使用我们的多项式荟萃分析，我们确定了所有表型中的2,143个相关基因座（前哨变体P< 5 × 10−9), with 3,823 independently associated variants (Extended Data Fig. 2, Supplementary Tables 2 and 3 and Supplementary Figs. 2 and 3). Of these, 1,346 loci and 2,486 independent variants were associated with SmkInit, 33 loci (39 variants) with AgeSmk, 140 loci (243 variants) with CigDay, 128 loci (206 variants) with SmkCes and 496 loci (849 variants) with DrnkWk. Approximately 64% (n = 1,364) of loci were phenotype-specific, five loci were associated with all four smoking phenotypes but not with DrnkWk, and five loci were associated with all five phenotypes. All sentinel variants within identified loci had high posterior probabilities that their effect would replicate in a sufficiently powered study according to a trans-ancestry extension of our GWAS cross-validation technique6. only 17 sentinel variants (0.7%) had such posterior probabilities of less than 0.99 and were therefore removed from the counts above and from further consideration (additional details on these 17 variants are shown in Supplementary Fig. 4).
Inclusion of diverse ancestry may improve the discovery of new variants through a combination of increased genetic variation, larger sample sizes and improved fine-mapping due to diverse patterns of linkage disequilibrium (LD). We quantified gains in power from the use of our multi-ancestry model over a simpler ancestry-naive fixed-effects model excluding the ancestry meta-regression. Comparing the number of associated variants, we found 721 additional independent variants that were identified only by the multi-ancestry meta-regression analysis. Both sets of models were fit to the same data, such that the larger number of significantly associated variants identified with the multi-ancestry model indicates increased power from accounting for axes of genetic variation and residual heterogeneity. Included among these 721 were newly associated variants in genes related to nervous system function (for example, NRXN1) including glutamatergic (GRIN2A) neurotransmission, which is of relevance to neurocircuitry in addiction7,8.
To isolate likely causal variants, we used a fine-mapping procedure (see Supplementary Note) that leverages variation in LD across ancestry groups to construct 90% credible intervals. We identified 597 loci (27.9%) in which the 90% credible intervals included fewer than five variants, including 192 loci (9.0%) with a single fine-mapped variant. Overall, credible intervals contained medians of 9–19 variants and median spans of 32–78 kb across phenotypes (Supplementary Table 4). Compared with the EUR-stratified GWAS (described in the next section), the trans-ancestry fine-mapping increased the number of 90% credible intervals containing fewer than five variants by 27.6%, and containing a single variant by 41.2%. Across all 2,143 loci, 1,330 (62.1%) loci had a reduced number of variants in the credible intervals in the multi-ancestry analysis. To determine the gain in resolution attributable to increased sample size (versus LD differences), we ‘downsampled’ the multi-ancestry analysis by removing EUR ancestry cohorts until the total sample size was approximately equal to that of the EUR-stratified analysis and regenerated fine-mapping results. Using the 1,330 loci with improved resolution in multi-ancestry analysis, we found that the credible intervals were reduced from a median of 22 variants in the EUR-stratified analysis to 12 variants in the downsampled multi-ancestry analysis, suggesting that approximately 55% of the observed improvement in fine-mapping is attributable to larger multi-ancestry sample sizes alone. These findings highlight the utility of both increased sample size and diverse ancestry in fine-mapping variants for these complex behavioural phenotypes. To characterize genes prioritized from fine-mapping, we conducted a series of functional enrichment analyses. We first selected intervals fine-mapped to fewer than five variants from the multi-ancestry results and mapped each variant to the nearest gene to identify ‘high-priority’ genes. Relative to genes mapped from variants with posterior inclusion probabilities (PIP) < 0.01, the high-priority genes were enriched across brain and nerve tissues (Extended Data Fig. 3a and Supplementary Table 5). Within the brain, cell-type enrichment of the high-priority genes was observed for projecting glutamatergic neurons from the cortex, hippocampus and amygdala (telencephalon excitatory projection neurons) and projection GABA neurons from medium spiny neurons of the striatum (telencephalon inhibitory projecting neurons), along with neurons in various subcortical structures such as the hypothalamus and midbrain, consistent with aspects of the mesolimbic theory of addiction7,8 (Extended Data Fig. 3b). Finally, these high-priority genes that were strongly associated with substance use were enriched in gene pathways related to neurogenesis, neuronal development, neuronal differentiation and synaptic function. The neurodevelopmental aspect of the high-priority genes could indicate a role for these genes in processes that predispose individuals to risk of substance use and/or may contribute to brain circuit rewiring during drug use.
The multi-ancestry meta-analysis method also allowed for tests of whether a variant effect size differed (that is, was moderated) by ancestry along four ancestry dimensions estimated from multidimensional scaling (MDS) of allele frequencies from each participating study (Fig. 1a). Roughly, the first axis represents an EAS ancestry cline, the second axis an AFR cline, the third a EUR cline (north to south EUR) and the fourth an AMR cline. There was minimal evidence of effect size moderation by ancestry for most independent variants, ranging from 76.6% (187 variants) in CigDay to 85.0% (175 variants) in SmkCes. Another 7.7–18.1% showed modest evidence for moderation. Finally, roughly 3.6% of all independent variants, reflecting 136 variants from 84 distinct loci, showed strong evidence of effect size moderated by ancestry (complete results are shown in Supplementary Table 2). Comparisons between the variants with strong evidence for effect size moderation by ancestry and those with no evidence suggested that the identification of these 136 variants was not driven to a large extent by differences in imputation quality, LD scores or Fst (fixation index) across ancestries.
