Large-scale genome sequencing of giant pandas improves the understanding of population structure and future conservation initiatives

Significance

The giant panda is facing accelerating habitat fragmentation, with several relatively isolated subpopulations in six mountain ranges of the Sichuan, Shaanxi, and Gansu Provinces in China. However, whether recently fragmented habitats have resulted in population differentiation and what risks wild giant pandas are facing are largely unknown, and this knowledge hereof is important for future conservation. Here, we performed a large-scale population genomic study covering all current populations, revealing a genetic structure, and providing a picture of population dynamics in the context of evolutionary history. The investigation of the genetic background and the evaluation of the risks of inbreeding and the accumulation of deleterious alleles provide crucial knowledge of importance for the future conservation of the giant panda.

Abstract

The extinction risk of the giant panda has been demoted from “endangered” to “vulnerable” on the International Union for Conservation of Nature Red List, but its habitat is more fragmented than ever before, resulting in 33 isolated giant panda populations according to the fourth national survey released by the Chinese government. Further comprehensive investigations of the genetic background and in-depth assessments of the conservation status of wild populations are still necessary and urgently needed. Here, we sequenced the genomes of 612 giant pandas with an average depth of ~26× and generated a high-resolution map of genomic variation with more than 20 million variants covering wild individuals from six mountain ranges and captive representatives in China. We identified distinct genetic clusters within the Minshan population by performing a fine-grained genetic structure. The estimation of inbreeding and genetic load associated with historical population dynamics suggested that future conservation efforts should pay special attention to the Qinling and Liangshan populations. Releasing captive individuals with a genetic background similar to the recipient population appears to be an advantageous genetic rescue strategy for recovering the wild giant panda populations, as this approach introduces fewer deleterious mutations into the wild population than mating with differentiated lineages. These findings emphasize the superiority of large-scale population genomics to provide precise guidelines for future conservation of the giant panda.

A cornerstone strategy for conserving small populations of a species, especially endangered species, is to build protected areas (PAs) (1). Delineating the fine-scale genetic structure of a species will greatly facilitate the rational planning of PAs and the conservation of genetic diversity (2, 3). If a genetic structure represents a remnant of ancestral standing variations, it is important to consider outbreeding risks when performing a translocation from one area to another, because adaptations formed over a long evolutionary history in response to local environments may differ between populations (3–5). In contrast, if the genetic structure of a certain species is the consequence of isolation due to fragmented habitats, gene flow between populations may better reduce the loss of genetic diversity (3, 6).

Genetic rescue has been considered a powerful means to increase fitness by preventing inbreeding and facilitating gene flow to increase genetic diversity (7, 8). Accordingly, understanding the basic genetic background and extinction risk of each population for a given species is key to formulating evidence-based genetic rescue plans, including information on genome-wide genetic diversity, genetic differentiation, historical and current gene flow, inbreeding, and potentially deleterious mutations across the genome (6, 9–11).

The giant panda (Ailuropoda melanoleuca) is a global flagship species for biodiversity conservation (12) and has been split into two subspecies: Ailuropoda melanoleuca qinlingensis, restricted to Qinling (QLI), and Ailuropoda melanoleuca melanoleuca, which lives in the Minshan (MSH), Qionglai (QLA), Daxiangling (DXL), Xiaoxiangling (XXL), and Liangshan (LSH) mountains. Challenges from both ecological and biological factors have pushed the giant panda to a state of endangerment (13–15). The Chinese government has passed a series of laws (Forest Law, 1984; Criminal Law, 1987; Law on the Protection of Wildlife 1988; Environmental Protection Law, 1989) in recent decades to improve the protection of the wild giant panda population. The number of wild giant pandas increased from 1,596 (1999 to 2003) to 1,864 (2011 to 2014) according to the report of the fourth National Giant Panda Survey released in 2014 (16), and the giant panda was downlisted from “Endangered” to “Vulnerable” by the International Union for Conservation of Nature (IUCN) in 2016 (17).

Current threats from human activity are, however, much greater now than they were decades ago, largely due to the development of infrastructure, agriculture, and tourism in the original habitats of the giant pandas (13, 14). As summarized in the fourth National Giant Panda Survey (18), the wild giant panda population is divided into 33 isolated subpopulations, 18 of which comprise fewer than 10 giant panda individuals (14). Obviously, some wild populations are still facing risks of local extinction, although the giant panda is not doomed to extinction in the near future (19). Building ecological corridors between isolated populations and rewilding captive-born pandas that represent wild lineages to recover the local wild populations are both conservation strategies that have been implemented and achieved initial success in the past two decades. In 2021, the Chinese government announced the opening of the Giant Panda National Park, which spans the Sichuan, Shaanxi, and Gansu Provinces (20) and further extended the protection of giant pandas. In addition, molecular evidence and genomic analysis have revealed a Qinling subspecies (21) and a preliminary population structure of Sichuan giant pandas (22). However, the genomic background of both wild and captive panda populations remains largely unexplored, and such information is essential to providing an in-depth understanding of their genetic status to aid in future conservation.

Recently, comprehensive genomic background investigations of endangered species have been reported to assist in protection and conservation (11, 23–30). Similarly, a systematic population genomic investigation is particularly important for determining the detailed population structure, evolutionary history, and extinction risks of the giant panda, all of which are critical for assisting short-term and long-term conservation plans. In this study, by carrying out a large-scale population genomics study involving 649 giant pandas (SI Appendix, Fig. S1) covering six mountain range populations and captive individuals, we comprehensively investigated and compared the genetic backgrounds of wild and captive giant pandas, which we envision will aid in evidence-based conservation of the giant panda in the future.

Results

Samples and Genome Sequencing.

We performed whole genome resequencing of 612 giant pandas, comprising 538 wild individuals and 74 captive individuals (21 wild-born and 53 captive-born), with an average sequencing depth and coverage of 26.08 ± 0.34-fold (lowest: 5.60-fold; 604 samples: ≥10-fold) and 98.10 ± 0.03%, respectively (Table 1, SI Appendix, Fig. S1, and Dataset S1). Considering that samples of old skin from wild pandas lack detailed identification information, we first performed species identification and individualization to facilitate subsequent genomic analysis. We found that all samples were from the giant panda, but 39 of them were closely related (twins, siblings, or parent–children) (SI Appendix, Table S1), and we randomly selected one from each pair, leaving 517 unrelated wild individuals, including 150 samples with known sampling sites (MSH: n = 99; QLA: n = 37; LSH: n = 11; QLI: n = 2; XXL: n = 1). Fortunately, we did not observe any deamination-introduced C-to-T changes at the ends of the reads in these historical skin samples (SI Appendix, Fig. S2).

