Massively parallel approaches for characterizing non-coding functional variation in human evolution

Abstract

The genetic differences underlying unique phenotypes in humans compared to our closest primate relatives have long remained a mystery. Similarly, the genetic basis of adaptations between human groups during our expansion across the globe are poorly characterized. Uncovering the downstream phenotypic consequences of these genetic variants has been difficult, as a substantial portion lies in non-coding regions such as cis-regulatory elements (CREs). Here, we review recent high-throughput approaches to measure the functions of CREs and the impact of variation within them. CRISPR screens can directly perturb CREs in the genome to understand downstream impacts on gene expression and phenotypes, while massively parallel reporter assays (MPRAs) can decipher the regulatory impact of sequence variants. Machine learning has begun to be able to predict regulatory function from sequence alone, further scaling our ability to characterize genome function. Applying these tools across diverse phenotypes, model systems, and ancestries is beginning to revolutionize our understanding of non-coding variation underlying human evolution.

Introduction

The rapid expansion of modern and ancient human genomes from around the globe [1,2] as well as the increasing scale of mammalian genome databases [3] is greatly advancing our understanding of human evolutionary genetics. The resulting catalog of genetic changes in our evolutionary history may underlie the unique phenotypes that characterize our species [4,5] or the distinctive local adaptations of different ancestry groups [6]. However, identifying which of these many genetic changes are functional, or how they mechanistically impart phenotypic change, remains a major challenge for human evolutionary studies.

This difficulty is due in part to the genome’s scale and the field’s limited ability to interpret the functional impacts of genetic changes. There are an estimated 35 million single nucleotide differences between humans and chimpanzees, our nearest extant relatives [5]. Even within modern humans, any two unrelated individuals are 99.9% identical, yet differ by ~2–4 million single nucleotide variants [7]. Secondly, causal trait-associated functional variants disproportionately occur in non-coding regions [8], modifying cis-regulatory elements (CREs), such as enhancers, promoters, and silencers, that regulate cell type-specific gene expression [9]. Predicting the phenotypic impact of non-coding variation has historically been challenging, but recent advances in high-throughput functional characterization tools have begun to reveal new insights into their consequences.

Unlike the genetic code for codons, we do not yet understand the context-specific, regulatory grammar of CREs well enough to predict non-coding variant effects from sequence. Mapping of epigenetic marks and transcription factor (TF) binding correlated with CRE activity can help nominate functional variants, but these marks’ widespread genomic prevalence means they cannot usually pinpoint causal non-coding variants [10]. Moreover, CREs can target distant genes through 3D chromatin interactions, making it hard to predict the gene expression effects of non-coding variation. Genome-wide association studies (GWAS) and expression quantitative trait loci (eQTL) studies can correlate variants with complex traits, diseases, or gene expression, but are only applicable to studying within-species variation, and current studies lack sufficient diversity to interpret global variation across human ancestries [11]. Linkage disequilibrium (LD), which is often especially high in recently adaptive loci, creates additional difficulties in distinguishing the few causal variants from the many linked non-causal variants.

This review highlights recent advances in massively parallel experiments and computational prediction tools that are beginning to transform human evolution research by linking causal evolutionary genetic changes to downstream impacts. Such tools must be capable of reading out non-coding functions, have genome-scale throughput, and be applicable across phenotypically relevant cell types and developmental time points. We focus on three major areas. Firstly, advances in CRISPR technologies allow researchers to directly perturb CREs and link them to target gene expression and phenotypes. Secondly, reporter assays enable testing of tens of thousands of non-coding regions and variants simultaneously. These functional tools can be further integrated with insights from transgenic and humanized mouse models [12,13] or be applied to an increasing variety of stem cell and organoid models [4], providing a powerful toolkit for comprehensively characterizing CREs. Finally, machine learning (ML) models are increasingly capable of predicting regulatory activity and genome function from sequence alone.

