Comparing Neanderthal Introgression Maps Reveals Core Agreement But Substantial Heterogeneity

Abstract

Statistical methods to identify Neanderthal ancestry in modern human genomes rest on varying assumptions and inputs. Nonetheless, most studies of introgression use only a single method to define Neanderthal ancestry. Due to a lack of “ground truth,” we have a limited understanding of the accuracy, comparative strengths and weaknesses, and the sensitivity of downstream conclusions for these methods. Here, we performed large-scale comparisons of 14 genome-wide introgression maps computed by 11 representative Neanderthal introgression detection algorithms: admixfrog, ArchaicSeeker2, ArchIE, ARGweaver-D, CRF, DICAL-ADMIX, hmmix, IBDmix, SARGE, Sprime, and S*. These algorithms span statistical approaches based on summary statistics, probabilistic modeling, and machine learning, and vary in their use of archaic, modern, and simulated genomes as input. Our results highlight a core set of regions predicted by nearly all methods, as well as substantial heterogeneity in commonly used Neanderthal introgression maps, especially at the individual genome level. Furthermore, we find that downstream analyses may result in different conclusions depending on the map used. Thus, we recommend careful consideration of map(s) chosen for downstream analysis and support the use of multiple maps to ensure robustness of conclusions. We make integrated prediction sets available, enabling further understanding of Neanderthal introgression's legacy on modern humans.

Introduction

One of the closest extinct relatives of modern humans, the Neanderthals, lived across western and central Eurasia for hundreds of thousands of years before the ancestors of modern Eurasians migrated out of Africa (Higham et al. 2014). Today, the Neanderthal legacy persists through artifacts, fossil remains, and Neanderthal genome sequences (Prüfer et al. 2014, 2017; Mafessoni et al. 2020). Analyses of Neanderthal and modern human genomes revealed multiple periods of interbreeding among the ancestors of modern humans, Neanderthals, and other ancient hominin groups, like the Denisovans. These events resulted in substantial introgression—the exchange and integration of genetic information between these groups (Ahlquist et al. 2021). Here, we focus on introgression from Neanderthals to modern humans.

Most Neanderthal ancestry observed in modern humans today originated from an introgression event that occurred 40,000 to 50,000 years ago, when humans and Neanderthals co-existed (Green et al. 2010; Iasi et al. 2024). The patterns of Neanderthal ancestry were shaped by natural selection, drift, and demographic processes over the subsequent ∼2,000 generations. In modern humans not from sub-Saharan Africa, 1% to 2% of each individual's genome consists of Neanderthal ancestry (Prüfer et al. 2014; Vernot et al. 2016). Here, we focus on the wide range of statistical approaches that have been applied to infer Neanderthal introgressed DNA in modern humans (Huang et al. 2025).

Neanderthal introgression detection methods can be broadly classified based on the statistical methods and genome inputs used. Probabilistic models, including hidden Markov models (HMMs) (Steinrücken et al. 2018; Peter 2020; Yuan et al. 2021) and Conditional Random Fields (CRF) (Sankararaman et al. 2014, 2016), rely on using a Neanderthal genome as an archaic reference and genomes from modern humans with limited Neanderthal introgression as outgroups. Commonly, Yoruba in Ibadan, Nigeria (YRI) genomes from the 1000 Genomes Project (1KG) (Auton et al. 2015) are used as the outgroup. There are also HMM models that do not use an archaic reference but still rely on genomes from Africa as nonintrogressed outgroups (Skov et al. 2018, 2020).

Dynamic programming-based approaches, like the S* statistic (Vernot et al. 2016), identify sets of SNPs in high linkage disequilibrium (LD) that differ from an introgression-free outgroup and match expected lengths based on the timing of introgression. For example, to identify Neanderthal introgressed segments, previous applications of the S* approach have used 1KG YRI genomes as outgroups (Vernot et al. 2016). The S* method was updated in Sprime (Browning et al. 2018), allowing lower levels of archaic introgression by removing the fixed-window approach and accounting for the frequency of introgressed haplotypes, mutation rates, and recombination rates. These approaches do not require a reference genome for the source introgressing population, but identified candidate regions can be compared to archaic reference genomes if available.

More recently, methods have been developed to detect introgression without requiring source reference genomes or modern human genomes from Africa as an outgroup. These methods have enabled the study of Neanderthal ancestry outside Eurasia. IBDMix (Chen et al. 2020; Li et al. 2024) detects Neanderthal introgression using shared identity by descent (IBD) between modern human and Neanderthal genomes, computing a log-odds score indicating the likelihood of shared alleles due to IBD. Notably, IBDMix does not rely on an introgression-free modern human reference. As a result, applications of this approach have revealed small amounts of Neanderthal ancestry in genomes of individuals from Africa, likely due to modern humans migrating back to Africa from Eurasia while carrying Neanderthal introgressed loci. Other methods, like ARGweaver-D (Hubisz et al. 2020) and Speedy Ancestral Recombination Graph Estimator (SARGE) (Schaefer et al. 2021), identified Neanderthal introgression by inferring ancestral recombination graphs (ARGs) to reconstruct the full genealogy of modern human and archaic genomes. Additionally, Archaic Introgression Explorer (ArchIE) (Durvasula and Sankararaman 2019) detected Neanderthal introgression using a logistic regression model trained on simulated introgressed genotypes to infer Neanderthal introgressed regions, without the use of archaic or modern nonintrogressed human reference genomes.

The output of these algorithms consists of predicted introgressed genomic regions and/or variants in modern individuals. We refer to these outputs collectively as introgression maps. Most maps agree on the approximate proportions of Neanderthal ancestry present in modern populations. However, agreement between introgression maps has not been comprehensively quantified. Differences between published introgression maps aggregated across individuals are expected due to true differences in individual-level introgression patterns, but limited cross-method comparisons suggest more substantial differences (Skov et al. 2018; Steinrücken et al. 2018; Durvasula and Sankararaman 2019; Chen et al. 2020; Schaefer et al. 2021; Yuan et al. 2021). For example, introgression deserts, or regions depleted of Neanderthal ancestry across individuals, are thought to have previously harbored introgressed alleles that were disadvantageous in modern humans and thus removed through negative selection (Harris and Nielsen 2016; Sankararaman et al. 2016; McCoy et al. 2017; Petr et al. 2019; Telis et al. 2020; Zhang et al. 2025). Several studies have identified introgression deserts, but the criteria used vary, resulting in different sets of desert regions available for further analysis (Sankararaman et al. 2016; Vernot et al. 2016; Chen et al. 2020; Skov et al. 2020; Yuan et al. 2021). These maps are commonly used to investigate evolution following admixture (Sankararaman et al. 2016; Vernot et al. 2016; Chen et al. 2020; Rinker et al. 2020; Skov et al. 2020) and the contribution of Neanderthal introgression to various modern human traits (Simonti et al. 2016; Zeberg and Pääbo 2020; McArthur et al. 2021; Dannemann et al. 2022; Koller et al. 2022; Reilly et al. 2022; Velazquez-Arcelay et al. 2023; Wei et al. 2023). However, the choice of introgression map could influence downstream analysis of evolutionary processes that shaped Neanderthal introgression and how Neanderthal ancestry contributes to modern human traits.