Table 1.

Summary of genome sequencing data of the 591 giant pandas included in this study

	Known wild individuals					Unknown wild individuals	Captive-born individuals	Total
	MSH	QLA	LSH	XXL	QLI
Sample size	102	53	12	1	3	367	53	591
Q30	91.3 ± 0.2	91.5 ± 0.2	90.2 ± 0.7	86.9	90.2 ± 1.1	91.5 ± 0.1	91.2 ± 0.3	91.4 ± 0.1
Bases (Gb)	73.1 ± 2.2	65.4 ± 2.8	74.9 ± 10.7	51.8	73.2 ± 18.0	75.7 ± 1.4	54.2 ± 1.9	72.3 ± 1.1
Depth (X)	26.4 ± 0.8	24.8 ± 1.1	26.9 ± 4.1	19.9	14.8 ± 4.7	27.0 ± 0.5	22.0 ± 0.8	26.2 ± 0.4
Coverage (%)	98.1 ± 0.1	98.2 ± 0.1	98.2 ± 0.1	98.6	96.2 ± 1.1	98.1 ± 0.1	98.4 ± 0.1	98.1 ± 0.1

Fine-Scale Population Genetic Structure.

Including previously published whole genome sequencing (WGS) data (n = 58, SI Appendix, Table S2), we explored the fine-scale population genetic structure of the giant panda based on 229 samples with known geographical origins (Fig. 1). We identified 17,069,700 population-level variants, including 13,535,492 single nucleotide polymorphisms (SNPs) and 3,534,208 short insertions or deletions (InDels), and summarized the characteristics of these variants (SI Appendix, Fig. S3 and Table S3). We identified four distinct genetic clusters by principal component analysis (PCA) based on genome-wide SNPs (Fig. 2A). Individuals from the MSH, LSH, and QLI formed distinct clusters without any overlapping samples. Unexpectedly, we were unable to identify clear borders between individuals from QLA, DXL, and XXL, indicating that these three neighboring populations might have formed a single metapopulation (QLA-DXL-XXL, hereafter QX) (SI Appendix, Fig S4). This also reflects the ability of larger sample sizes to explore genetic structures. The maximum likelihood (ML) phylogenetic tree (Fig. 2B) and admixture analysis (Fig. 2C and SI Appendix, Fig. S5) were consistent with the PCA clustering results, with four distinct genetic groups: MSH, LSH, QLI, and QX. To eliminate the clustering bias from different sample sizes, we further performed population structure analysis by randomly selecting the same sample size with varying gradients from each population to support our conclusion (SI Appendix, Fig. S6).

Fig. 1.

Sampling information for the 649 giant panda individuals included in this study. This map covers all current giant panda habitats in six mountain ranges (red: Qinling; green: Minshan; purple: Qionglai; pink: Daxiangling; blue: Xiaoxiangling; orange: Liangshan). The numbers in triangles and circles represent the number of published and newly sequenced individuals, respectively. The sampling sites (breeding centers) for captive individuals are pointed out by yellow six-pointed stars, with four individuals from the Qinling Giant Panda Research Center, and 49 individuals from China Conservation and Research Center for the Giant Panda (Ya'an base: n = 11, Dujiangyan base: n = 18; Wolong base: n = 20). Large rivers are marked: a: Baishui River; b: Duobu River; c: Fu River; d: Min River; e: Dadu River; f: Niri River.

Fig. 2.

Population structure of wild giant pandas with known sampling locations. (A) PCA plot for the first (PC1) and second (PC2) component revealing four distinct clusters of giant panda individuals from the MSH, LSH, QX, QLI populations. (B) ML phylogenetic tree constructed by IQ-TREE. (C) Inferred admixture proportions for K = 4 to 6 with colors representing ancestry components. (D) PCA plot showing the genetic structure of the three MSH subpopulations. (E) Phylogenetic relationships of the MSH subpopulations shown by the ML tree. F_ST values between different populations/subpopulations are indicated by black numbers in the PCA plots (A and D). The gray triangles and diamonds in (A–E) represent the four exceptional individuals in the MSH population.

Interestingly, upon closer inspection of the PCA clusters, we noticed that the MSH population was composed of three subgroups (MSH1, MSH2, MSH3) (Fig. 2A). Admixture analysis also supported that the MSH population split into three clusters when K = 6 (Fig. 2C), and the ML tree showed the corresponding three clades, but even more subclades could be identified (Fig. 2B), revealing complex structures in the MSH population. Focusing on samples only from MSH, we found nearly the same result with identification of three distinct subpopulations (Fig. 2 D and E). Interestingly, except for four samples, the three subpopulations matched with their geographical distribution, with the MSH1, MSH2, and MSH3 subpopulations matching to Minshan G, J, and K, respectively (31) (Figs. 1 and 2D).

Based on this genetic structure, we attempted to assign the remaining 367 individuals with unknown sampling information to current giant panda populations using both supervised and unsupervised clustering algorithms (SI Appendix, Fig. S7). Ultimately, we identified 256 individuals who most likely originated from the MSH and 16 individuals who most likely originated from the LSH. The remaining 95 individuals could not be assigned to a distinct population, because these individuals associated with QLA, DXL, and XXL, supporting the above-mentioned QX population (SI Appendix, Fig. S7). We reproduced the existence of three subpopulations in the MSH by adding these 367 samples to the analysis. This information on population structure is valuable for the future management of wild populations, but should be further corroborated by additional genomic data, particularly from XXL and DXL.

Historical Gene Flow and Admixture in Wild Populations.

Phylogenetic analysis supported the most ancient split of the QLI population from other Sichuan populations (SI Appendix, Fig. S8), but the historical gene flow and admixture events of these populations needed to be further elucidated. We first examined the evolutionary topology (((MSH, QX), LSH), QLI) by combining evidence from a phylogenetic tree, identity by descent (IBD) fragments, and TreeMix analysis (SI Appendix, Figs. S8–S10). Based on this topology, we disentangled the gene flow among populations with the polar bear as an outgroup (32).

