Genome-wide association studies have identified common variants associated with risk of esophageal squamous cell carcinoma (ESCC). However, these common variants cannot explain all heritability of ESCC. Here we report an exome-wide interrogation of 3,714 individuals with ESCC and 3,880 controls for low-frequency susceptibility loci, with two independent replication samples comprising 7,002 cases and 8,757 controls. We found six new susceptibility loci in CCHCR1, TCN2, TNXB, LTA, CYP26B1 and FASN (P = 7.77 × 10−24 to P = 1.49 × 10−11), and three low-frequency variants had relatively high effect size (odds ratio > 1.5). Individuals with the rs138478634-GA genotype had significantly lower levels of serum all-trans retinoic acid, an anticancer nutrient, than those with the rs138478634-GG genotype (P = 0.0004), most likely due to an enhanced capacity of variant CYP26B1 to catabolize this agent. These findings emphasize the important role of rare coding variants in the development of ESCC.
Esophageal squamous cell carcinoma (ESCC) is the fourth leading cause of cancer death in both men and women in China1. Approximately half of the world’s 500,000 new cases of ESCC each year occur in this country1,2. Through four genome-wide association studies (GWAS), we and others have identified several variants associated with ESCC risk in Chinese populations, some of which showed significant interaction with lifestyle, including alcohol consumption and tobacco smoking3,4,5,6,7. However, most of these susceptibility variants are common variants, with minor allele frequency (MAF) > 0.10 in these populations, and confer relatively small effect sizes, with odds ratios ranging from 1.1 to 1.5. These variants thus explain only a small fraction of the genetic risk of ESCC and improve the area under the curve (AUC) by only 7.0% for a risk model as compared with the AUC of a model using the four nongenetic factors (sex, age, smoking status and drinking status)8. Therefore, other rare loci that are undetected by commonly used commercial SNP chips may have a relatively larger effect size and contribute to the ‘missing’ heritability. The Illumina HumanExome Beadchip (referred to as the ‘exome chip’ hereafter) platform has thus been developed to capture low-frequency or rare variants in coding regions and has been demonstrated to be an effective complementary approach for identifying genetic susceptibility loci for complex diseases or traits9,10,11,12.
In this study, we first scanned 3,721 affected individuals (cases) and 3,889 cancer-free individuals (controls) from Beijing using an exome chip (Supplementary Fig. 1 and Supplementary Table 1); after quality control, 3,714 cases and 3,880 controls were retained for the association analysis. No significant population stratifications were observed for the cases or controls in this stage (Supplementary Figs. 2 and 3). All of the variants identified in the previous GWAS7 were successfully verified by direct genotyping or imputation in the present study except for rs35597309 in 6p21.32 (P = 0.1865) in this study (Fig. 1 and Supplementary Table 2). This variant was also not significant (P = 0.3590) in the replication sample from Beijing in the previous GWAS7. The overall association results are shown in Fig. 1, and 30 SNPs were identified as significantly associated with ESCC risk with P < 1.00 × 10−4 after quality control (Supplementary Fig. 1 and Supplementary Table 3). We then replicated these SNPs in two independent replication sample sets, Replication I from Hubei province and Replication II from Hebei province, with 7,002 ESCC cases and 8,757 controls in total (Supplementary Table 1). Upon combining the results from the discovery and replication stages (Supplementary Table 4), we found three low-frequency (MAF < 0.05) coding variants that were significantly associated with risk of ESCC: rs117353193 (NC_000022.10: g.31010417 G > A) at TCN2 (p.Arg170Gln; odds ratio (OR) = 1.67, 95% confidence interval (CI) 1.51−1.84, P = 1.74 × 10−23); rs138478634 (NC_000002.11: g.72360331 G > A) at CYP26B1 (p.Arg323Trp; OR = 1.82, 95% CI 1.58−2.09, P = 9.45 × 10−17); and rs17848945 (NC_000017.10: g.80040513 C > T) at FASN (p.Val1937Ile; OR = 1.67, 95% CI 1.44−1.94, P = 1.49 × 10−11) (Table 1 and Supplementary Table 5). Additionally, we also identified three common missense variants that were significantly associated with ESCC susceptibility in the 6p21.3 region: rs130079 (NC_000006.11: g.31112737 C > A) at CCHCR1 (p.Gly664Cys; OR = 1.39, 95% CI 1.31−1.49, P = 7.77 × 10−24); rs204900 (NC_000006.11: g.32056580 A > C) at TNXB (p.Ser921Ala; OR = 1.34, 95% CI 1.26−1.42, P = 3.16 × 10−23); and rs1041981 (NC_000006.11: g.31540784 C > A) LTA (p.Thr60Asn; OR = 0.82, 95% CI 0.80−0.86, P = 9.37 × 10−22) (Table 1 and Supplementary Table 5).