For this review, we will focus on how these tools have been used to study genetic changes spanning several major periods of human evolutionary history (Figure 1A), though they can be applied to additional evolutionary questions and species. These periods include: 1) the origins of humans following our divergence from the chimpanzee and bonobo lineage, during which we evolved the derived anatomical, physiological, and cognitive traits that differentiate us from other great apes (Figure 1B-D) [4,5]; 2) the evolutionary divergence and subsequent introgression events between anatomically modern humans and our now-extinct archaic human relatives such as Neanderthals and Denisovans (Figure 1E,F) [14,15]; and 3) recent local adaptation within modern human ancestry groups to different environments, pathogens, and diets (Figure 1G) [6].

Figure 1. Non-coding variation in human evolution.

(A) Simplified human evolutionary history showing human-specific variants underlying divergence from chimpanzees/bonobos (blue), modern human divergence from archaic humans and archaic introgression into modern human ancestry groups (oranges), and variation between modern human ancestry groups (green): AFR: African, EUR: European, SAS: South Asian, EAS: East Asian, AMR: Admixed American, MEL: Melanesian. (B-G) Evolutionarily relevant non-coding variation types, including: human-specific variation, such as rapidly evolving (B) human accelerated regions (HARs) and (C) human ancestor quickly evolved regions (HAQERs), and (D) human-specific deletions in highly conserved regions (hCONDELs); modern v. archaic human genetic differences, including (E) modern-specific variants and (F) archaic introgressed variants; and (G) recent selective sweeps underlying local adaptation in modern human ancestry groups. Causal variants for each non-coding variation type highlighted in labeled color, while non-causal variants are in gray.

Mapping non-coding regions to target genes and phenotypes with CRISPR

For the vast majority of loci of interest in human evolution, the target gene(s) and phenotypes remain unknown (Figure 2A). A prime example of this challenge is seen in human accelerated regions (HARs), which are non-coding regions that are conserved across species but show accelerated substitution rates in the human lineage (Figure 1B) [16]. HARs are enriched in CREs and have been explored across multiple studies for their role in human-specific traits, especially during embryonic, skeletal, and neurodevelopment [17–19]. Many studies have used 3D chromatin maps, such as 4C, Hi-C, and PLAC-seq, to link HARs to potential target genes [13,19–23]. However, 3D chromatin methods do not establish direct causal links between CREs and gene expression.

Figure 2. High-throughput approaches for characterizing non-coding variation in human evolution.

(A) Linking non-coding variation in human evolutionary history to their downstream functional impacts is a major challenge for the field. New experimental tools are helping overcome challenges at each stage in this process: identifying relevant variation, prioritizing variants with effects on cis-regulatory elements (CREs), relating those effects to cell type-specific molecular changes, and ultimately characterizing adaptive phenotypes. (B) CRISPR perturbations modalities, including cutting/knockouts (KO), gene expression interference (i), and activation (a), enable characterizing of CREs endogenously, and can be paired with measurement of downstream gene targets and/or cellular and organismal phenotypes. (C) Massively parallel reporter assays (MPRAs) enable high-throughput screens of non-coding variants for differential regulatory activity. (D) Machine learning models trained to predict many functional assays across many cell types from sequence can enable genome-scale analyses. TF: transcription factor binding, 3D: 3D chromatin folding, GE: gene expression.