The most intensive gene flow was found from the QLI to the MSH population, which was supported by D-statistics, TreeMix (m = 1 to 20), and mitochondrial haplotype sharing between the QLI and MSH populations (Fig. 3A and SI Appendix, Figs. S10–S12). In particular, gene flow between the QLI and the MSH1 population was more frequent than that between the QLI and the MSH2 or the MSH3 populations. Among the Sichuan populations, gene flow between the LSH and the QX population was more frequent than that between the LSH and the MSH population. Interestingly, the QX shared higher gene flow with the MSH3 population than that with the MSH1 and the MSH2 populations, consistent with their geographical distribution (Fig. 1). Although IBD analysis revealed frequent gene flow among giant panda populations since 20 to 30 thousand years ago (kya), gene flow between the QLI and the Sichuan populations, as well as between the LSH and MSH populations has become minimal to nonexistent in the most recent 1,000 y (SI Appendix, Fig. S9).

Fig. 3.

Historical gene flow, admixture graph, and demographic fluctuations in extant giant panda populations. (A) D-statistic under the model of (H1, H2, QLI, Polar bear) or (H1, H2, LSH, polar bear), where polar bear serves as an outgroup. Each group was calculated by 1,000 independent runs with 10 randomly selected individuals from each giant panda population. Significant differences with an absolute Z score above 3 are noted by gray points. (B) Admixture graph modeling with qpGraph illustrating the formation process of the current giant panda populations and admixture events contributing to the emergence of the MSH and the QX populations. (C) The inferred effective population size (N_e) history over the past 30,000 -1,000 y for the wild giant panda populations. (D) Recent demographic fluctuations over the past 2,000 y of each wild population inferred by PopSizeABC. The colored dotted lines indicate the 90% CI. These plots are scaled with a generation time of 12 y and mutation rate of 1.29 × 10⁻⁸ per site per generation.

The full landscape of the admixture scenarios inferred by qpGraph modeling was complex (Fig. 3B). An ancient lineage of the QLI population migrated to the ancestor of the MSH population. After the almost simultaneous split of the MSH population into three subclades, genetic components from the QX population entered the MSH3 population, and then contributed to the MSH2 population. Thus, the formation of the MSH and the QX populations involved ancient hybrid events with the QLI population and the LSH population, respectively, which provided an evolutionary basis for recovering the past gene flow.

Demographic History of Giant Pandas.

We inferred that the QLI population separated from the Sichuan populations from 12 kya to 7 kya, after which the earliest divergence in the Sichuan populations occurred in the LSH population (from 11 kya to 6.2 kya) (SI Appendix, Figs. S13 and S14). Separation between the MSH and the QX populations occurred from 6 kya to 3.5 kya, which is in line with previous estimates (15), and the most recent separation among the three MSH subpopulations occurred between 3 kya and 2 kya.

We then revealed postdivergence demographic trajectories of the Qinling and the Sichuan subspecies. The three Sichuan populations (MSH, LSH, and QX) underwent a similar demographic history characterized by a continuous decline before their separation from 30 kya to 10 kya, and then experienced different Ne changes during the past 10,000 y (Fig. 3C and SI Appendix, Fig. S15): 1) The MSH and the QX populations exhibited a slow decrease from 10 kya to 5 kya, after which both began to increase; 2) the LSH population experienced a continuous decrease from 10 kya to 3 kya after separation from the other Sichuan populations and then began to increase; and 3) the Ne of the MSH population remained higher than that of the other populations since the beginning of their separations (10 kya). For the three MSH subpopulations, the changes in Ne over time were very similar before their separation (Fig. 3C and SI Appendix, Fig. S15). Within the past 2,000 y, the Ne of the MSH, LSH, and QX populations were similar, with a shared declining trajectory (Fig. 3D and SI Appendix, Figs. S16–S18). More seriously, compared to that of the Sichuan populations, the QLI population decreased quickly from 30 kya to the present day without any population rebound (Fig. 3 C and D). However, the Ne of the QLI population was found to be the largest before 7 kya, then converged with that of the Sichuan populations and remained the lowest until now.

Inbreeding Profile of Giant Panda Populations.

Inbreeding in giant panda populations was evaluated by screening runs of homozygosity (ROH) across the genome. The average and the longest lengths of all ROH fragments were 316.23 Kb and 8.79 Mb, respectively, in wild pandas, and 332.52 Kb and 10.89 Mb, respectively, in captive-born captive (CBC) individuals (SI Appendix, Table S4). ROHs shorter than 1 Mb were dominant in both the wild and CBC giant pandas (wild: 95.98%; CBC: 94.85%), indicating that recent inbreeding was not severe. We detected 23.52 ± 3.79% and 18.30 ± 6.56% F_ROH in the wild and CBC giant pandas, respectively (Fig. 4A), indicating that CBC individuals are facing remarkably less inbreeding risk than wild pandas. We also found different distribution patterns of the length and number of ROHs on different chromosomes (SI Appendix, Fig. S19).

Fig. 4.

Characterization of inbreeding in giant panda populations. (A) Individual inbreeding coefficients inferred from the proportion of the genome within ROHs (F_ROH). The F_ROH for ROH ≥ 100 Kb is shown here. (B) Relationship between the F_ROH and genetic differentiation (F_ST) between parental linages of captive pandas. (C) Distribution of ROHs among different length classes in the wild and CBC populations. The ROH length is categorized by their expected generations.

We further investigated the inbreeding level in each wild and CBC population. We found that the QLI population harbored the most ROHs (1,849.20 ± 293.06) among these populations, with the longest average ROH length per individual (360.96 ± 80.78 Kb), which was longer than that of the other populations (SI Appendix, Fig. S20). As expected, the F_ROH in the QLI population ranged from 18.14 to 37.34% (average: 29.09 ± 4.66%), which was greater than those of the Sichuan populations (LSH: 26.65 ± 6.48%, QX: 23.15 ± 2.39% and MSH: 22.89 ± 3.34%). Similarly, the genetic diversity (π) of the QLI population was also the lowest among the wild panda populations (SI Appendix, Fig. S21). The ROH distributions in the three MSH subpopulations were very similar, with F_ROH values of 23.12 ± 3.17%, 22.19 ± 3.97% and 23.05 ± 2.32% in the MSH1, MSH2, and MSH3 populations, respectively (Fig. 4A).

For CBC individuals, it is worth noting that the inbreeding level of hybrid CBC individuals (hCBC, the ancestor of parents from different wild populations) was much lower than that of inbred CBC individuals (iCBC, the ancestor of parents from the same wild population) (SI Appendix, Fig. S22), showing the value of reducing inbreeding by artificial mating management. Additionally, the F_ROH of CBC was inversely related to the F_ST between the paternal and maternal lineages (Fig. 4B).