Here, we compare representative Neanderthal introgression maps to understand variation in introgression predictions across modern human populations, facilitate further method development, and establish a resource to support the robustness of future results. Due to the lack of ground truth, we focus on identifying similarities and differences in available introgression maps at base-pair resolution across individuals and in aggregate. We also investigate the evolutionary and phenotypic properties of introgressed regions in each map. Our results provide a resource for analyzing Neanderthal ancestry in the modern human genome, and a deeper understanding of Neanderthal introgression patterns will help reveal structural and functional differences between Neanderthals and modern humans that have yet to be fully uncovered. Our approach also provides an evaluation framework that can be applied to introgression events from other groups and time periods, such as from Denisovans into modern humans.

Results

Introgressed regions differ in length and genomic coverage between maps

To quantify genomic patterns of existing introgression predictions, we integrated 14 published introgression maps across 11 algorithms (Table 1; Methods; Figure S1) and compared their genome-wide coverage. Given that these introgression maps are based on very different numbers of individuals and populations (Table 1), we caution that the comparisons in this section are strongly influenced by the number and diversity of individuals considered by each method. For example, previous applications of IBDMix, which does not rely on nonintrogressed outgroups, identified the largest number of introgressed base pairs based on 2,504 1KG genomes. ARGweaver-D identified the smallest amount of introgression since it uses only four individuals, two of whom are from Africa, for introgression inference. The total number of base pairs detected as introgressed by a method begins to plateau after considering a few hundred individuals, although the dynamics vary across populations (Kerdoncuff et al. 2024).

Table 1.

The introgression maps analyzed in our study.

Introgression map	Genomic scale	Reporting level	N	Populations represented in target genomes
ArchaicSeeker2 (Yuan et al. 2021)	Regions	Individual	1,511	EAS, SAS, EUR, Oceania (1KG, SGDP)
ArchIE (Durvasula et al. 2019)	Regions	Individual	99	CEU (1KG)
ARGweaver-D (Hubisz et al. 2020)	Regions	Individual	4	Africa, Oceania, West Eurasia (SGDP)
CRF14 (Sankararaman et al. 2014)	Variants and Regions	Population	665	EUR, EAS (1KG)
CRF16 (Sankararaman et al. 2016)	Variants and Regions	Population	257	AMR, Central Asia, EAS, Oceania, SAS, EUR (SGDP)
DICAL-ADMIX (Steinrücken et al. 2018)	Regions	Individual	282	CEU, CHBS (1KG)
IBDMix (Li et al. 2024)	Regions	Individual	2,504	AFR, AMR, EAS, EUR, SAS (1KG)
SARGE (Schaefer et al. 2021)	Regions	Individual	279	Oceania, West Eurasia, East Asia, Africa, Africa2, America, Central Asia, Siberia, South Asia (SGDP)
S* (Vernot et al. 2016)	Variants and Regions	Individual	1,531	SAS, ASN, PNG, EUR (1KG), Melanesia
Sprime (Browning et al. 2018)	Variants and Regions	Population	1,843	EAS, SAS, EUR, AMR, Oceania (1KG, SGDP)
hmmix20 (Skov et al. 2020)	Variants and Regions	Population	27,566	Iceland (deCODE)
hmmix18 (Skov et al. 2018)	Variants and Regions	Individual	3,134	EUR, EAS, AMR, Central South Asia, Oceania, Middle East (1KG and HGDP)
admixfrog (Iasi et al. 2024)	Regions	Individual	2,116	EUR, EAS, AMR, SAS, Oceania, West Eurasia, East Asia, Africa, America, Central South Asia, South Asia (1KG and SGDP)

For each map, we describe whether Neanderthal introgression is reported at the region or variant level. Additionally, we indicate whether loci are provided at the individual level or only as population-level introgression maps. Lastly, we describe the number of individuals included in individual-level analyses, as well as the number reported for each original study in population-level data. For Sankararaman et al. (2016), we consider two sets of predictions: (i) using Simons Genome Diversity Panel (SGDP) African and Denisovan genomes as outgroups and (ii) using only SGDP African genomes as an outgroup. See Figure S1 for additional details on the algorithms used to generate the maps.

Despite these caveats, comparisons of published introgression maps combined across individuals are still valuable given their common use in studies of Neanderthal introgression. To mitigate the impact of the number of individuals considered in each introgression map, we focus in subsequent sections on predictions made in the same individuals for each map. Due to the small number of individuals analyzed by ARGweaver-D, we excluded this method in map-wide comparisons.

Overall, 2,240,223,121 bases, or ∼84.5% of the autosomal human genome (without assembly gaps), were called as introgressed in at least one map (Fig. 1a). This is substantially greater than the proportion predicted by any single map (61.4%), and 16.6% of the introgressed base pairs were unique to a single map (Fig. 2b). The introgression maps also vary substantially in the lengths of their predicted introgressed regions (Fig. 1b).

Figure 1.

The genomic coverage and length distribution of each introgression map. a) Amount of the genome (in millions of base pairs, or Mb) identified as having Neanderthal introgression in at least one individual from each map's published predictions. The coverage values reflect the union of introgressed loci across individuals and populations. Percentages outside each bar reflect the percentage of nongap autosomal genome coverage for each map. Note that there are differences in the number and diversity of individuals analyzed, which contribute substantially to the variation (see subsequent sections for individual-level comparisons). b) Boxplots of the distribution of introgressed region lengths in log₁₀(base pairs) for each map. Ns to the right of each boxplot reflect the number of distinct, noncontinuous regions in each map. Boxes reflect the interquartile range, and whiskers denote 1.5 times the interquartile range.

Figure 2.

