Evidence of abundant purifying selection in humans for recently-acquired regulatory functions

Initial sequencing of the human genome revealed that 98.5% of human DNA does not code for protein (1), raising the question of what fraction of the remaining genome is functional. Mammalian conservation suggests that ~5% of the human genome (2–3) is conserved due to non-coding and regulatory roles, but more than 80% is transcribed, bound by a regulator, or associated with chromatin states suggestive of regulatory functions (4–6). This discrepancy may result from non-consequential biochemical activity or lineage-specific constraint (7–8). Similarly, evolutionary turnover in regulatory regions (9–11) may be due to non-consequential activity in neutrally-evolving regions in each species, or turnover in functional elements associated with turnover in activity. To resolve these questions, we need new methods for measuring constraint within a species, rather than between species.

Single nucleotide polymorphisms (SNPs) within human populations have been identified only every 153 bases per average (12), compared to 4.5 substitutions per site among the genomes of 29 mammals (2), making it impossible to detect individual constrained elements (13). Instead, aggregate measures of human diversity across thousands of dispersed elements are needed. Such measures have been used to show that human constraint correlates with mammalian conservation (4, 14–17), mRNA splice sites (18), regulatory elements (19), and that similar selective pressures act in human and across mammals (2). However, differences between mammalian and human constraint remain unresolved. Recent positive selection has been detected by unexpectedly many recent substitutions (20) or extreme patterns of linkage disequilibrium (LD) and population differentiation (21). However, recent negative selection has not been investigated, as the paucity of variants segregating in the global population makes a selective decrease in the diversity of any given locus indistinguishable from a fortuitous one.

Combining population genomic information from the 1000 Genomes Project (12) and biochemical data of the ENCODE project (5) we estimated constraint associated with diverse genomic functions in aggregate over 1567 Mb of `previously-unannotated' regions encompassing 4.7 million SNPs, excluding exons, proximal promoter regions, and artifact-prone regions (22) (Fig. 1A). On the basis of SNP density, heterozygosity, and derived allele frequency (DAF), we developed a statistical procedure for measuring genome-wide constraint accounting for mutation rate biases and interdependence of allele frequencies due to LD (22). All P values are derived from this test unless otherwise noted. To distinguish whether the increased human constraint in active regions (5) could be due solely to mammalian conservation (Figs. 1B, S1), rather than lineage-specific constraint, we specifically studied regions not conserved across mammals.

Remarkably, non-conserved active regions showed significant evidence of purifying selection: SNP density was 10% lower (P<10⁻⁶⁴), heterozygosity 13% (P<10⁻⁸⁵), and DAF 5% (P<10⁻⁶⁵), compared to reductions of 28%, 33%, and 16% respectively for conserved regions. As non-conserved regions cover a >10-fold larger fraction of the genome, this suggests that a significant fraction of human constraint lies outside mammalian-conserved regions. The observed decrease in diversity is not due to undetected conserved regions or the threshold used to defined conserved elements (Fig. S2), nor to background selection (23) (Fig. 1C,D), biased gene conversion (Table S1), or decreased mapping to non-reference alleles (22) (Table S2).

The level of human-specific constraint varies with the observed biochemical activity (Figs. 2, S3–S5, Table S3–S4). Short non-coding RNAs are as strongly constrained as protein-coding regions. Long non-coding RNAs (lncRNAs) are significantly constrained in human, even though they lack significant mammalian conservation (5), suggesting primarily lineage-specific functions. These results are not explained by local mutation rate variation nor transcription-mediated repair, as DAF is robust to both.

We also found human-specific constraint across non-conserved regulatory features (Fig. 2C,D). Regulatory motifs bound by their regulators show constraint similar to coding regions, and consistently higher than for non-bound instances (P = 9.5 × 10⁻⁷, binomial test) (Fig. 3). Regulatory regions defined by different assays, including DNase hypersensitivity and transcription factor binding, show significant and similar levels of human constraint. Different chromatin states (5, 24) show levels of constraint according to their roles (Fig. 2E,F), with promoter states similar to previously-annotated TSS-proximal regions, enhancer states significant but weaker, and insulators similar to background regions, consistent with enhancer and promoter regions requiring a larger number of motifs than insulator regions. In contrast, regions that do not overlap with active ENCODE elements and inactive chromatin states show even lower constraint than ancestral repeats (Fig. 2B,D,F), suggesting they may provide a more accurate neutral reference than repeats that can have exapted functions (25).

Comparison with primate constraint suggests evolutionary turnover. Mammalian-conserved regions lacking ENCODE activity show reduced human constraint relative to active regions (SNP density P<10⁻⁴¹, heterozygosity P<10⁻⁵², DAF P<10⁻¹⁴) (Fig. 1B, S1), suggesting recent loss in function and activity. These also show higher primate divergence relative to active regions, suggesting some loss of constraint likely predates human-macaque divergence. Conversely, a fraction of lineage-specific elements likely arose in the common ancestor of primates, as human-macaque divergence mirrors human diversity for both active and inactive non-conserved regions (Fig. S6).

To gain insights into the functional adaptations likely involved in this turnover, we applied our aggregation approach to regulatory regions associated with genes of different functions (22). We found that highly-constrained non-conserved enhancers are associated with retinal cone cell development (P<10⁻⁴ in GO) and nerve growth (P<10⁻⁵ in GO, Reactome, and KEGG; Fig. S7). This evidence of recent purifying selection for regulation of the nervous system and color vision is intriguing given their accelerated evolution in primates (20, 26–27).

We next studied how the number of aggregated regions affects the ability to discriminate functional elements based on their increased human constraint (Fig. S8). We found no discriminative power for individual elements, despite a significant global reduction in heterozygosity (P<10⁻²⁰, Mann-Whitney-Wilcoxon test on heterozygosity of individual elements), but discriminative power increased significantly as the sample size grew (22).

We estimated the proportion of the human genome under constraint (PUC) after correcting for background selection (Fig. S9), and found remarkable agreement between our orthogonal metrics (Fig. 4A). We estimate that an additional 137 Mb (4%) of the human genome is under lineage-specific purifying selection (Table S6), consistent with a recent cross-species extrapolation (28).

Our results suggest that almost half of human constraint lies outside mammalian-conserved regions, even though the strength of human constraint is higher in conserved elements. Protein-coding constraint occurs primarily in conserved regions while regulatory constraint is primarily lineage-specific (Fig. S10), as proposed during mammalian radiation (29). While differences in activity between mammals (10–11) can be interpreted as lack of functional constraint (30), our results suggest instead that turnover in activity is accompanied by turnover in selective constraint. A minority of new regulatory elements lie in recently-acquired primate specific regions (5) but the bulk lies in mammalian-aligned regions that provided raw materials for regulatory innovation.

Genome-wide association studies suggest that 85% of disease-associated variants are non-coding (8), a fraction similar to the proportion of human constraint we estimate lies outside protein-coding regions (Table S6). This suggests that mutations outside conserved elements play important roles in both human evolution and disease, and that large-scale experimental assays in multiple individuals, cell types and populations can provide a means to their systematic discovery.