CRISPR genome manipulation has emerged as a powerful tool for investigating the function of CREs in the context of human evolution. It enables the direct, endogenous measurement of CRE function, often identifying target genes or phenotypes. CRISPR technologies encompasses a range of approaches, including enzymatically active CRISPR-Cas used for gene/CRE knockouts (CRISPR KO) or targeted allelic replacement, and catalytically deactivated Cas protein fused to a repressor/activator domain used to modulate genomic elements in methods called CRISPR interference/activation (CRISPRi/a) (Figure 2B) [24]. However, it’s important to recognize that the majority of these studies have been conducted in cell lines, with limited high-throughput in vivo applications to date. They have also generally been used to interrogate individual candidate loci from other prioritization approaches, such as those prioritized by conservation, epigenomics, chromatin contacts, association studies, reporter assays, and machine learning (ML) predictions [23,25–31]. For example, human thermoregulatory adaptation is distinguished from that of other great apes and primates by our nearly hairless bodies and high densities of eccrine sweat glands. Aldea et al. [20] identified candidate enhancers of the EN1 TF gene, screened them in skin cells using transgenic mouse assays, and validated a skin-specific enhancer of EN1 that increased eccrine gland density in a CRISPR-Cas9 humanized enhancer knock-in mouse model. CRISPR genome editing has also been used to identify downstream trans-regulatory effects of extinct archaic coding variants in a transcription factor affecting skeletal morphology (GL13 variant R1537C) [32] and a splicing factors affecting neurodevelopment (NOVA1 variant I200V) [33], thus enabling studies of the relatively few protein-coding variants underlying regulatory evolution in humans.

Pooled CRISPR screens now enable high-throughput in vitro screening of CRISPR perturbations, by combining multiplexed guide libraries with phenotypic readouts such as cell proliferation and single-cell RNA-seq (Perturb-seq) [34]. Researchers have used CRISPR screens to probe the genetic basis of evolving human phenotypes at genome scales. A pooled CRISPR-KO screen in human neural stem cells (hNSCs), which are important for determining cortical size, identified thousands of genes and enhancers that impact hNSC proliferation [35]. These included many HARs, supporting their importance in the evolution of human neurodevelopment, and an enrichment for regions linked to neurodevelopmental disorder genes, such as microcephaly and autism spectrum disorder (ASD). Similarly, a pooled CRISPRi knockdown screen found a small number of human or chimpanzee-specific essential genes, including cell-cycle progression gene regulatory networks that may contribute to increased neural proliferation in humans [36]. Finally, a pooled CRISPRi screen with cell proliferation and single-cell RNA sequencing readouts has been used to mimic large human-specific deletions (hDels) by knockdown of orthologous regions in chimpanzee pluripotent stem cells, allowing the authors to discover both cis and trans-regulatory gene targets of deleted CREs [37].

So far, pooled CRISPR screens for human evolutionary questions have focused on perturbing entire genes or non-coding regions in individual cell lines. However, emerging CRISPR technologies such as base and prime editors allow for highly precise and versatile edits, including insertions, deletions, and point mutations. Pooled base/prime editing screens have so far been used to study disease-associated enhancer variants at scale [38–40], but could also be applied to variants in evolutionary biology. For example, they could be used to interrogate the downstream effects of individual substitutions in HARs or identify target genes of small human-specific insertions and deletions. In parallel, the development of pooled CRISPR screens in organoids and animal models will enable high-throughput in vitro and in vivo CRISPR screens across multiple cell types and tissues at once [4,41]. Thus, these tools present new opportunities to directly interrogate the downstream gene regulatory and phenotypic impacts of evolutionarily important regions and variants across multiple cell types and tissue systems.

Identifying functional non-coding evolutionary variation with reporter assays

Traditional low-throughput reporter assays, such as luciferase reporters, have been used to understand the evolutionary importance of genetic changes [42], but do not scale well. In contrast, high-throughput, sequencing-based reporter assays, such as MPRAs and STARR-seq, can functionally screen thousands of non-coding variants in parallel for their effects on gene expression in a cell line of interest [43–45]. Generally, a library of putative cis-regulatory sequences is cloned upstream (in MPRA) of a minimal promoter, or in the 3’UTR (STARR-seq) where they can act to drive transcription of a reporter gene. These reporters are transfected into particular cell lines [46], or integrated into the genome for hard-to-transfect primary cells [47], and cis-regulatory output quantified by comparing the number of reporter RNA transcripts to DNA molecules (Figure 2C). By comparing alleles, MPRAs can identify variants with differential cis-regulatory effects, albeit outside their endogenous epigenomic context; and combining MPRAs with endogenous maps of CREs provides a powerful method to prioritize functional variants [48]. MPRAs typically test synthetic oligos, and thus are particularly useful for studying non-coding variants in evolutionary biology [49], such as ancestral, fixed, extinct, rare, from underrepresented ancestry groups, or completely novel variation. High-throughput reporter assays can also be used to study other non-coding processes, though they have been rarely applied to evolutionary questions, with examples for 3’ UTRs [50] and splicing [51].