The overlap between introgression maps. a) Upset plot for genomic overlap amounts among regions identified by different combinations of introgression maps. The top 15 combinations are shown. b) Histogram of the number of introgression maps supporting each base pair predicted as introgressed. c) The number of unique base pairs predicted as introgressed by each map. d) Jaccard similarities—the intersection divided by the union at the base pair level—are given above the diagonal. Values in the lower triangle reflect normalized Jaccard similarity, where the size of the smaller set is the denominator. Introgression maps are ordered by hierarchical clustering based on their similarity, and distance values reflect the maximum distances between two sets, computed using farthest point clustering.

Introgression maps have core agreement, but substantial discordance

Next, we quantified the number of maps that identified each genomic base pair as introgressed (Fig. 2a to c). The most common scenarios are (i) regions predicted only by ArchIE (131.80 Mb), (ii) regions predicted as introgressed by all maps except ArchIE (128.97 Mb), (iii) regions predicted as introgressed by all maps except ArchIE and Sprime (103.24 Mb), and (iv) regions predicted as introgressed by all maps (84.76 Mb). More broadly, regions were either predicted by many maps or only a few (Fig. 2b). This suggests a core of regions identified by most introgression maps. Methods varied substantially in the amount of the genome they alone predicted as introgressed (Fig. 2c).

To quantify the genome-wide agreement between introgression maps, we computed the Jaccard similarity—the size of the intersection divided by the size of the union—at the base pair level between pairs of maps (Fig. 2d). Additionally, we computed normalized Jaccard similarity, where the size of the smaller set is the denominator. Raw Jaccard similarities ranged between 0.19 and 0.76, with ArchIE as the most discordant map. The normalized values are higher due to size differences between maps but show similar trends to the raw Jaccard similarities. To evaluate the amount of overlap observed between a pair of maps, we estimated the null distribution of overlap expected if the regions were randomly distributed, generating 100 shuffles for each pair of maps. All observed Jaccard similarities were significantly greater than expected (Figure S5).

We caution that this comparison is influenced by the amount of the genome predicted as introgressed by a method, and that if introgression maps identify the same regions in their highest-confidence predictions, this pattern would be obscured in the whole-map comparison. To account for differences in introgression map size, we reduced each introgression map to include only the highest-scoring regions, matching the smallest autosomal map (∼591 Mb from Sprime). This resulted in lower overall similarities, but in some cases, higher similarity between maps generated by similar methods (Figure S6).

Outside of autosomal regions, the X chromosome has been previously observed to be depleted in Neanderthal introgression due to potential hybrid incompatibilities (Sankararaman et al. 2014) and greater negative selection (Harris and Nielsen 2016; Skov et al. 2023). Only 6 out of 12 maps provide introgression predictions for the X chromosome: hmmix20, CRF16 (1) and (2), CRF14, DICAL-ADMIX, and an additional map included for this analysis, hmmix23, which uses hmmix to identify introgression in the X chromosome using 162 male genomes from SGDP (Skov et al. 2023). ARGweaver-D also provided X chromosome predictions, but we again excluded this method due to the small number of individuals considered. Across the X chromosome, 28.9% (43,686,164 bp) is called as introgressed in at least one map. Introgression is depleted in the X chromosome compared to autosomes across methods: Each map has lower introgression coverage percentages in the X chromosome compared to autosomes. For introgressed regions in the X chromosome, agreement is more varied than in autosomal-only comparisons after matching to the smallest set, hmmix20 (Figures S7 to S10). Introgression predictions in the X chromosome agree more when the comparison is not limited to top-scoring regions (Figure S9).

In addition to region-level introgression maps, six methods call specific Neanderthal introgressed variants: CRF16 (1) and (2), CRF 14, hmmix20, hmmix18, S*, and Sprime. At this variant level, raw Jaccard values range from 0.12 to 0.66, and CRF16 (1) and CRF16 (2) have the highest raw Jaccard similarity of 0.66. hmmix18 and hmmix20 form another cluster with a raw Jaccard of 0.32, followed by a cluster between CRF14 and S*, with a raw Jaccard value of 0.17. Sprime is an outgroup compared to all other variant maps, with raw Jaccards ranging from 0.12 to 0.27 (Figure S11). Four methods identify variants in the X chromosome: hmmix20, CRF16 (1) and (2), and CRF 14. X chromosome introgressed variants have the highest agreement among CRF methods (Jaccards = 0.25 to 0.61) (Figure S13).

Individual-level comparisons highlight discordance between introgression maps

Next, we investigated how introgression maps from different methods vary in their predictions on the same individual genomes. We focus on autosomal introgressed regions predicted in several sets of shared individuals between maps.

These comparisons reveal that different introgression detection methods often yield substantially different predictions when applied to the same individual's genome, with average Jaccard similarities ranging from 0.05 to 0.64 across 84 individuals from Utah with Northern and Western European ancestry (CEU individuals) from 1KG (Fig. 3a). ArchIE's maps are also the most discordant at the individual level, with an average Jaccard with other maps lower than 0.1 and was the only map where empirical P-values were not significant for all observed Jaccard values (Fig. 3a). IBDMix was the next most discordant map, but it had much higher significant agreement with other maps, with Jaccard statistics ranging from 0.40 to 0.51 (Fig. 3b). Even though DICAL-ADMIX and ArchaicSeeker2 were in different clusters in the genome-wide comparison, their individual-level introgression maps agreed more than any other pair of methods for these individuals. This highlights the importance of individual-level comparisons. We did not include maps that lacked predictions for these individuals.

Figure 3.

Comparison of introgression maps at the individual level highlights significant overlap and substantial discordance. a) Average Jaccard similarities (lower left triangle) and average z-scores (upper right triangle) between individual-level introgression maps for all pairs of maps for 84 individuals from the CEU 1KG population. The z-scores for observed Jaccard similarities were computed from null distributions based on 100 shuffles for each pair of maps. All individual-level comparisons for each pair of maps yielded significantly more overlap than expected, except for ArchIE (empirical P-values < 0.05, bolded cells). The size of the dot indicates the standard deviation (SD) of the Jaccard similarities (lower triangle) or z-scores (upper triangle) for each pairwise comparison. b) Violin plot of pairwise Jaccard similarity distributions for the individual-level DICAL-ADMIX introgression maps compared to all other maps over the 84 CEU individuals, for the red highlighted row in a. c) Violin plot of pairwise z-score distributions for the individual-level DICAL-ADMIX introgression maps compared to all other maps over the 84 CEU individuals. d) Empirical P-value distributions for the individual-level DICAL-ADMIX introgression maps compared to all other maps over the 84 CEU individuals. e) The same as a, but for 230 shared SGDP individuals across methods that made predictions for these individuals. f) The same as a, but for 271 shared CEU and CHBS 1KG individuals across methods.