Inbreeding History.

To examine the inbreeding history, we dissected the ROH distribution by generation. A large proportion of ROH fragments were restricted to <1 Mb, indicating that the sharing of ancestral components mainly occurred at least 50 generations ago (QLI>LSH>MSH>QX) (Fig. 4C). Thereafter, the inbreeding level in giant panda populations seemed to be modest, even in the past 10 generations, because the occurrence of ROHs larger than 5 Mb was very limited in each population, suggesting a very modest negative effect from recent inbreeding. Again, the QLI population exhibited more intensive inbreeding than the Sichuan populations across its evolutionary history. An alarming trend was that inbreeding in the LSH and the QX populations increased and became greater than that observed in the MSH population during the past 50 generations. Within the MSH population, the inbreeding level in the MSH3 subpopulation was lower than that in the MSH1 and MSH2 subpopulations before 50 generations ago, and then the MSH3 subpopulation experienced an increase in inbreeding, which was ultimately greater than that of the other two MSH subpopulations. Interestingly, the observed F_ROH in CBC individuals before 30 generations, particularly before 50 generations, was much lower than that in wild individuals, and subsequently became comparable to that in wild populations.

Genetic Load in Giant Pandas.

We screened for missense mutations, loss-of-function (LoF) mutations, and deleterious nonsynonymous SNPs (dnsSNPs) across the entire genome to investigate putatively derived deleterious mutations. For the homozygous load, we detected an average of 8,014.78 ± 247.80 missense mutations, 533.26 ± 19.13 LoFs and 288.58 ± 14.89 dnsSNPs in the wild population. Not surprisingly, the QLI population harbored the most deleterious homozygous mutations, with a ratio significantly higher than that in the three Sichuan populations (Fig. 5A and SI Appendix, Fig. S23). Within the MSH population, the MSH2 subpopulation harbored fewer deleterious mutations than did the MSH1 and the MSH3 subpopulations. We further found that homozygous LoFs were positively correlated with F_ROH in all wild populations, and this correlation was stronger in the QLI and LSH populations (Fig. 5B). We found many high-frequency LoFs distributed in genes related to the immune system and reproduction, with the QLI population carrying the most LoFs in reproduction-related genes, and the LSH population carrying the most LoFs in immune-related genes (SI Appendix, Table S5).

Fig. 5.

Genetic load in giant panda populations. (A) The ratio of homozygous LoF alleles in each wild individual genome. Comparisons between populations are shown by the P values above. (B) Linear regression of homozygous LoF alleles against F_ROH for each wild population. (C) The distribution of LoF alleles in ROH regions and outside ROH regions, normalized by the number of synonymous alleles in the same region for each individual. Significance comparisons between ROH and non-ROH regions in each population: p_QLI = 5.98e-04, p_LSH = 3.50e-05, p_QX = 6.46e-16, p_MSH1 < 2.2e-16, p_MSH2 = 1.34e-14, p_MSH3 = 6.22e-04, p_CBC = 8.79e-15. (D and E) Prediction of the change in the genetic load accumulation in the offspring during the population recovery stage after reintroducing three donors to the QLI population under three scenarios, presented by heterozygous (D) and homozygous (E) damaging alleles (LoF and dnsSNP associated with HGMD).

There were fewer homozygous deleterious mutations in captive pandas than in wild pandas, with an average of 7,565.43 ± 359.10 missense, 500.68 ± 23.88 LoF, and 271.15 ± 15.21 dnsSNP mutations detected in this study. We also found a significant positive correlation between LoFs and F_ROH in the captive individuals (SI Appendix, Fig. S24). Interestingly, the correlation was much stronger in iCBC individuals than in hCBC individuals, which indicated that the genetic load could be further relieved by mating management in future generations.

To substantiate these findings, we performed GERP analysis to identify the most conserved sites across the genome (26), providing information on the potentially most deleterious mutations. The results of this analysis were consistent with those predicted by SnpEff (SI Appendix, Fig. S25). In addition, we blasted genes carrying putatively deleterious alleles to the Human Gene Mutation Database (HGMD) (9) to narrow down mutations that are likely associated with pathological disorders. The HGMD screening showed that the QLI population suffered from the highest detrimental genetic load, followed by the LSH population (SI Appendix, Fig. S26). Again, the overall detrimental genetic load in CBC pandas was lower than that in wild pandas.

Genetic Purging.

Large-effect deleterious alleles are usually recessive and masked in large populations, but could lead to inbreeding depression when they are exposed in small populations (33). Genetic purging always occurs but occurs more often in small populations and is facilitated by inbreeding to reduce large-effect deleterious mutations in homozygous form that affecting fitness and viability under purifying selection (34). Since all giant panda individuals sampled in this study were either mature juveniles or adults, the severely deleterious mutations affecting survival are less likely to persist in the homozygous state in the ROH region, which is suitable for evaluating the efficiency of genetic purging. Here, we found a significantly lower frequency of LoF and dnsSNP inside ROH than outside ROH regions in all giant panda populations (Fig. 5C and SI Appendix, Fig. S27), which indicated that many recessive strongly deleterious mutations still existed in the heterozygous state harbored in the non-ROH regions and had not been effectively removed from the panda populations. Otherwise, there would be no difference in the frequency of deleterious mutations in ROH vs. non-ROH regions because the severely deleterious mutations have already been purged from the population and the rest should be less harmful or at least not fatal; thus, their frequency would not be affected by repeated inbreeding and homozygosity (9). We obtained the same result by focusing on only heterozygous deleterious mutations (SI Appendix, Fig. S27). This phenomenon in giant pandas may result from low inbreeding, which is less efficient in unmasking such mutations (9, 24, 35). However, the difference in ROH vs. non-ROH regions (measured by P-value) was smaller in the QLI and LSH populations than that in the MSH and QX populations. This indicated that the purging of severely deleterious mutations is more efficient in the QLI and LSH populations due to higher inbreeding levels than that in the MSH and QX populations, which was also evidenced by the R_XY analysis (SI Appendix, Fig. S28). GERP analysis revealed that the difference in relative mutational load between panda populations with high GERP scores was similar to that calculated with relatively low GERP scores, which indicated that the efficiency of genetic purging was not significantly greater for the most deleterious alleles (24) (SI Appendix, Fig. S29). Finally, we still found that putatively damaging alleles were present at a lower frequency than neutral alleles across giant panda populations, indicating that purifying selection had been working, although purging might not have been strong (SI Appendix, Fig. S30). Similarly, we did not detect obvious genetic purging in captive pandas, which is also expected considering the improved breeding management of captive giant pandas (Fig. 5C).