MPRAs have been frequently used to characterize the regulatory functions of non-coding regions identified through computational analysis of evolving sequences. For example, MPRAs have compared the cis-regulatory output of human sequences to their chimpanzee or ancestral orthologs, both at HARS (Figure 1B), as well as CREs with species-specific epigenetic features [22,23,30,52]. Many of these studies also used the base-pair resolution of MPRAs to identify which variants functionally drive species differences, finding a variety of additive, synergistic, and compensatory interactions [22,23]. To avoid oligo synthesis limitation (200bp), which is shorter than most HARs, Girskis et al. [21] used capture-based MPRA (caMPRA) to target longer sequences from human and chimpanzee genomic DNA (~500 nts). By being able to comprehensively test all HARs and their chimpanzee orthologs, they identified a PPP1R17-regulating HAR functional during neurodevelopment. Overall, MPRA studies of HARs to date have primarily focused on the human brain (see Norman et al. [53] for an MPRA of HARs in testis).

Reporter assays have also interrogated other rapidly evolving human sequences beyond HARs. Mangan et al. [54] studied human ancestor quickly evolved regions (HAQERs), previously neutral regions that have evolved even faster than HARs (Figure 1C). Using in vivo single-cell STARR-seq (scSTARR-seq) in embryonic mouse cerebral cortex, the authors compared modern, archaic, chimpanzee, and ancestral HAQER sequences, identifying those that introduce de novo, cell-type specific CRE activity and were enriched in neurodevelopmental disease variants. These included HAQERs involved in rapid divergence following human-specific segmental duplications of FOXD4, involved in neuronal differentiation, and NBPF, associated with brain size. Another important class of putatively adaptive human-chimpanzee non-coding differences are human-specific deletions, rather than substitutions, in highly conserved regions (hCONDELs) (Figure 1D). hCONDELs may disrupt deeply conserved CRE function and thus could underlie distinctly human traits. While large hCONDELs have been studied previously [55], Xue et al. [29] used the scale and nucleotide-resolution of MPRAs to characterize thousands of small hCONDELs more amenable to reporter assays that compare the activity of two alleles. Following these screens with CRISPR genome editing and single-cell RNA-seq, the authors were able validate two regulatory hCONDELs’ effects on the neurodevelopmental genes PPP2CA and LOXL2.

MPRAs have been used to screen non-coding variants in more recent human evolution, such as those identified by comparing modern human versus Neanderthal and Denisovan genomes (Figure 1E) [14,15]. Most modern-archaic phenotypic differences, such as gene expression and soft tissue differences, are not preserved in the archaeological record, but can be investigated by functional genomics. Weiss et al. [56] found hundreds of modern human-specific single-nucleotide changes with regulatory differences using MPRAs in neural progenitor cells (NPCs), embryonic stem cells, and bone osteoblasts, chosen to reflect the evolutionary importance of the brain, development, and skeletal anatomy. They found that modern-specific cis-regulatory changes were enriched in genes of the vocal tract and brain, possibly reflecting divergence in speech production and cognition despite similar overall brain size. They also found that many modern-specific cis-regulatory changes in the brain were predicted to disrupt binding sites for ZNF281, which inhibits neuronal differentiation, and highlighted a modern-specific promoter variant of SATB2, a TF associated with brain and craniofacial phenotypes, that down-regulated expression in all three cell types. Archaic introgression also contributed to phenotypic variation in modern non-Africans, particularly phenotypes related to sun exposure, high altitudes, cold environments, and endemic pathogens of Eurasia (Figure 1F) [6,57,58]. Jagoda et al. [28] used MPRA to screen Neanderthal introgressed variants for regulatory effects in immune cells, and validated two of these variants, using CRISPR-Cas9 enhancer deletions to show they target the influenza infection responsive genes ELMSAN1 (rs11624425), and PAN2 and STAT2 (rs80317430). They also used MPRA to functionally fine-map causal non-coding variants underlying an adaptive Neanderthal-like haplotype at the COVID-19 risk-associated CCR1/5 locus [59]. Smaller-scale reporter assays have also validated regulatory variants on the adaptive OAS1/2/3 Denisovan-like haplotype involved in innate immunity in Papuans [60]. These MPRAs complement efforts to catalog immune-related eQTLs based on infection-stimulation of immune cells across human ancestry groups, which currently lack substantial Denisovan-introgressed variation [61,62].