To investigate whether the association with ESCC risk at each of the four susceptibility regions can be completely explained by the index SNP, we then performed imputation analysis for the identified four regions. After imputation, we tested 9,411 SNPs (1,150 directly genotyped and 8,261 well-imputed SNPs) for association, and 201 SNPs (7 directly genotyped and 194 imputed SNPs) on chromosome 6p21.3 satisfied the significance threshold in this study (P < 1.00 × 10−4). These SNPs were in linkage disequilibrium (LD) with genotyped SNPs identified at rs130079 (r2 = 0.32−1.00), rs204900 (r2 = 0.35−1.00) and rs1041981 (r2 = 0.31−1.00) (Supplementary Fig. 4 and Supplementary Table 6). After conditioning on each of the three SNPs, the association P values for the SNPs in this region were no longer less than 0.01, suggesting that the association signals in this region probably point toward these three genotyped SNPs. We further imputed HLA alleles in 6p21 region with a HAN-MHC reference panel on the basis of deep sequencing of 10,689 Han Chinese individuals13. Our results indicated that no classical HLA alleles are in LD with rs130076 or rs204900 (both r2 < 0.1). Four HLA amino acids are also in low LD with rs1041981 (r2 = 0.14−0.16) (Supplementary Table 7). We then performed conditional logistic regression analysis and found that, after conditioning on rs1041981, the association P values for the four HLA amino acids were not significant (all P > 0.05), suggesting that the association signals in this region may come from rs1041981. These results indicate that the three variants in 6p21 region identified in this study are independent of HLA alleles. In the other three chromosomal regions that contained associated low-frequency variants, no imputed SNPs satisfied the significance threshold in this study (Supplementary Fig. 4). However, two variants, rs139481399 and rs137943394, were found to be in high LD (r2 > 0.7) with the typed SNPs rs117353193 and rs138478634, respectively (Supplementary Figs. 5–7).
We next analyzed the associations of the six newly identified variants stratified by age, sex, smoking status and drinking status. In the stratified analysis, the risks of ESCC associated with rs130079 and rs204900 were significantly different among individuals with different ages. The risks of ESCC associated with rs1041981, rs117353193 and rs1041981 were significantly different among different replication stages (Supplementary Fig. 8). Although the exact reasons for these differences need to be further addressed, the heterogeneity might be due to different sample sizes and reflect different gene–environment interactions where samples were collected. Most interestingly, the association of rs138478634 at CYP26B1 was lifestyle dependent, with a more pronounced enhancement of risk observed in smokers and drinkers (Supplementary Fig. 8e). Interaction analyses showed that rs138478634 interacted multiplicatively with tobacco smoking (P = 0.0239) and alcohol consumption (P = 0.0405) to contribute to ESCC risk. Smokers or drinkers with the risk allele of rs138478634 had an OR 2.68-fold or 2.62-fold higher, respectively, than the ORs for smokers or drinkers without the risk allele and 2.34-fold or 3.05-fold higher than the ORs for nonsmokers or nondrinkers who had the risk allele but did not smoke or drink (all P < 0.001 for heterogeneity test, Fig. 2). We then examined whether the rs138478634 variant is associated with smoking or drinking status. The rs138478634 variant was not significantly associated with smoking or drinking status in controls, indicating that the interaction may be toward ESCC carcinogenesis (Supplementary Table 8). However, the gene–lifestyle interaction of rs138478634 was driven mostly by replication II (Supplementary Table 9), and whether it is true needs future verification with larger sample size.