Simulations of Releasing Captive Pandas to the QLI and LSH Populations.

To predict the potential accumulation of deleterious mutations in the recipient wild populations over generations by introducing captive pandas originating from different genetic lineages (different mountain ranges), we performed forward-in-time simulations for the two smallest populations (QLI and LSH) under three scenarios: 1) captive pandas originating from the same mountain ranges as the recipient wild populations; 2) hybrid captive pandas originating half from the recipient wild populations; and 3) captive pandas originating from mountain ranges different from the recipient wild populations (Fig. 5 D and E and SI Appendix, Fig. S31). The results indicated that introducing captive individuals originating from the same mountain ranges as the recipient population seemed to be the best strategy compared to the other two scenarios. Here, in both the QLI and LSH populations, we observed a decreasing trend of heterozygous deleterious mutations and an increasing trend of the homozygous deleterious mutations in the offspring under all the three scenarios during the population recovery stage (36). Interestingly, we found that the number of heterozygous deleterious mutations was positively correlated with the genetic relationship between the introduced CBC population and the recipient wild population, and introducing captive individuals originating from the same mountain ranges as the recipient population (e.g., CBC_QLI to QLI, CBC_LSH to LSH) would introduce the lowest level of genetic load during the population recovery.

Discussion

Large-Scale Population Genomics Reveals a Distinct Genetic Structure.

Previous studies agree that the QLI population and the MSH population are two genetically distinct populations (22, 37–39). However, the genetic relationships among other giant panda populations are controversial. Microsatellite data supported that the DXL, XXL, and LSH populations belonged to the same genetic cluster (37), but genome-wide data supported that the QLA, DXL, XXL, and LSH populations comprised unique genetic clusters (QXL) (22). With deep resequencing of a large sample size, we found that the QXL population can be clearly divided into two parts, the QLA-DXL-XXL (QX) and the LSH populations, which is supported by multiple lines of evidence (Fig. 2). The habitats of the QLA, DXL, and XXL populations were naturally connected before the 1950s according to the 3rd national giant panda survey (40), which could also support the findings of our study. Two other possibilities could be envisaged: 1) The LSH population might have been separated into two subpopulations, and one subpopulation was mixed with the QLA population; 2) we may not have included enough individuals from DXL and XXL. However, both of these possibilities could be ruled out: 1) Previous genetic evidence supported that giant pandas in the Liangshan Mountain are from a single population without any substructure (41, 42); 2) the published WGS data contained samples from DXL and XXL, which were both mixed with the QLA population (22). We then concluded that the current giant panda population should be divided into four main genetic clusters: the Qinling population, the Minshan population, the Qionglai-Daxiangling-Xiaoxiangling population, and the Liangshan population.

Previous studies did not find any genetic substructure within each wild population (15, 22, 37). Here, we first identified a genetic substructure within the MSH population that was consistent with the geographical distribution of the MSH1, MSH2, and MSH3 subpopulations in the Minshan G, J, and K populations, respectively (31). This genetic differentiation in the MSH population might be a result of both natural barriers and human disturbance: 1) The MSH1 and MSH2 subpopulations were isolated by the Duobu River, a large tributary of the Fu River, and the MSH2 and MSH3 subpopulations were isolated by the Fu River (Fig. 1), which might be the natural barriers for giant pandas; 2) Early human settlements often occurred along rivers (43), reflected by the discovery of ancient culture remains along rivers in the Sichuan Province (44). Moreover, the Fu River has been a main waterway since the Qin Dynasty, and many ancient towns were established along the river and have been developing since then (43, 45). Thus, human disturbance along natural barriers might further hinder the connection between giant panda populations. We found that one individual sampled from MSH3 habitat was clustered into the MSH2 population, and three individuals from MSH2 were clustered into the MSH1 population. By further inspection, we found that three of the four individuals were females. It is possible that these four individuals reflect the natural migration of giant pandas between subpopulations, considering the female-biased dispersal of giant pandas (46), although we cannot completely exclude the possibility of sampling record errors.

A concern in this study is that many samples in this study were collected several decades ago, which may not reflect the current population status. A previous study compared several key genetic parameters between historical and current giant pandas and revealed no significant differences (47), supporting the robustness of our inference to the current status of the giant panda. We delineate the fine-grained genetic structure of the giant panda revealed by the analysis of a larger population comparable to almost 30% of living wild giant pandas. We envisage that this study will serve as a valuable reference for future conservation.

Moderate Inbreeding with Less Efficient Genetic Purging.

The IUCN has down-listed the giant panda from “Endangered” to “Vulnerable” (17), but its wild habitats are contracted and fragmented. The giant panda is still under first-class protection in China. Fortunately, the inbreeding level in giant pandas is moderate, with an average F_ROH in the wild giant panda population of ~0.22, which is much lower than that in tigers [captive South China tiger: F_ROH ≈ 0.40 (25); wild Amur tiger: F_ROH ≈ 0.51 (26, 48)]. Furthermore, the prevalence of inbreeding in captive pandas was much lower than that of wild pandas (F_ROH ≈ 0.18). Although captive pandas are under genetic management, this may not be the only contributor to their lower F_ROH because the accumulative length of ROHs traced 30 generations back in captive individuals was much lower than that in the wild populations (Fig. 4C). This may reflect the heterogeneous genetic background of their parental ancestors (Dataset S1). Importantly, by obtaining complete pedigree information on captive giant pandas over ~70-y of breeding history, we found that the hCBC individuals often presented lower inbreeding levels than did iCBC individuals, which provides potential guidance for future captive breeding management.

Inbreeding promotes the exposure of deleterious mutations in the homozygous state in small populations, which further enhances genetic purging under purifying selection (34). Here, we did not find obvious genetic signals of purging in either wild or captive giant pandas, which is not surprising because it is difficult to efficiently unveil deleterious alleles with relatively mild inbreeding. Nonetheless, this may be a positive signal for the giant panda, because many deleterious alleles may still exist in the heterozygous state. With the finding of a moderate to high level of genome-wide genetic diversity, the extinction risk of giant pandas may not be imminent.

The QLI and LSH Populations Need More Conservation Efforts.