Within modern humans, selection scans have nominated hundreds of loci with signatures of recent natural selection underlying locally adaptive traits (Figure 1G) [6,63]. Yet for only a few exemplary selective sweep loci do we currently have a good understanding of the underlying causal variants and their functions [9,64]. Moreover, high LD in recent sweep regions, potentially spanning megabases, complicates efforts to fine-map causal variants [63]. These problems mirror those present in GWAS and eQTL causal variant fine-mapping, which have been successfully tackled by reporter assays and CRISPR genome editing [46,65], suggesting these tools may also be useful for dissecting selective sweeps. Three recent studies used reporter assay fine-mapping and CRISPR validation of evolutionary loci: skin barrier adaptation at the involucrin locus in Northern/Western Europeans [25], pulmonary fibrosis and testosterone at the positively selected IVD locus in Japanese ancestries [26], and hypoxia adaptation at the Denisovan introgressed EPAS1 locus in high altitude Tibetans [27]. Alternatively, starting with a phenotype already known to underlie local adaptation, Feng et al. [31] identified candidate GWAS variants and variants with extreme allele frequency differences underlying skin pigmentation differences in Africa, then applied MPRA and CRISPR to identify regulatory single nucleotide variants in melanocytes and validate their downstream impacts on gene expression and melanin levels. This revealed repeated mutations (rs6497271-G, nearly fixed in Africans, and rs12913832-A, nearly fixed in Europeans) in the same OCA2 enhancer contributing to global diversity in hair, skin, and eye color (the AA haplotype, which had the highest enhancer activity, is most commonly found in Africans and South Asians, GA in East Asians, and GG, which had the lowest, in Europeans), and mutations in MITF (rs111969762), LEF1 (rs11939273 and rs17038630), TRPS1 (rs11985280), and BLOC1S6 (rs72713175) enhancers that contribute to light skin color in the San people of southern Africa. Combining selection scans with massively parallel functional approaches provides a powerful lens to pinpoint phenotypically impactful variation even in underrepresented ancestry groups [11].

Advancing beyond experimental characterization to prediction

While the experimental techniques discussed thus far provide crucial insights into the functional impacts of genomic elements in human evolution, they are often constrained by scale, time, and resources. Here, we transition to exploring how machine learning (ML) approaches are complementing and extending these experimental methods. ML models offer the ability to make predictions across vast genomic landscapes and diverse cellular contexts, potentially identifying targets for more focused experimental validation. Importantly, ML predictions can guide experimental design, while experimental results inform and improve ML models. We examine how ML is enhancing our predictive capabilities, while recognizing that these computational insights will ultimately require experimental validation. .

Machine learning (ML) models of genome function from DNA sequence alone are beginning to accurately predict context-specific epigenomes [66–68], 3D chromatin [69], and gene expression [70]. These sequence-to-function models aim to learn underlying cis-regulatory grammar [71], allowing them to predict variant effects in silico at scales that cannot be achieved with CRISPR or MPRA (Figure 2D). However, they require large high-quality training data in relevant contexts [71], can be challenging to adapt to biological data [72], and often struggle with distal interactions, especially in predicting gene expression [73,74]. Nonetheless, ML capabilities are rapidly improving, and current models are already achieving success in interpreting human evolution.

