their TSS correspond to decreased expression—we first classified genes as either consistent with repressive methylation (either Ch>Hu methylation paired with Hu>Ch expression, or Hu>Ch methylation paired with Ch>Hu expression) or not consistent. Gene sets with more than 200 genes or fewer than 10 genes with corresponding methylation and expression data were excluded from the analysis. We then reran the sign test using a new background probability defined as the overall fraction of genes with repressive methylation in each cell type. FDRs of the resulting binomial p-values were estimated by randomly permuting the signs of all genes and performing the binomial sign test 10⁴ times. To calculate FDR, we compared each pathway’s empirical binomial p-value against the distribution of permuted p-values. The expected number of false positives was estimated as the average number of pathways reaching that significance level per permutation. FDR was then defined as this expected count of false positives divided by the observed count of pathways with empirical p-values at least as extreme.

High-confidence human-chimpanzee single nucleotide variants (SNVs) were identified using a dual-reference approach with stringent filtering criteria as described previously²⁷. Allele-specific methylation within different genomic ranges (25bp, 50bp, 100bp) of CpG-disrupting variants was calculated. Proportions of regions with Hu>Ch and Hu<Ch methylation were calculated for SNVs in three categories: SNVs that disrupt CpG dinucleotides in humans, SNVs that disrupt CpG sites in chimpanzees, and SNVs that lead to no interspecies changes in CpG content between the species. The significance of the difference in relative proportions of Hu>Ch and Hu<Ch of SNVs in the three categories was assessed using a two-proportion test. For proximity tests, the Mann-Whitney U test⁶⁷ was used to compare the distribution of distances from all cis or all trans-DMRs to the closest CpG-disrupting SNV. HOMER’s findMotifsGenome.pl was used to test for motif enrichment in the trans-DMR set⁵².