With several decades of conservation efforts, the captive population has increased to more than 600 individuals, while the wild population has reached ~1,900 individuals, with 73% growth compared to the population size in the 1980s. Although the overall extinction risk of this species is not imminent, the QLI and LSH populations seem to be more fragile than the MSH and QX populations. The fourth National Giant Panda Survey also reported much less suitable habitats (49) in the Qinling and Liangshan Mountains (16).

The QLI population has experienced a continuous population decline over the whole Holocene period (Fig. 3C). The Yellow River Valley is the birthplace of the Chinese civilization, and Neolithic cultures have thrived on the Loess Plateau along the Yellow River since 9000 BP (50–52). The human population continuously expanded at this location with the development of agriculture, primarily based on millet during the Yangshao Culture (7000-5000 BP) (52), which made the Qinling region the earliest area in China to be disturbed by human activity. Over the following thousands of years, Xi’an has consistently become one of the most important capital cities in China. The increasing human population, prolonged extensive human activity, and frequent warfare might have been the most significant factors damaging giant panda habitats in Qinling, similar to the sympatrically distributed snub-nosed monkeys (53). Although the climate became colder and drier at ~5,000 kya (44, 51, 54), the Sichuan panda population still expanded (Fig. 3C); therefore, climate may not be the direct driver of the decrease in the QLI population. However, climate change combined with enhanced human activity pushed the northern boundary of bamboo from ~N40° to N35° in China (55), and the reduction in natural bamboo forests in Qinling might be another important factor for the continuous decline of the QLI population. As the population decreased, the QLI population reached its highest inbreeding level and accumulated the highest homozygous genetic load, which implied that the QLI population may face the greatest fitness cost among giant panda populations. Coincidentally, gene flow between the QLI population and Sichuan populations appears to have been very limited in the past 1,000 y (SI Appendix, Fig. S9). Therefore, special attention should be given to QLI pandas, instead of downlisting the conservation status of giant pandas as a whole.

The LSH population also represents a relatively isolated and ancient population. With a much smaller effective population size, the genetic load in the LSH population ranked second only to that in the QLI population (Figs. 3C and 5B). From the 1950s to the 1990s, the fastest population decrease was also found in LSH (56). Although wild LSH pandas have been well protected in the last two decades, both the population increase (7.8%) and the decrease in human interference (48.4%) in LSH were the lowest among the wild populations (16). Currently, the Liangshan population consists of five isolated populations (13), the two largest populations (Liangshan A and Liangshan B) are both close to developed lands (31), and it is very difficult for Liangshan D and Liangshan E (only three individuals) to connect with neighboring large populations. Even worse, the Liangshan habitat is not included in the Giant Panda National Park (20), which may increase the risk of inbreeding in the LSH population and greatly elevate the possibility of extinction (13). The LSH population undoubtedly represents another giant panda population that deserves special protection.

Future Genetic Rescue.

Considering the issues discussed above, genetic rescue to improve the recovery of giant pandas is still necessary. Captive breeding programs are very successful in China, with lower inbreeding and genetic load than those of wild giant pandas. The Chinese government has successfully released more than 10 individuals to the wild, and some of them have reproduced in the wild. We are convinced that a scientific genetic assessment is a prerequisite for efficient reintroduction and could further improve the success rate. Outbreeding depression due to local adaptation represents a serious risk for rescued small populations (10, 29, 30). In this study, we found a small number of genes that were under recent positive selection in both captive and wild giant pandas (SI Appendix, Fig. S32 and Table S6). We did not find positively selected genes that were directly related to the environmental adaptation, although it is very difficult to accurately evaluate these genes without further experimental validation.

Another concern regarding genetic rescue is that assisted gene flow from long-term diverged populations is likely to result in outbreeding depression by introducing private deleterious mutations (10). Fortunately, we found that a very small number of derived deleterious mutations were directly introduced into the wild population in the next generation, when suitable CBC pandas with similar genetic background (e.g., the same mountain ranges) were selected for reintroduction (SI Appendix, Fig. S33). The results from the forward-in-time simulations also showed that introducing captive individuals originating from the same genetic lineage as the recipient wild population will accumulate the least genetic load over generations during population recovery (Fig. 5 D and E and SI Appendix, Fig. S31). Thus, this approach is considered to be the best strategy for reintroduction, providing great hope for the recovery of the endangered subpopulations at high risk of local extinction (13). Our results indicated that rewilding the captive individuals should be a reasonable strategy for genetic rescue, and we considered iCBC individuals of the same wild origin as the recipient populations to be the best candidates for reintroduction. Finally, further empirical genomic and fitness monitoring of the offspring in reintroduction programs would be valuable for current conservation programs, particularly for high-risk populations, such as the QLI and LSH populations.

Materials and Methods

Samples, Sequencing Data, and Ethics Statements.

Seventy-four blood samples from captive giant panda individuals were collected from the China Conservation and Research Center for the Giant Panda, Qinling Giant Panda Research Center, and other zoos in China (Fig. 1). Historical skin samples of wild giant pandas were collected from individuals who died of natural causes during the bamboo flowering events in the early 20th century or from the confiscation before the major protection measures were enacted in 1988. All samples were used for whole genome DNA isolation for resequencing. Blood samples were promptly placed into anticoagulant tubes and transported to the laboratory on ice packs. Upon arrival, these samples were stored at −80 °C until DNA extraction. Skin samples were kept in a dry, low-temperature environment until DNA isolation. In addition, whole-genome sequencing data from 58 giant panda individuals (15, 22) were downloaded from the National Center for Biotechnology Information (NCBI, SI Appendix, Table S2) for downstream population genomic analyses. The collection of samples, execution of experiments, and design of the research in this study were all conducted under the approval and oversight of the Institutional Review Board of BGI (BGI-IRB E22017), ensuring that the study was conducted ethically and with the proper measures in place to safeguard the welfare of the animals involved.

DNA extraction and library preparation for blood and skin samples were conducted according to the manufacturer’s instructions of commercial kits (SI Appendix, SI Methods). Paired-end (100 bp) sequencing with an insert size of ~350 bp was performed on a DNBSEQ T1 sequencer (MGI, Shenzhen, China).

Genome-Wide Variant Calling and Quality Control.

Adapter sequences and low-quality bases were removed from the raw sequencing reads using Trimmomatic (v0.33.0) (57). The WGS files were then aligned to the giant panda genome [Ame_Sichuan.fa (15)] using the Burrows–Wheeler algorithm (58) implemented in Sentieon. Based on the alignment files, we first used mapDamage2 (v2.2.1) (59) to investigate whether the deamination-introduced C-to-T transition had occurred in the old skin samples. After sorting and deduplication of the resulting alignment files, variant calling was performed via the Sentieon Haplotyper pipeline, which is similar to the Genome Analysis Toolkit (GATK) HaplotypeCaller pipeline. Joint genotyping was performed using the Sentieon DNAseq GVCFtyper to produce our final multisample Variant Call Format (VCF) file.