One area where ML models have been applied in evolution is to predict 3D chromatin contact maps. Keough et al. [75] investigated whether HARs emerged from conserved CREs where 3D chromatin folding had been rewired by human-specific structural variants (hsSVs) by using comparisons of Hi-C contacts between human and chimp NPCs, MPRAs in human midgestation telencephalon primary cells, and 3D chromatin contact map prediction of 1 Mb human and chimpanzee sequences from the ML model Akita [69]. These implicated enhancers that target neurodevelopmental genes, such as NECTIN3 and MAF, associated with hsSVs that rewire 3D chromatin folding, and HARs with MPRA activity sharing topologically associating domains (TADs) with developmental genes EFNA5, EN1, PBX3, and GBX2. The Akita model has also been applied to predict 3D chromatin differences between modern and archaic humans and within modern human populations, identifying 392 genomic regions with divergent 3D chromatin folding between modern and archaic genomes, including one containing a Neanderthal introgressed CTCF mutation that creates two TADs and is an eQTL for the insulin-like growth factor-binding protein 3 gene, IGFBP3 [76,77], identifying many loci expected to display 3D chromatin divergence even when lacking substantial sequence divergence, including modern and archaic differences in overlap with genes associated with eye, hair, skeletal, lung function, brain, and immune phenotypes. As in many of these ML approaches, functional experiments are needed to validate these predictions.

ML models have also achieved highly accurate predictions of non-coding CRE activity. Whalen et al. [23] employed MPRAs and the ML model Sei [68] to systematically interrogate the cis-regulatory effects of individual substitutions within HARs. They found that these substitutions have larger effect sizes than human polymorphisms in the same HARs. They also observed that HARs harbor compensatory mutations, which may reflect a history of adaptive trade-offs between novel human cognitive abilities and neuropsychiatric disorders, such as ASD and schizophrenia [19,78]. Li et al. [79] employed an ML model to predict thousands of putative de novo neurodevelopmental enhancers, and found that they could often arise from just a single substitution in a previously neutral sequence, in contrast to HAQERs which have many. Finally, Kaplow et al. [80] developed the ML model TACIT to predict cell type-specific enhancer conservation across 222 mammals using open chromatin data from just a few species. This was used to link non-coding regions to phenotypes that vary across mammals, such as brain size, vocal learning, and social behavior.

Looking forward, we anticipate a wealth of opportunities for ML predictions in human evolution studies. Expanding functional genomic annotations across variants [81], cellular contexts [10], and, importantly, diverse ancestry groups [11] will result in critical community resources for developing highly accurate non-coding variant effect predictors. Incorporating information about distal interactions will be key to improving gene expression prediction from sequence [73,74]. Finally, massively parallel functional experiments, including CRISPR screens and MPRAs on synthetic sequences, will be crucial for generating training and benchmarking data for sequences beyond those seen in the human genome [71].

Conclusions

Rapid progress is being made to link non-coding regulatory variants to downstream functions and phenotypes, a long-standing goal of genomics. This revolution has been enabled by high-throughput experimental systems for functional characterization, such as MPRA and CRISPR screens, as well as computational prediction tools. The field of human evolution is increasingly using this powerful toolkit to distinguish causal non-coding variants and to understand the regulatory genetic basis of adaptive traits in human evolution.

Novel CRISPR-based tools, such as pooled base/prime editing, combinatorial editing, and Perturb-seq, offer new capabilities for large-scale characterization of non-coding variants driving human evolution. These have been primarily applied in individual cell lines, such as NPCs relevant to human and chimp neurodevelopment, but combining these with emerging stem cell-derived models, such as organoids and embryoids [4], will allow us to comprehensively interrogate variant effects across multiple tissue systems and developmental time points. However, the transition from high-throughput variant discovery to organismal phenotypes remains challenging due to the low-throughput nature of in vivo studies.