For other groups of individuals with predictions from multiple methods (271 CEU and CHBS individuals, 230 SGDP individuals, 14 Papuan SGDP individuals, and 1 Papuan SGDP individual), we observed similar patterns of statistically significant, but moderate in magnitude of agreement between methods (Jaccard similarities ∼0.39 to 0.65; Figs. 3e, f, and Figures S15 to S18).

We also analyzed agreement between individual-level maps of introgressed regions in African genomes. Based on two shared individuals from SGDP between admixfrog, ARGweaver-D, and SARGE, raw Jaccard similarities were statistically significant and ranged from 0.032 to 0.103 (Fig. 4a; Figure S19). Compared to non-African individuals, the similarity of introgression maps is lower for African genomes (Fig. 4a), even after accounting for the lower expected overlap between African maps due to much lower levels of introgression (Fig. 4b). Introgression prediction in genomes from Africa remains a challenge due to the low levels of Neanderthal ancestry present and their common use as an outgroup in introgression methods. Although our comparisons are limited to a small number of methods and individuals, they reveal consistently more discordance between introgression maps for African individuals compared to Eurasians.

Figure 4.

Individual-level introgression map Jaccard similarity and z-score distributions across superpopulations. The distribution of a) Jaccard similarities and b) z-scores for each pair of maps for shared individuals separated by superpopulations from 1KG and SGDP. These include genomes from Africa (AFR), Oceania, South Asia (SAS), Central Asia, Europe (EUR), admixed individuals from the Americas (AMR), and East Asia (EAS). Asterisks denote Bonferroni-corrected P-values: *P < 0.05, **P < 0.01, and ***P < 0.001 for pairwise Mann–Whitney U tests comparing pairwise Jaccard similarity and z-score distributions for individuals from Africa vs. each other population. The similarity of introgression maps predicted by different methods on the same individuals from Africa is significantly lower than for other populations, even after accounting for the lower expected overlap due to lower levels of Neanderthal ancestry. All Jaccard similarities between individuals have P-values < 0.05, based on an empirical null distribution computed from 100 shuffles for each pair of maps.

Introgressed regions in all maps have less evidence of background selection compared to regions without introgression

Analyses of the first Neanderthal introgression maps found that regions with high levels of Neanderthal ancestry have less evidence of background selection from linked functional elements than regions with low levels of Neanderthal ancestry (Sankararaman et al. 2014). As a first step in exploring the robustness of conclusions about patterns of Neanderthal ancestry to the introgression map used, we tested this result across the introgression maps considered here. We computed the average B statistic, which quantifies the expected fraction of neutral diversity at a site, across regions with introgression versus those without introgression for each map. Lower B values indicate stronger background selection (McVicker et al. 2009).

Each introgression map had significantly less evidence of background selection, or higher average B values, in regions with predicted introgression compared to regions without inferred introgression (Fig. 5). IBDMix and ArchIE had the largest differences (ratio of averaged introgressed vs. nonintrogressed B values: 1.075 and 1.069, respectively), while SARGE and ARGweaver had the smallest differences (ratio of averaged introgressed vs. nonintrogressed B values: 1.019 and 1.004, respectively). Thus, the previously documented relationship between Neanderthal ancestry and the strength of background selection is robust to the introgression map used. While power to detect Neanderthal introgressed regions depends on many factors, simulations suggest that selection against Neanderthal alleles, rather than differences in power, explains the depletion of Neanderthal ancestry in regions with many constrained functional elements (Sankararaman et al. 2014). However, more work is needed to understand the power and sensitivity of these methods in different genomic contexts.

Figure 5.

Regions from each Neanderthal introgression map have less evidence of background selection. Average background selection scores as quantified by the B statistic (McVicker et al. 2009) for introgressed and nonintrogressed autosomal regions from each map. B values were summarized in 500 bp bins (Telis et al. 2020). Higher B values indicate less background selection. Error bars reflect 95% confidence intervals based on 1,000 bootstraps. Asterisks denote P-values: *P < 0.05, **P < 0.01, and ***P < 0.001 for Mann–Whitney U tests comparing B statistic values between introgressed and nonintrogressed windows.

Phenotype annotations enriched in different introgression maps are discordant

Gene set annotation enrichment analysis is commonly performed on genes in regions with or without evidence of introgression. These analyses have yielded hypotheses about properties of the genome that retain Neanderthal ancestry versus regions that do not. For example, genes in regions with introgressed Neanderthal DNA have been associated with immune function, skin and hair, metabolism, and ultraviolet radiation sensitivity (Vernot et al. 2016; McCoy et al. 2017; Chen et al. 2020). Some of these enrichments may reflect benefits of Neanderthal ancestry in regions encoding traits relevant for adaptation to Eurasian environments and pathogens (Racimo et al. 2015). However, further analyses are required to interpret the causes of phenotypic enrichment or depletion.

To test the robustness of annotation enrichment results across different introgression maps, we performed gene set enrichment analyses for each map and compared the results. Due to ArchIE's high discordance with other maps and ARGweaver-D's small number of individuals used, we excluded ArchIE and ARGweaver-D from phenotype enrichment analysis, resulting in 12 total maps included in these analyses. Results including ArchIE and ARGweaver-D are reported in Table S2.

We used rGREAT, a tool that performs gene annotation enrichment analysis by assigning input genomic regions to genes based on their proximity. rGREAT accounts for different probabilities of regions being assigned to a gene based on each region's length and distribution across the genome. For each introgression map, we computed enrichment compared to the whole genome background for phenotypes from the Human Phenotype Ontology (HPO) (Files S1 and S2).

The number of significantly enriched phenotypes for each introgression map varies substantially—from 0 to 734 (Fig. 6c). Overall, we discovered 1,427 significant enrichments (spanning 1,183 unique phenotypes) between a phenotype and an introgression map. However, 993 of the enriched phenotypes are observed for only one introgression map (Fig. 6a). Several maps—admixfrog, CRF16 (1) and (2), hmmix18, hmmix20, S*, and Sprime—do not have any HPO phenotype enrichments.

Figure 6.