Subsequently, SNP and InDel variant sets were extracted from the comprehensive VCF file using GATK (v4.0.3.0) (60) with the “SelectVariants” parameter. Hard filtering was applied to eliminate low-quality variants in the primary VCF file with the parameters “QD < 2.0 || FS > 60.0 || MQ < 40.0 || MQRankSum < −12.5 || ReadPosRankSum < −8.0” (61) for SNPs and “QD < 2.0 || FS > 200.0 || ReadPosRankSum < −20.0” for InDels, which are recommended by GATK as best practices. Then, multiallelic variants and variants on sex chromosomes were excluded from the variant set. Additionally, variants with either extremely low (<0.5%) or extremely high (>99.5%) sequencing depth across individuals were removed. Furthermore, we assessed the genotype of each individual by examining Phred-scaled likelihood (PL) values for the three possible genotypes (0/0, 0/1, 1/1), and only genotypes where one of the PL values equaled 0 and the other two were 20 or higher were retained (62). SNPs with a missing rate exceeding 20% were filtered out from the VCF file.

We identified closely related individuals by computing the kinship coefficient using KING (v2.2.4) (63) software, as detailed in SI Appendix, SI Methods. We applied BCFtools (v1.11) (64) statistics to obtain the summary information of variants for each population and the whole population.

Population Structure Analysis.

For PCA, we used the PLINK (v1.9) (65) software to convert the VCF files into PLINK files. Then, PCA was carried out using the genome-wide complex trait analysis (GCTA) (v1.92.2) (66) with default parameters. To construct the population-based phylogenetic tree, we first conducted LD pruning of the SNP dataset using PLINK and utilized vcf2phylip (v2.7) (67) to convert the pruned VCF file into PHYLIP format. Afterward, we constructed a ML phylogeny for the giant panda individuals using IQ-TREE (v1.6.12) (68) with the recommended nucleotide substitution model “GTR+F+G4” calculated by jModelTest (v2.1.10) (69). Population structure was analyzed by the model-based clustering method ADMIXTURE (v1.3.0) (70), with cluster numbers (K) ranging from 2 to 10. We calculated the F_ST between populations by using the vcftools (v0.1.16) (71) software with the following parameters: “vcftools --gzvcf vcf.gz --weir-fst-pop pop1.list --weir-fst-pop pop2.list --fst-window-size 50000 --fst-window-step 10000 --out result.” We also used vcftools to calculate the genome-wide π with the following parameters: “vcftools --gzvcf vcf.gz --window-pi 500000 -out result.” For the wild samples without detailed geographic origins, we first performed a population assignment with genotype likelihoods using LASER (v2.0) (72) for PCA, and we further performed supervised admixture modeling to confirm the assignment with the ADMIXTURE software by using samples of known origins as reference individuals with the parameters of “--supervised.” Unsupervised PCA clustering of all wild samples was also conducted with GCTA software to confirm the results. We finally assessed the robustness of population structure analysis by employing the bootstrap strategy for samples from known sampling sites (SI Appendix, SI Methods).

Gene Flow and Admixture Graph.

Short-read sequencing data of a polar bear (Ursus maritimus, SRA: SRR15170755) were mapped to the giant panda genome to generate variants, which served as an outgroup. TreeMix (v1.13) (73) was used to detect population migration events with the parameter “-m 1-20 -k 1000 -root Polarbear.” The refined-ibd software (16May19.ad5.jar) (74) was used for calculating the identity by descent (IBD) fragment with the following parameters: length = 0.01 trim = 0.005. The IBD fragments of pairwise individuals were divided into different expected generations (g) according to the equation l = 100/(2 g), where l is the genetic distance of the IBD in cM (75). Here we applied the typical recombination rate of large mammals (1 Mb = 1 cM) (76).

We performed D-statistics analysis (A, B; X, Y) by using qpDstat in ADMIXTOOLS (v5.1) (77), where we set U. maritimus as Y, based on phylogenetic topology ((((QX, MSH), LSH), QLI), polar bear). Next, we used the qpGraph program (77) to model the population splitting and admixture events among diverse giant panda populations with the following parameters: “outpop, NULL; useallsnps, YES; blgsize, 0.005; lsqmode, YES; diag, 0.0001; hires, YES.” We assembled mitochondrial genome sequences for wild samples with detailed origins using NOVOPlasty (78) (v4.3.1) and constructed a haplotype network map using PoPART software (v1.7) (79) (SI Appendix, SI Methods).

Population Demographic History.

We combined theories of multiple sequentially Markovian coalescent (MSMC), approximate Bayesian computation (ABC), and linkage disequilibrium (LD) to infer the change in the effective population size (Ne) of the giant panda over generations. First, we randomly selected four individuals from each population to infer population history by using MSMC2 (80) with the parameters “-R -i 20 -t 6 -p ‘10*1 + 15*2’.” The input VCF files were phased by BEAGLE (version 5.0) (81) with default parameters, and the uncovered regions were masked with bamCaller.py. The final result was visualized with a generation time of 12 y and a mutation rate of 1.29 × 10⁻⁸ substitutions per site per generation (15). To further explore more recent changes in effective population size, we used SNPs from 10 randomly selected individuals to run PopSizeABC (v2.1) (82) with the following parameters: mac (minor allele count threshold for AFS and IBS statistics computation) of 0; mac_ld (minor allele count threshold for LD statistics computation) of 3 and 5 respectively; L (size of each segment, in bp) of 4000000; nb_rep (number of simulated datasets) of 500; nb_seg (number of independent segments in each dataset) of 30. We used the LD-based method implemented in GONE (v1.0) (83) to estimate recent N_e of the past 50 to 100 generations. We equalized the sample size to be 10 for each wild population. We kept the maximum number (50,000) of SNPs per chromosome to be analyzed with the parameter hc of 0.02 and Haldane’s correction. The analysis was repeated 500 times for each jackknife of different sets of 10 individuals.

Population divergence time was inferred using four randomly selected samples from each population by MSMC2 (v2.1.2) with the following parameters: --skipAmbiguous -i 20 -t 6 -p ‘1*2 + 15*1 + 1*2’. Then we used a python script (combineCrossCoal.py) to obtain the relative cross-coalescent rate (RCCR) and visualized it by a Perl script with the same mutation rate and generation time as the Ne inferred by MSMC2.