Moving forward, we anticipate a multi-pronged strategy to bridge this gap. First, refining variant prioritization methods, such as MPRAs and machine learning predictions, will be crucial for identifying the most promising candidate variants for in-depth study. Second, linking prioritized variants to downstream genes, cell types, tissues, and pathways, using techniques such 3D chromatin maps, pooled CRISPR screens, and single-cell multi-omics [82,83]. Third, the development of high-throughput in vivo methods, such as advanced CRISPR techniques in transgenic that can make multiple edits or complex humanized mouse models, will be essential for scalable characterization of organismal phenotypes, such as those involved in skeletal development [12,13]. Combining these approaches will allow researchers to balance detailed characterization of candidate variants with screening approaches, thus linking cellular mechanisms to complex organismal phenotypes.

These genome editing and perturbation tools will be crucial for tackling the wide range of adaptive human traits, as well as answering fundamental questions about adaptive and antagonistic pleiotropy in human evolution. Adaptive pleiotropy occurs when a genetic variant positively affects multiple traits, while antagonistic pleiotropy refers to variants that benefit one trait while negatively impacting another. For example, variants that increased brain size in human evolution may have also led to increased difficulty in childbirth [84,85]. Understanding these trade-offs is crucial for a complete picture of human adaptive evolution. Moreover, these approaches could allow us to identify non-coding variants underlying the many physiological differences between humans and other great apes, such as bipedalism, digestion, metabolism, and life history [4,5]. In the future, new CRISPR engineering approaches may also be used to characterize understudied variation types, such as human-specific tandem repeats, transposable elements or complex genomic rearrangements, which have been linked to evolutionary innovation.

MPRAs have enabled unprecedented precision into understanding the regulatory effects of individual mutations across evolution. However, they have mostly been applied to individual cell lines, with the exception of the relatively low-throughput in vivo scSTARR-seq assay used to characterize HAQERs [54]. They also test sequences outside their endogenous context, and thus should be interpreted in the context of endogenous maps of CRE activity and gene targets, and CRISPR experimental validations. Increasing the throughput of single cell and in vivo MPRAs will enable large-scale comparative regulatory insights across many tissues and cell types. For recent adaptations that vary across modern humans, beneficial variants may be functional only in the context of particular environmental stimuli, such as immune stimulation [59], hypoxia [27], or sun exposure [31]. MPRAs can be used to uncover these gene-by-environment interactions and identify differential regulatory effects under environmental perturbations, such as pathogens, micronutrients, and environmental toxins [86]. Integrating MPRA and CRISPR screens across cellular and environmental contexts will greatly expand the number of selective sweep loci with known causal variants and function.

Finally, ML models are allowing researchers to move beyond experiments to in silico predictions, thus greatly expanding the scope of variants and cell types that can be examined in a single study. Future ML models will show rapid advances in capabilities, such as predicting distal interactions and individual gene expression from sequence [65,68,69]. These models will benefit from expansion of bulk and single cell functional genomic resources, including CRISPR screens and MPRAs, across variants [75], cell types [10], diverse ancestries [11], and closely related species [3]. Moreover, understanding complex adaptive phenotypes will continue to require biobank-scale association studies, and may be aided by ML models for image-based phenotyping, such as for skeletal form [18]. Combining these ML tools for variant prioritization with experimental characterization validation in organisms provides a roadmap to transform our understanding of human origins and adaptive evolution.

Highlights.

• Human evolution/adaptation is substantially driven by non-coding, regulatory changes

• Novel high-throughput methods can characterize the function of evolutionarily relevant loci and the variation within them

• CRISPR genomic perturbation screens endogenously measure molecular/cellular phenotypes

• Massively Parallel Reporter Assays (MPRAs) assay for regulatory impact of non-coding variants across cell types

• Machine learning predicts regulatory function from sequence at genome-scale