Because rs138478634 (G > A change in exon 5 at CYP26B1) had the highest effect size (OR = 1.82) and significant gene–lifestyle interactions with ESCC risk, we performed a series of analyses to determine its function. CYP26B1 is a critical regulator of cellular all-trans retinoic acid (atRA) levels that specifically inactivates atRA to its hydroxylated forms. Because atRA is an important tumor suppressor, we conjectured that rs138478634 might affect cellular atRA levels by altering the catabolic activity of the enzyme. To test this hypothesis, we overexpressed different CYP26B1 variants in KYSE30 and KYSE150 cells by transfecting pcDNA3.1-CYP26B1 plasmid containing CYP26B1 rs138478634[G] or rs138478634[A] and measured cellular levels of atRA by reverse-phase high-performance liquid chromatography (HPLC) 48 h after the addition of atRA to culture medium. We found that, as compared with cells that expressed CYP26B1 from the rs138478634[G] allele, cells that expressed CYP26B1 from the rs138478634[A] allele had significantly lower cellular atRA levels (Fig. 3a), indicating that rs138478634 SNP may enhance the enzymatic activity of CYP26B1 in catabolizing atRA. We further measured serum levels of atRA in 404 individuals with different CYP26B1 rs138478634 genotypes using ultra-performance liquid chromatography–tandem mass spectrometry. We found that individuals with the rs138478634-GA genotype had significantly lower serum atRA levels than those with the rs138478634-GG genotype (P = 0.0004; Fig. 3b). We also examined whether CYP26B1 affects ESCC cell proliferation, and the results showed that overexpression of CYP26B1[A] significantly enhanced cell proliferation compared with overexpression of vector control or CYP26B1[G] (Fig. 3c,d), whereas knockdown of CYP26B1 significantly reduced cell proliferation (Fig. 3e,f).
In summary, through an exome-wide association study, we identified six ESCC susceptibility variants located in four genomic regions. Among these variants, rs138478634 in CYP26B1 had the highest effect size. CYP26B1 acts as a critical regulator of atRA levels by specifically inactivating atRA to hydroxylated forms. Retinoic acid is an important regulator of cell development, differentiation and the immune system, and it suppresses carcinogenesis in animal models of skin, oral, lung, breast, bladder, ovarian, prostate and esophageal cancers14,15,16,17. Furthermore, the retinoic acid synthesis–related gene ALDH1A2 (encoding aldehyde dehydrogenase) and retinoic acid–inducible gene I (RIG-I) may be candidate tumor suppressors; in contrast, overexpression of the atRA-catabolizing enzyme CYP26A1 promotes colorectal tumorigenesis18,19,20,21,22. However, little is known about the role of CYP26B1 in human cancer. In the present study, we demonstrated that a genetic variant in CYP26B1, which enhanced the catabolic activity for atRA nearly 35% (estimated based on the reduction of cellular atRA level), is associated with ESCC. In addition, individuals with the CYP26B1 risk genotype had a lower level of serum atRA, and knockdown of CYP26B1 expression significantly reduced ESCC cell proliferation. Significant SNP–smoking and SNP–drinking interactions were also observed for the CYP26B1 variant, and the results could be explained by interactions between beta-carotene and tobacco smoking and alcohol consumption23,24,25,26.
Besides the functional variant at CYP26B1, we also identified two other low-frequency variants (rs17848945 at FASN and rs117353193 at TCN2) that are significantly associated with the risk of developing ESCC. FASN catalyzes the terminal steps in the de novo biogenesis of fatty acids, and SNPs in FASN have been associated with types of neoplasms other than ESCC, such as prostate cancer, uterine leiomyoma and breast cancer27,28,29,30,31,32. TCN2 is an essential carrier for the cellular uptake of vitamin B12, and genetic polymorphisms in TCN2 are associated with LINE-1 methylation, CpG island methylator phenotype (CIMP) status and risk of colorectal cancer33,34,35,36,37. Additionally, we identified three moderate-frequency susceptibility variants located in 6p21.3, which contains the major histocompatibility complex (MHC), a susceptibility hot spot for multiple types of cancer38,39,40,41. The most significant SNP we identified in this region is rs130079 in CCHCR1, which is a psoriasis susceptibility gene42, and previous studies have identified another coding variant at CCHCR1, rs130067 (r2 = 0.05 with rs130079), that is associated with increased prostate cancer risk, possibly indicating an important link between autoimmune diseases and cancer43,44. Two other significant markers identified in 6p21.3 were rs204900 in TNXB and rs1041981 in LTA. TNXB has been shown to regulate collagen synthesis, and its deposition and absence enhance invasion and metastasis in melanoma cells; additionally, TNXB interacts with vascular epidermal growth factor B (VEGF-B) and enhances the ability of VEGF-B to stimulate cell proliferation45,46,47. LTA (also known as TNFB) belongs to the tumor necrosis factor family and plays an important role in inflammatory and immunologic response48. Genetic polymorphisms in LTA, including rs1041981, which was identified in this study, have already been found to be correlated with susceptibility to multiple types of human cancer, including cancer of the stomach, breast, cervix and colorectum and non-Hodgkin’s lymphoma49,50,51,52,53,54,55,56,57.