Different phenotypes are enriched in different introgression maps. a) Number of Human Phenotype Ontology (HPO) phenotypes significantly enriched in different numbers of introgression maps. Bars in blue reflect phenotypes supported by 3+ maps, as shown in Fig. 7. Gene set annotation enrichment for each introgression map was performed using rGREAT with gene annotations from the HPO. The significantly associated phenotypes for each method and highly supported regions are given in Files S1 to S4. b) Upset plot of the number of HPO phenotypes enriched in different combinations of introgression maps. For example, 615 phenotypes are enriched in both highly supported regions present in 11+ maps (excluding ArchIE and ARGweaver-D) and regions identified by SARGE. c) Number of HPO phenotypes associated with each introgression map from b, along with highly supported regions.

We identified 41 phenotypes enriched in at least three distinct maps (Fig. 7). Many of these shared phenotypes influence a few specific bodily systems, including the eyes, skull, limbs, muscle, and reproductive tract. Patterns of similarity among the enrichments between maps differed from their overall genomic similarity (Figure S23).

Figure 7.

Human Phenotype Ontology (HPO) phenotypes significantly supported by at least three methods’ introgression maps. HPO phenotypes are ordered by the number of introgression maps supporting enrichment and then by phenotypes with the lowest P-value across tests. Three phenotypes highlighted in bold are also enriched in highly supported introgressed regions.

To explore whether highly supported introgressed loci, or regions supported by 11+ maps, had similar enrichments to those found across multiple methods, we performed rGREAT phenotype enrichment analysis on sets of highly supported regions (Table S1). These regions are associated with a higher number of phenotypes (up to 1,082) compared to each introgression map individually. Moreover, only three of the 41 phenotypes associated with 3+ maps are present in the set of phenotypes associated with highly supported regions, suggesting that the highly supported regions capture a set of loci with differing phenotypic relationships.

To explore the robustness of these observations to other phenotype enrichment methods, we also carried out phenotype enrichment analyses using a stricter criterion for mapping introgressed regions to genes, requiring that the introgressed region overlaps a gene's exon. Based on these gene sets identified by Annotatr, we used EnrichR to perform gene set enrichment analyses on phenotypes from the GWAS Catalog 2023 and HPO. This stricter mapping resulted in fewer phenotypic associations, but the heterogeneity of the associations and a greater number of enrichments for the highly supported introgressed regions remained (Table S3; Figures S25 and S26; Files S3 and S4).

Desert regions vary in levels of introgression across maps

Long genomic regions with very low levels of Neanderthal introgression have been identified in several previous introgression maps (Fig. 8a). However, the criteria used to call deserts varied substantially (Table S4), complicating comparisons and interpretation. To evaluate the support for five sets of deserts called in previous studies of introgression maps, we intersected all maps with these desert regions to quantify levels of introgression across maps in the deserts (Fig. 8b).

Figure 8.

Overlap of introgression deserts across maps. a) Locations of Neanderthal introgression deserts identified by five previous methods across the genome. b) Levels of introgression called in each introgression desert set. For each of the five sets, we computed the amount of introgression called in each desert by other introgression maps. The fraction of base pairs in each desert called as introgressed by a map was divided by the total fraction of the genome called as introgressed by that map (x axis). Thus, each box-and-whisker plot represents the relative depletion of introgression in each desert compared to the genome-wide proportion for each map. Desert windows in the X chromosome from hmmix20 and CRF16 are excluded in b due to most maps only reporting introgression in autosomes.

Given the substantial differences in criteria used to define deserts in previous studies (Table S4), we did not expect complete agreement between study-specific deserts. Nonetheless, eight distinct genomic regions on chromosomes 3, 7, and 8 harbor deserts identified across all five maps (Fig. 8a). Notably, this includes a region on chromosome three that overlaps ROBO1, which is involved in neurodevelopment and language abilities (Hannula-Jouppi et al. 2005), and was previously highlighted in Neanderthal desert analysis (Chen et al. 2020). An additional 36 desert regions are supported by at least three maps; these include regions overlapping FOXP2 and ROBO2, genes previously proposed to be involved in human-specific traits such as language (Sankararaman et al. 2016; Vernot et al. 2016; Chen et al. 2020). However, genes in these 36 loci are not significantly enriched for any HPO phenotypes using rGREAT.

To evaluate introgression patterns in the annotated deserts across all maps, we compared introgression levels called in each map in desert regions (Fig. 8b). For each desert region, we computed the proportion of introgressed bases called by other maps. To account for differences in the overall amount of introgression called by each method, we then normalized each proportion by each map's genome-wide introgression proportion. Introgression levels were relatively low for many of the deserts, but there were substantial differences between maps. Introgression levels in the hmmix20 deserts deviated most from other desert sets, likely due to inferring deserts based on a map of introgression in a large cohort of only modern Icelandic individuals. Additionally, we identified Neanderthal-free, 50 kb windows across the genome and found that these introgression-free windows mirror the similarity observed between the introgression maps (Figure S27). These results suggest several deserts supported by most methods but also highlight substantial variation in what is called an introgression desert, dependent on both the method and the set of individuals analyzed.

Discussion

Neanderthal introgression inference methods vary in their use of Neanderthal and outgroup genomes, statistical approaches, and target genomes. Here, we compare published introgression maps from a representative set of Neanderthal introgression inference methods applied to different modern human genomes. We find that support for specific predicted introgressed regions is bimodal, with about half of the regions highly supported by most maps and the remainder supported by only one or a few maps. There is substantial variation in the introgression predictions of different methods, even when applied to the same individual. Focusing on the effects of this variation on downstream analyses, we find that some evolutionary patterns are conserved across all maps, such as evidence for weaker background selection in regions with introgression. However, functional annotations of genes in introgressed regions and the identification of introgression deserts depend on the introgression map used.

We hope that the comparison and synthesis of Neanderthal introgression maps provided here will enable future studies to easily evaluate the robustness of their conclusions to the introgression map used and support the development of more accurate methods for calling introgression. To this end, we suggest that studies focused on a specific introgressed variant or region ensure that it is supported by maps based on different input data, algorithms, and assumptions. If a region is only supported by one method, this does not necessarily mean that it is not introgressed, but it suggests that further locus-specific modeling is required to resolve its history. For analyses of genome-wide patterns of introgression, we suggest evaluating whether results are qualitatively similar across multiple maps and the core set of highly supported regions. At the individual level, we provide introgression maps for individuals that have been used in many previous studies of introgression. Researchers developing new algorithms for detecting Neanderthal introgression can use these individual-level introgression maps as an additional source of comparison to existing methods. We encourage the further development of algorithms to infer archaic hominin ancestry in individuals from Africa. Finally, we strongly encourage the developers of new methods to provide open-source, well-documented implementations; the challenge of running existing methods is a substantial roadblock to comprehensive evaluation (Huang et al. 2025).