Across phenotypes, 88 of these 136 variants showed moderation by the first axis of ancestry variation (approximate EAS cline; Fig. 1b, left), 29 variants by the second axis (approximate AFR cline; Fig. 1b, middle) and 10 variants by the fourth axis (approximate AMR cline; Fig. 1b, right). Nine variants showed differences in effect size moderated by the third axis (EUR cline). only the effect of one variant was moderated by three or more ancestry clines (EAS, AFR and AMR): rs1229984, a missense variant in the alcohol dehydrogenase gene ADH1B, which has been shown to be protective against alcohol consumption9. An increase on any of these clines was associated with a reduced effect size of this allele, on average. For example, if there are two people who both carry one copy of the protective T allele for this variant but are separated by 1 s.d. on MDS component 1 (EAS cline), the person with a lower value on that MDS cline would be expected to drink 0.3 fewer drinks than the person with a higher MDS value, despite the same rs1229984 genotype in ADH1B.
To further evaluate causal genes and relevant tissues through which associated variants may be operating, we applied a trans-ancestry transcriptome-wide association study (TWAS) analysis to each phenotype across 49 tissues derived from the GTEx Consortium10. Using a P value threshold Bonferroni-corrected for the total number of genes and tissues within a phenotype, we found 1,167 genes significantly associated with SmkInit, 21 genes with AgeSmk, 203 genes with CigDay, 188 genes with SmkCes and 504 genes with DrnkWk (resulting in 1,705 unique genes across phenotypes; Supplementary Table 6). For each of our five phenotypes, matrix decomposition parallel analysis11 of the per-tissue P value correlation matrix suggested two components: one explaining 53.7–55.2% of the variance in P values, and another explaining 3.5–3.8% of the variance in P values. Similar loading patterns were observed for all phenotypes such that all tissues loaded strongly (all loadings > 0.12) on the first component, suggesting that it represents a general effect across tissues, whereas only brain tissues had strong loadings on the second component (all loadings >0.12），表明脑特异性基因表达对这些烟草和酒精使用表型的重要性。对TWAS相关基因的途径富集分析确定了跨组织广泛富集的表型的1,029个独特的基因途径（补充表7），包括许多与神经传递和神经发育的明显相关性。
为了进一步说明感兴趣的基因中的几种变体，我们将上述的发现集成在一起，以选择跨分析方法有关联的变体，并且对这些变体的可用性显然相关。以与上述富集分析相似的方式选择说明性变体：（1）我们从含有少于五个变体的多型精细图表的可靠间隔中提取了变体，（2）我们将所得的变体与多种癌变的twas-cis-cis-cis-cis-cis-twas-cis-cis-cis-national national nortialational dentiticational dentiticational tartic tartic tartic tartialtials相关景点进行了交叉。我们强调了这一过程产生的52个基因中的五个。
我们发现，烟碱基因簇CHRNA5 – A3 – B4与Smkinit12显着相关，并以90％的可靠间隔进行精细映射，从欧元分层的53种变体缩小到多2个变体的多疗程结果中的两个变体（RS2869055和RS2869055和RS28438420; RS28438420;补充表4）。这些变体不在该基因簇中的众所周知的变体RS16969968中，具有高LD（对于两个变体的R2 = 0.31）。相比之下，该基因座被细化为高LD中的两个变体，CIGDAY的RS16969968（R2 = 0.84和0.86），这表明该信号为吸烟引发的信号所基于的变体可能与每天的香烟不同。我们还发现了SMKINIT和CACNA1B之间的新型关联，该钙通道（CAV2.2）控制着神经元神经递质释放，并与可卡因的再施加13相关，并增加了攻击性和警惕性，以及减速和疾病和探索14。CACNA1B与多种精神病疾病有关，包括精神分裂症，躁郁症和自闭症谱系障碍15,16,17。
CIGDay与神经蛋白酶（NRTN）的变体有关，这是一种涉及多巴胺神经元发展和存活的神经胶质细胞系衍生的神经营养因子18。该基因已与帕金森氏病有关，其潜力恢复了多巴胺神经记录19。同样，PAK6是另一个与TWAS结果中CIGDAY密切相关的新型基因，在90％可靠的间隔中仅绘制了三个变体。PAK6编码在纹状体和海马中高度表达的P21激活的激酶，与GABA能中神经元20的迁移以及多巴胺能神经转移的调节有关，并参与了大型运动和认知功能22。PAK6与精神分裂症23和神经退行性疾病有着牢固相关的24,25，例如帕金森氏病和阿尔茨海默氏病，进一步强调了其在突触变化中的作用。最后，我们发现了ECE2和Drnkwk之间的新型关联。ECE2参与了皮质发育26以及包括神经辛和物质P27在内的几种神经内分泌肽的加工，并且也可能在淀粉样蛋白-β加工中起作用28。ECE2还会产生诸如BAM 12（显示κ阿片受体选择性）和BAM 22（显示μ-阿片受体选择性）之类的肽，这表明与疼痛透射率有联系27。