The present study is the first exome-wide association study for ESCC with a multistage design involving discovery and replications from three independent sets of cases and controls, which should reduce false positive findings from exome-wide screening. We have identified six new ESCC susceptible variants located in exonic regions, including low-frequency variants that have not been identified in the previous GWAS with common variants. Therefore, this study highlights the important roles of low-frequency genetic variants in the development of this malignancy. A limitation of the study is that we did not systematically investigate all six of the identified variants. However, functional analysis was performed for a CYP26B1 variant, rs138478634, supporting the biological plausibility of this association. These results extend previous findings and advance our understanding of the genetic etiology of ESCC, which might be useful for risk assessment, early detection and targeted treatment of ESCC.
Study subjects and genotyping analysis
We conducted a three-stage study, and the characteristics of the subjects are summarized in Supplementary Table 1. In the discovery stage, 3,721 ESCC cases and 3,889 controls who were recruited from Beijing were genotyped using the Illumina HumanExome Beadchip system to identify potential susceptibility variants. The case and control samples were mixed and randomly allocated in the plates. All initial genotyping of cases and controls was done at the same time on the same platform, and researchers performing the assays were blind to case/control status. The sample size has 50% to 99% power to detect variants with MAF ranging from 0.01 to 0.05 (OR = 1.5). Some of these individuals were also included in our previous GWAS5,6,7. Genotypes were called by the Illumina GenomeStudio software, and the selected variants were re-called by zCall58. Quality control of the raw genotyping data was performed to filter unqualified genetic variants and samples (Supplementary Fig. 1).
A total of 158,389 variants were excluded from subsequent analysis because they (i) had duplicate variants on the chip (831 variants), (ii) were mitochondrial variants or were located on the X or Y chromosome (1,494 variants), (iii) were monomorphic in our study subjects (155,216 variants), (iv) had a call rate of < 95% (372 variants), or (v) presented a P value of <0.0001 in a Hardy–Weinberg equilibrium test among the control subjects (476 variants). A total of 7 ESCC cases and 9 controls were excluded because they (i) had an overall genotyping rate of <95% or (ii) had an extreme heterozygosity rate more than 6 s.d. from the mean. Ancestry and population stratification were detected with a method based on principal-component analysis using EIGENSOFT59 and 4,604 autosomal informative ancestry markers included on the exome chip. We determined identity-by-state similarity to estimate the cryptic relatedness or duplication for each pair of samples using PLINK software, and no duplicated individuals (PI_HET > 0.25) were found in this study. As shown in Supplementary Figs. 2 and 3, no individuals were excluded as outliers, and the ESCC cases and controls were genetically matched with a small inflation factor (lambda = 1.032). Nearly half of the variants interrogated in this study were low-frequency or rare variants (MAF < 0.1) (Supplementary Fig. 9). Genotyping consistency was assessed based on of 300 replicate samples genotyped using both exome chip and OpenArray system for 30 promising SNPs, and the concordance rate of each variant was between 99.7–100% (Supplementary Table 10).
In the first replication stage (Replication I), 30 promising SNPs were genotyped using the OpenArray genotyping platform in 3,120 ESCC cases and 3,919 controls recruited from Wuhan. In the second replication stage (Replication II), nine promising SNPs were genotyped using TaqMan assays platform (ABI 7900HT system, Applied Biosystems) in 3,882 ESCC cases and 4,838 controls recruited from Hebei province. These cases were recruited from the Han Chinese population through collaboration with multiple hospitals in Beijing, Wuhan and Hebei province. All controls were cancer-free individuals selected from a community nutritional survey in the same region during the same period as when the cases were collected, some of which were also included in our previous studies60,61. Several genotyping quality controls were implemented in the replication stage, including (i) case and control samples were mixed in the plates, and individuals who performed the genotyping assay were unaware of case or control status; (ii) positive and negative (no DNA) samples were included on every 384-well assay plate; and (iii) we further employed direct sequencing of PCR products to replicate sets of 50 randomly selected, OpenArray-genotyped and TaqMan-genotyped samples, and the concordance rate of OpenArray and TaqMan platforms for each variant was 98.0–100% (Supplementary Table 10).