Inbreeding Estimation.

ROH were detected using PLINK with the following parameters: --homozyg-window-snp 20 --homozyg-kb 100 --homozyg-density 50. The individual inbreeding coefficient, F_ROH, was calculated as the proportion of ROH fragments relative to the total length of the 20 autosomes. The coalescent times (generation) of ROH were estimated by applying the same method and parameters for the above IBD analysis. The two-sided pairwise t test was conducted in R (version 4.1.2) (84).

Screening of Genetic Load.

We estimated the genetic load in giant panda genomes using two approaches. First, we annotated synonymous and nonsynonymous mutations within coding regions using SnpEff (v4.3) (85). Before this, we used all published bear genomes to infer ancestral alleles in the giant panda genome for screening derived alleles. Whole-genome sequencing data from Ursus maritimus (SRA: SRR15170755), Tremarctos ornatus (SRA: SRR16086837), Ursus arctos (SRA: SRR22801817), Ursus americanus (SRA: SRR20985020), Helarctos malayanus (SRA: SRR13167983), and Ursus thibetanus (SRA: DRR250459) were all mapped to the giant panda genome, and the major alleles at each locus were used to represent the ancestral state. Sites without full alignment of the six bear genomes were excluded from the downstream analysis. After replacing the reference allele with the ancestral allele using a custom Perl script, we retained a total of 12,302,938 SNPs. We identified three different categories of mutations in the SnpEff software, including 1) synonymous mutations, 2) missense mutations, and 3) loss of function (LoF) mutations. Here, we considered “stop_gained,” “start_lost,” “stop_lost,” “splice_donor_variant,” and “splice_acceptor_variant” as LoF mutations. We then used the ANNOVAR (v2020Jun08) software (86) to annotate nonsynonymous SNPs with the parameter “--aamatrixfile grantham matrix.” Deleterious nonsynonymous SNPs (dnsSNPs) were diagnosed by calculating the Grantham score (GS) (87), a measurement of the physical/chemical properties of amino acid changes. Nonsynonymous SNPs with a GS score greater than 150 were considered as the dnsSNPs (88). Next, we counted the number of derived mutations per individual in the homozygous and heterozygous states. The proportion of homozygous-derived SNPs was calculated with the following formula: 2 × homozygous sites/(2 × homozygous sites + heterozygous sites) (89). The significance of the difference between giant panda populations was estimated by a two-sided pairwise t test in R. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed in R with the package “clusterProfiler” (90, 91). We further aligned giant panda protein-coding sequences to the human genome and compared them against the Human Gene Mutation Database v2017.1 (HGMD® http://www.hgmd.org/) to screen for LoF mutations and dnsSNPs that were more likely to be deleterious.

Second, we measured mutations at sites under strict evolutionary constraints that may have essential functional roles. The genomic evolutionary rate profiling score (GERP) of each locus in the giant panda genome was calculated with the GERP++ software (92), as detailed in SI Appendix, SI Methods. Mutations occurring at highly conserved sites (i.e., with higher GERP scores) are likely to be more deleterious. The individual relative mutational load was measured as the sum of all derived alleles multiplied by their GERP score, including only the derived alleles with high GERP scores (>4, >5, and >6) divided by the total number of derived alleles per individual.

Detection of Genetic Purging.

To detect genetic purging, we first compared the number of LoF mutations and dnsSNPs occurring in ROH and non-ROH regions for individuals in each population, normalized by synonymous mutations in the same region (9). We then used the R_XY method to estimate the relative excess of genetic load (9) by randomly selecting 10 individuals from each population to conduct 100 rounds of jackknife calculations. In each round of calculation, we randomly selected 85 to 90% of the loci for synonymous, missense, LoF, and dnsSNP mutations, and for intergenic variants, 100,000 loci were randomly excluded (26). To detect the purifying selection on maintaining the genetic purging, we compared the site-frequency spectra (SFS) of deleterious mutations with those of neutral mutations. We equalized the sample sizes across the populations and across loci within each population by randomly subsampling 20 nonmissing alleles from each locus in each population before estimating the derived neutral (GERP score < 1) and derived damaging (LoF and dnsSNP mutations combined) allele frequencies.

Evaluation of Reintroduction Strategies.

Recent positively selected genes that potentially relevant to local adaptation were identified in the wild and CBC populations by applying the Cross Population Extended Haplotype Homozygosity (XP-EHH, v20090727) (93) method, as detailed in SI Appendix, SI Methods. We predicted the current risk of assisted gene flow between giant panda populations by counting newly introduced deleterious alleles from the source population to the recipient population (10). The deleterious mutations that existed in the source populations, but were absent in the recipient population, were defined as newly introduced deleterious mutations. To normalize the effect of different wild population sizes, we randomly selected 10 individuals to represent each recipient population. To predict the future impact of reintroduction on the change and accumulation of deleterious mutations, we performed the forward-in-time simulations with SLiM (v4.1) (94) software. We used the giant panda reference genome and the phased VCF file as the two input files in the simulation with the following parameters: a non-Wright–Fisher (nonWF) model, a mutation rate (1.29e-8), a recombination rate (1e-8), sex ratio (1:1), reproduction age (6 ~ 20 y old, mean = 12), and life table with higher mortality of the cub and the old. The survival rate in each generation was scaled by the individual age and environment carrying capacity (K). Eight wild pandas and three captive pandas were selected to serve as founders in each simulation and this process was repeated five times for each scenario. The genomes of the descendants were generated for every generation to predict the accumulation of derived deleterious mutations in the population recovery stage (1 ~ 30 generations). The “generation” concept here is similar to a reproduction cycle according to the SLiM manual (~2 y for giant pandas) (95).

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

The whole genome resequencing data of the giant pandas included in this study are publicly available in the CNGB Sequence Archive (CNSA) (96) of the China National GeneBank DataBase (CNGBdb) (97) under accession number CNP0001699. The giant panda reference genome (CNA0007300) was downloaded from https://db.cngb.org/search/assembly/CNA0007300/ (98). Previously published data used for this work (CNP0000785, SRP013618) were downloaded from https://db.cngb.org/search/project/CNP0000785/ (99) and https://www.ncbi.nlm.nih.gov/sra/?term=SRP013618 (100). All other data are included in the manuscript and/or supporting information.