Our comparison of introgression desert regions called by different introgression maps found that previously highlighted genes, including ROBO1, ROBO2, and FOXP2, were included in most desert sets, with the ROBO1 gene supported by all five desert sets. The consistency of the desert regions overlapping these genes supports their potential role in modern human-specific biology, and further investigation into these desert regions may reveal the underpinnings of modern human evolution after Neanderthal introgression. However, we also observed that introgression levels predicted in many previously called desert regions vary greatly by method.

Our analyses have several limitations that should be kept in mind when interpreting our conclusions and using the synthesized introgression maps. First, there is no “ground truth” available to evaluate introgression predictions. We do not know the exact history of interactions between Neanderthals and modern humans and how these events influenced our genomes. Moreover, existing maps are based on genomes from modern human populations with diverse demographic histories that are challenging to infer and integrate into these analyses. Given this challenge, previous analyses have relied on different metrics and thresholds to identify Neanderthal introgression, adding another source of heterogeneity when comparing introgression maps. Some previous studies, such as ArchIE, have used simulations to create genomes in which the true history of all loci is known; however, such approaches are heavily dependent on the assumptions of the simulations. Given the lack of ground truth, we focus on quantifying differences and similarities between introgression maps and are unable to definitively determine whether observed discordance reflects the inaccuracy of a method. Nonetheless, the overall disagreement of ArchIE's predictions with other methods suggests miscalibration of the model or other technical issues (Huang et al. 2025).

Second, the set of introgressed regions detected depends on the number and genetic diversity of genomes studied. Given the challenges of running some of the methods, we were unable to generate predictions for all methods on the same set of individuals. Specifically, these challenges include: lack of available implementations (CRF), potential technical issues (ArchIE), and limited access to human data previously used to generate introgression maps (hmmix20). Considering these challenges, we performed several different comparisons of maps: (i) on full published integrated maps from all methods, (ii) on introgression maps filtered to contain the same number of highest-confidence predictions, and (iii) on individual genomes for a reduced set of methods. The map-wide and filtered comparisons should be interpreted carefully due to differences in the number and diversity of target genomes used by each method. Overall, while the results of each of these analyses differed, they consistently showed significantly more agreement between maps than expected. Nonetheless, given the substantial differences between maps, even when generated for the same individual, the choice of introgression map can impact both the identification of Neanderthal ancestry at a locus of interest and downstream conclusions.

Third, for downstream analyses, multiple methods and ontologies exist to perform phenotype enrichment on a set of genomic loci, which may yield different enrichment strengths and phenotypes implicated. As an illustrative test case, we quantified enrichment in regions with Neanderthal introgression in modern humans using HPO, which consists of gene annotations inferred primarily from human monogenic disease phenotypes. The enrichments we observe reflect many of the previously reported enrichments for phenotypes relevant to known differences between Neanderthals and humans, such as optical (Sankararaman et al. 2014), reproductive (Weaver and Hublin 2009), and skeletal (Colbran et al. 2019; McArthur et al. 2021; Wei et al. 2023) morphology. While these are present in introgressed regions supported by multiple methods, they are far from consistent across methods. We also note that other annotation sets and enrichment testing methods could yield different phenotype enrichment patterns.

Finally, all our analyses were carried out in the context of the hg19 genome assembly. This was necessary given the focus of previous Neanderthal variant calls and introgression analyses on this assembly. To facilitate future analyses, we provide our integrated prediction sets lifted over to both the hg38 and the CHM13 telomere-to-telomere (T2T) assembly. We anticipate that our findings will hold for these newer assemblies; however, additional introgressed loci would likely be found if the full detection pipelines were applied to these genome assemblies. Recent studies have begun to focus on the complete human reference genome (Liang et al. 2025), T2T-CHM13, and we encourage future work on Neanderthal introgression to focus on this assembly.

To build on our conclusions, future studies should consider various factors that drive differences between identified introgressed loci between methods. These factors include method-specific sensitivity to demographic models, amount of introgression inferred, types of selection, bias toward Neanderthal reference genomes, and the number and diversity of target samples used. For each of these factors, future work should use simulated genome data, where the ground truth is known, and alter one factor at a time. This framework will investigate how each of these factors contributes to variation in detected introgressed loci and how introgression detection methods may be biased toward certain conditions. Moving forward, we also encourage method developers to make all software and data publicly accessible and the greater community to make a coordinated effort to benchmark new methods.

Given our limited understanding of the legacy of Neanderthals on humans today and the continued discovery of archaic genomes, we anticipate further development of algorithms to identify introgressed DNA and new applications for these methods. We hope that our results will guide these studies and help ensure robust conclusions. Additionally, introgressed DNA identified between methods from other ancient hominin species, such as Denisovans, has yet to be formally evaluated. By quantifying differences in introgression inference methods from other archaic groups, we will gain a greater understanding of how humans evolved and how our ancestors shaped our history.

Methods

Defining and harmonizing introgression maps

For each introgression inference method, we obtained the predicted introgressed map from the original study. Each introgression map was then processed in a study-specific manner to generate standardized human genome build hg19-referenced bed files consisting of introgressed regions and an associated confidence score. For each method, introgressed fragments were merged across population groupings using BEDtools (Quinlan and Hall 2010) v2.31.1 merge. The highest scores across merged regions were retained. To generate method-specific introgression maps, all detected introgressed loci across populations were merged. All introgression maps were built in hg19 due to its overwhelming use in ancient DNA studies.

admixfrog (Iasi et al. 2024): Introgression maps for admixfrog were downloaded from https://doi.org/10.5061/dryad.zw3r228gg. ALL_called_ancestry_segments_updated.csv was used to identify Neanderthal introgressed fragments for 1KG and SGDP individuals. Only present-day individuals were included in the final introgression map, as listed in Meta_Data_individuals.csv, as well as fragments labeled as the “state” type with a Neanderthal target. All fragments were at least 0.05 cM (∼50 kb) long, and hg19 positions were obtained from the “pos” and “pos_end” columns. Scores for each fragment reflect the certainty of an introgressed segment, normalized by its bin size (“nscore”).