Demographic characteristics, including age, sex and history of smoking and drinking, were obtained from the medical records of these individuals. The ESCC diagnosis was confirmed histopathologically or cytologically by at least two local pathologists according to the World Health Organization classification. All the cancer cases had no previous diagnosis of another type of cancer. We confirm that our study is compliant with all relevant ethical regulations. Informed consent was obtained from all participants at recruitment, and this study was approved by the Institutional Review Board of the Chinese Academy of Medical Sciences Cancer Institute and Tongji Medical College, Huazhong University of Science and Technology. We confirm that our study is compliant with the “Guidance of the Ministry of Science and Technology (MOST) for the Review and Approval of Human Genetic Resources.”
Quality control and association analysis
In this study, we mainly focused on low-frequency but not rare variants. Therefore, we excluded variants with MAFs less than 0.1%. Assuming an additive genetic model, we performed a single-variant association analysis by using a logistic regression model adjusted for age, sex, smoking status, drinking status and the first three principal components as implemented in PLINK. Based on the following criteria, we then selected 30 promising variants for further genotyping in the replication stages: (i) the single-variant association P value was less than 0.0001 and (ii) variants were in low LD (r2 < 0.8) with previous GWAS-identified variants. Variants with association P values between 0.001 and 0.0001 are also shown in Supplementary Table 11 for further evaluation in the future. In the replication stage, association analyses were performed using a logistic regression model adjusted for age, sex, smoking status and drinking status. Nine variants with a FDR q < 0.05 in the first replication stage showed consistent associations with variants found at the discovery stage and were further genotyped in the second replication stage (Supplementary Table 3). During the second replication stage, we genotyped these nine variants and identified six with associations that were consistent with the discovery stage and the first replication stage and had a FDR q < 0.05 (Supplementary Table 4).
Genotype imputation LD pattern analysis
We performed imputation using MaCH62 software to impute ungenotyped SNPs in a region of 1 Mb or 500 kb centered on six significant SNPs. This analysis was based on the LD and haplotypes information from 1000 Genomes Project Phase 3 ASN samples as references. Imputed variants with MACH Rsq < 0.3 were excluded for the further analysis. HLA alleles were imputed using SNP2HLA63 with a HAN-MHC reference panel based on deep sequencing of 10,689 Han Chinese individuals13. LD structures and haplotype block plots were generated using LocusZoom64.
Two human ESCC cell lines, KYSE30 and KYSE150, were gifts from Y. Shimada at Kyoto University and were used for functional analysis. Both cell lines used in this study were authenticated by short tandem repeat (STR) profiling and tested for the absence of mycoplasma contamination (MycoAlert; Lonza). Both cell lines were maintained in Dulbecco’s modified Eagle’s medium (DMEM; Gibco) medium with 10% FBS at 37 °C in 5% CO2.
Construction of reporter plasmids and transient transfections
To construct a vector expressing human CYP26B1 (Gene ID: 56603), full-length CYP26B1 cDNA containing the rs138478634[G] allele or rs138478634[A] allele was commercially synthesized (Genewiz) and subcloned into the BamHI and XhoI sites of the pcDNA3.1 vector (Invitrogen). The resulting vectors were named CYP26B1[G] and CYP26B1[A]. KYSE30 and KYSE150 cells were seeded at 1 × 106 cells per well in 6-well plates, and 5 μg of plasmid was co-transfected into cells using Lipofectamine 3000 (Invitrogen). After 24 h, the growth medium was exchanged 2 h before incubation with 500 μg all-trans retinoid acid (Sigma-Aldrich). Cells were collected and frozen at –80 °C after 18 h incubation.