ARGweaver-D (Hubisz et al. 2020): Introgressed haplotypes from ARGweaver-D were obtained from http://compgen.cshl.edu/ARGweaver/introgressionHub/. Introgressed regions from the ooaM1A directory were used, reflecting a model with migration bands from Neanderthals and Denisovans into four modern humans from SGDP: two Africans (Mandenka and Khomani San), one Papuan, and one Basque individual. Files used to generate a combined file were ooaM1A/Papuan_1F.bed, ooaM1A/Basque_2F.bed, ooaM1A/Mandenka_2F.bed, and ooaM1A/Khomani_San_1F.bed. All loci labeled as neaTo(Human) were included, and scores represent heterozygous (500) vs homozygous (1,000) introgressed regions.

ArchIE (Durvasula et al. 2019): Neanderthal introgressed haplotypes in CEU 1KG individuals were provided by Dr. Arun Durvasula, by request. Haplotypes with a probability of Neanderthal introgression greater than 0.9984 were subsetted, reflecting ∼2% Neanderthal introgression per individual. To avoid overlapping 1 bp windows, we subtracted 1 from each end coordinate.

ArchaicSeeker2 (Yuan et al. 2021): Introgressed fragments for 1KG and SGDP Papuan individuals were obtained from .seg files in the IntrogressedSeg directory at https://github.com/Shuhua-Group/ArchaicSeeker2.0. Segments with “Neanderthal” for the BestMatchedPop columns were retained to generate introgression maps.

DICAL-ADMIX (Steinrücken et al. 2018): DICAL-ADMIX output for CEU, CHB, and CHS 1KG individuals was downloaded from https://dical-admix.sourceforge.net. All fragments provided have posterior probabilities of Neanderthal introgression greater than 0.42, as recommended by the original authors, and were used to generate introgression maps.

CRF14 (Sankararaman et al. 2014): Neanderthal introgressed regions and SNPs were downloaded from https://drive.google.com/drive/folders/1jRymYI3GyNrBXKtd6pT9pqoikZ6I8P3Z. Each entry is associated with scores reflecting the probability of introgression from the CRF. All haplotypes have a probability greater than 0.9, as provided in the downloaded data. Population labels are based on 1KG categorizations.

CRF16 (1) and CRF16 (2) (Sankararaman et al. 2016): Introgression maps were downloaded from https://drive.google.com/drive/folders/1DyhMw0E1mXQUDNeGQbQrcQbHe0uTyK2h. Introgressed regions were processed separately for two directories available: The first directory only uses African individuals as a non-Neanderthal introgressed outgroup, whereas the second directory uses both African and the Denisovan genome as a non-Neanderthal introgressed outgroup, which is the focus of the 2016 paper. We refer to these two methods, respectively, as CRF16 (1) and CRF16 (2). For each directory, the average predicted probability for each haplotype was obtained to generate introgression maps. All CRF haplotypes have introgression probabilities greater than 0.9, and variants have Neanderthal ancestry probabilities greater than 0.5 for the derived allele. Additionally, variant information was generated by separating the SNP ID into chromosome and position. Finally, the “start” column was added by subtracting one from the variant position column.

hmmix18 (Skov et al. 2018): Introgression maps for HGDP (Cavalli-Sforza 2005) and 1KG individuals were downloaded from https://zenodo.org/records/14136628 and lifted to hg19 using LiftOver. To subset for Neanderthal introgressed fragments, we selected fragments with ND_type labeled “Neanderthal” and kept fragments with a mean probability greater than 0.8.

hmmix20 (Skov et al. 2020): Dataset S1 from the original publication (Skov et al. 2020) was downloaded, containing all introgressed fragments with an introgression probability greater than 0.9. Using LiftOver (Perez et al. 2025), we converted the genomic coordinates to hg19 from hg38 to allow for direct comparison with other methods. To focus on Neanderthal introgressed regions, fragments with an archaic label of Altai or Vindija were subsetted.

hmmix23 (X chromosome only) (Skov et al. 2023): Introgression predictions in the X chromosome were provided by Dr. Laurits Skov, by request. Segments with MeanProb greater than 0.8 and with a greater number of SNPs in the “Altai” and “Vindija” compared to “Denisova” columns were used. This map was only used in the X chromosome analysis reported briefly in the X chromosome-specific comparison (Figures S7 and S8).

IBDMix (Li et al. 2024): Introgressed segments were downloaded from https://github.com/PrincetonUniversity/IBDmix/blob/main/IBDmix_calls_using_3_archaics.tar.gz, as recommended by the authors and used in Li et al. (2024). IBDMix uses LOD scores, reflecting the likelihood that a region is shared with the reference Neanderthal through identity by descent. Here, all introgressed loci have LOD scores greater than 3 and fragment lengths greater than 50 kb, as provided by the original authors. Population labels are based on 1KG. Desert regions used for IBDMix in this study reflect calls from an earlier implementation of IBDMix on the same individuals (Chen et al. 2020).

SARGE (Schaefer et al. 2021): Data were provided by Dr. Nathan Schaefer, by request. Data provided were already in the hs37d5 genome build. Scores reflect similarity scores between haplotypes, depending on how many ancestral recombination events are shared, and all data provided are filtered according to the original implementation (Schaefer et al. 2021).

Sprime (Browning et al. 2018): Sprime output for 1KG (non-African) and SGDP Papuan individuals was downloaded from https://doi.org/10.17632/y7hyt83vxr.1, consisting of variant-level and fragment-level files. For variant files, a “start” column was added by subtracting one from the variant position column. Fragment files were generated by taking the first and last variant positions for each Segment_ID. To call putative Neanderthal introgressed loci, we first identified Sprime outputs where NMATCH contained “match.” Next, we calculated match rates for each segment with at least 30 putative Neanderthal and Denisovan variants and divided the number of variants with NMATCH containing “match” by the total number of putative Neanderthal and Denisovan variants. Finally, to obtain high-confidence Neanderthal introgressed regions, we kept segments with match rates higher than 0.6 to the Neanderthal and match rates less than 0.4 to the Denisovan genome, as described in the original publication (Browning et al. 2018). Sprime scores reflect the confidence of introgression associated with each locus, and all Sprime loci met a minimum threshold value of 150,000.

S* (Vernot et al. 2016): S* Neanderthal haplotypes and variants were downloaded from https://drive.google.com/drive/folders/0B9Pc7_zItMCVWUp6bWtXc2xJVkk?resourcekey=0-Cj8G4QYndXQLVIGPoWKUjQ. Neanderthal-specific population and chromosome-specific merged files in the introgressed_haplotypes folder were used to generate Neanderthal introgression maps. Scores reflect the posterior probability of Neanderthal introgression for each haplotype, labeled post.p.alt1 in the provided data. Variants were obtained from introgressed_tag_snp_frequencies.

Jaccard similarity and hierarchical clustering of introgression maps

Jaccard similarity was computed by dividing the number of shared base pairs by the union of all base pairs between pairs of maps $u,v$ with NumPy v1.26.3. Normalized Jaccard similarity was computed by using the size of the smaller map as the denominator, rather than the union between the two sets. Additionally, for each Jaccard similarity comparison, we generated a null distribution by shuffling each pair of maps 100 times across the hg19 genome, with gaps removed, and computing Jaccard similarities for each shuffled pair of maps. This null distribution quantifies the expected overlap between a pair of maps if their regions were randomly distributed. We compute a z-score for each observed Jaccard similarity, compared to the null distribution:

$$z\text{-}\mathrm{score}=\frac{\mathrm{observed}\mathrm{Jaccard}-\mathrm{mean}\mathrm{Jaccard}\mathrm{from}\mathrm{shuffled}\mathrm{distribution}}{\mathrm{standard}\mathrm{deviation}\mathrm{of}\mathrm{Jaccards}\mathrm{from}\mathrm{shuffled}\mathrm{distribution}}$$

We also compute empirical P-values for each observed Jaccard similarity by counting the number of Jaccard similarities in the null distribution that are at least as extreme as the observed value + 1, divided by the number of shuffles (100) + 1.

$$\mathrm{empirical}P\text{-}\mathrm{value}=\frac{\mathrm{sum}(\mathrm{Jaccards}\mathrm{from}\mathrm{shuffled}\mathrm{distribution}\ge \mathrm{observed}\mathrm{Jaccard})+1}{100\mathrm{shuffles}+1}$$

Hierarchical clustering was performed using the linkage function in SciPy v1.14.1 on the Jaccard distance matrix created by subtracting each Jaccard similarity value from 1. Clustering was performed using the Farthest Point Algorithm, where $d(u,v)=\max (\mathrm{dist}(u[i],v[j])$.

To obtain the top introgressed loci for each method when subsetting the maps, introgressed regions for each introgression map were sorted by the highest associated scores.

Results were visualized using upsetplot v0.9.0, seaborn v0.13.2, and matplotlib v3.9.2 in Python v3.11.5.

B statistic scores in introgressed versus nonintrogressed regions

B statistic scores originally described in McVicker et al. (2009) were obtained from Telis et al. (2020), containing B statistic values ranging from 1 to 1,000. B statistic values were quantized into multiples of 50 and lifted to hg19 by Telis et al. (2020). To compute B statistic values across introgression maps, we assigned each 500 bp window across the hg19 genome as either introgressed or nonintrogressed. A window was introgressed if it contained at least one introgressed region, as we hoped to observe whether windows with tolerance for some introgression had different B statistic scores compared to nonintrogressed windows.

Next, we performed bootstrapping with 1,000 iterations to obtain a bootstrapped mean and confidence intervals. We visualized results using matplotlib v3.9.2 and computed differences between B statistic values between introgressed and nonintrogressed windows using a Mann–Whitney U test.

Phenotype annotation enrichment in introgression maps

Gene set enrichment analysis for each introgression map was performed using rGREAT v2.6.0 (Gu and Hübschmann 2023) with the default hg19 genome (gaps removed) as the enrichment background and gene sets in the Human Phenotype Ontology, which we selected to associate loci with interpretable human features. Each gene in the ontology was extended to capture short and long-range transcription start site (TSS) associations using default parameters (basal domain and extension around the TSS, 5 kb upstream and 1 kb downstream, then extension of the basal domain in both directions to 1 Mb, or until the neighbor gene's basal domain is reached to capture long-range associations). Next, rGREAT computes the fraction of genome overlap between the set of input regions and gene-associated extended regions for each phenotype's gene set, then computes an enrichment P-value under a binomial model.

Additionally, we used EnrichR v3.4 (Chen et al. 2013) as a complementary gene set enrichment method for further comparison. To obtain an input gene set consisting of genes that overlapped with exonic regions for a set of input loci, we used Annotatr v.1.30.0 (Cavalcante and Sartor 2017). We annotated our input bed file with hg19 genes where an input region falls in an exon (hg19_genes_exons), and annotations were built by Annotatr's build_annotations function using data from the TxDb.Hsapiens.UCSC.hg19 database. The exon-specific gene list from Annotatr was used as an input for EnrichR, and gene set libraries were limited to GWAS Catalog 2023 and HPO.

Principal components analysis (PCA) was performed in R v4.4.3 on adjusted P-values from EnrichR and rGREAT with the prcomp function from stats v4.4.3. Adjusted P-values were standardized, and phenotypes with constant or “NA” values were excluded, resulting in 386 phenotypes included in the PCA.

Visualizations were generated using ggplot2 v3.5.1 and UpSetR v1.4.0.

Comparing Neanderthal introgression deserts

Neanderthal introgression deserts were obtained for each respective study. IBDMix and S* deserts were manually entered into bed format from Table S8 (Chen et al. 2020). ArchaicSeeker2 deserts were obtained from Table S6 (Yuan et al. 2021), and hmmix20 deserts were downloaded from Dataset S4 (Skov et al. 2020) and lifted to hg19 from hg38. Sankararaman16 deserts were provided by Dr. Sriram Sankararaman. No additional modifications to these data were performed.

To observe the amount of introgression for each desert window, we used BEDtools’ coverage function to compute the fraction of introgressed loci for each method's introgression map overlapping each desert window.

Desert regions across the genome for each method were plotted using karyoploteR (Gel and Serra 2017).

Author contributions

Yaen Chen (Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing—original draft, Writing—review & editing, Visualization), Keila Velazquez-Arcelay (Data curation, Writing—review & editing), and John A. Capra (Conceptualization, Methodology, Investigation, Writing—original draft, Writing—review & editing, Supervision, Funding acquisition)

Supplementary material

Supplementary material is available at Molecular Biology and Evolution online.

Funding

This work was supported by the National Institutes of Health (NIH) General Medical Sciences award R35GM127087 to J.A.C. K.V.-A. was supported by NIH award T32CA108462.

Data availability

The publicly available data used for analysis are available in the following repositories: https://github.com/yaenchen/NeanderthalIntrogressionMaps.