The transfection efficiency was detected by western blot and qRT-PCR (Supplementary Fig. 10). The antibodies against CYP26B1 (Cat# 21555-1-AP) and GAPDH (Cat# 10494-1-AP) were purchased from Proteintech. Both antibodies were validated by the company using western blot in human cell lines. Total RNA from cells was extracted with TRIzol (Invitrogen) according to the manufacturer’s instruction. First-strand cDNA was synthesized using the PrimeScript 1st Strand cDNA Synthesis Kit (Takara). Relative RNA expression levels determined by qRT-PCR were measured using the SYBR Green method on an ABI Prism 7900 sequence detection system (Applied Biosystems) with SYBR Premix Ex Taq (Takara), 50 ng cDNA and 1 μM gene-specific primers in a 25 μl reaction mixture. Reactions were performed with an initial 30-s denaturation step at 95 °C, followed by 40 cycles of 95 °C for 5 s and 60 °C for 30 s. Independent experiments were done in triplicate. GAPDH was employed as an internal control.
Detection of cellular atRA level
Cells were washed with PBS and frozen at –80 °C until analysis. The atRA content was separated and detected by reverse-phase high-performance liquid chromatography (HPLC), which was accomplished using the Luna C18 column (5 μm, 150 × 4.6 mm) (Phenomenex). The samples were eluted with acetonitrile/water/acetic acid (70:30:0.05, mobile phase) at a flow rate of 1 ml/min and a 25 °C column temperature using 20 μl sample size.
Detection of serum level of atRA
We measured serum levels of atRA in 394 individuals with the rs138478634-GG genotype and 10 individuals with the rs138478634-GA genotype using ultra-performance liquid chromatography–tandem mass spectrometry (UPLC-MS/MS). These individuals were cancer-free individuals recruited from a physical examination center in Wuhan; their demographic characteristics, including age and sex, are shown in Supplementary Table 12. UPLC-MS/MS was performed using the ACQUITY UPLC HSS T3 C18 column (1.8 µm, 2.1 × 100 mm, Waters), and the sample was eluted with acetonitrile/water/acetic acid (50:50:0.04, mobile phase) at a flow rate of 0.35 ml/min and 40 °C column temperature using 10 µl sample size. Information was collected with the Applied Biosystems API 4500 QTRAP LC/MS/MS and analyzed using Analyst version 1.6.1 software in MRM (multiple reaction monitoring) mode65,66. The RA was extracted from fresh human serum samples without freeze-thawing using the acetonitrile liquid–liquid extraction method67. We determined serum atRA levels in duplicate samples from each person. The data presented for each person is an average of duplicate detections with coefficient of variation (CV) of <20%.
siRNA oligonucleotides targeting CYP26B1 and non-targeting siRNA (Supplementary Table 13) were purchased from RiboBio. Transfections with siRNA (100 nM) were performed with Lipofectamine 3000.
Analysis of cell proliferation
Cells were seeded in 96-well flat-bottomed plates, with each well containing 2,000 cells in 100 μl of cell suspension. After a certain time in culture, cell viability was measured using CCK-8 assays (Dojindo Laboratories). Each experiment with six replicates was repeated three times.
For the functional analyses, results presented means ± s.e.m. or means ± s.d.; the mean values between two groups were compared using two-sided unpaired Student’s t-test. All statistical analyses for the functional analyses were performed using R (3.3.0). For the association analyses, we conducted logistic regression model adjusted sex, age, smoking status, drinking status and the first three principal components in the discovery stage and sex, age, smoking status, drinking status in the replication stages or combined samples. The association analyses were performed using PLINK (1.90b) and R (3.3.0).
Life Sciences Reporting Summary
Further information on experimental design is available in the Life Sciences Reporting Summary.
The data that support the findings of this study have been deposit in the Genome Sequence Submission (Gsub) of Beijing Institute of Genomics, Chinese Academy of Sciences (http://bigd.big.ac.cn/gsub/) with accession number PRJCA000453.
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was supported by the National Key Research and Development Plan Program (2016YFC1302702 to X.M., 2016YFC1302701 to C.W. and 2016YFC1302703 to R.Z.); the National Program for Support of Top-notch Young Professionals, National Natural Science Foundation of China (81171878, 81222038 to X.M.); the Fok Ying Tung Foundation for Young Teachers in the Higher Education Institutions of China (131038 to X.M.); and the Program for HUST Academic Frontier Youth Team (to X.M.).