Model training and validation/test set.

Genome-wide association studies (GWAS) have successfully uncovered numerous associations between genetic variants and disease traits to date. Yet, identifying significantly associated loci remains a considerable challenge due to the concomitant multiple-testing burden of performing such analyses genome-wide. Here, we leverage the genetic associations of molecular traits – DNA CpG-site methylation status and RNA expression – to mitigate this problem. We encode their co-association across the genome using PinSage, a graph convolutional neural network-based recommender system previously deployed at Pinterest. We demonstrate, using this framework, that a model trained only on methylation quantitative trait locus (QTL) data could recapitulate over half (554,209/1,021,052) of possible SNP-RNA associations identified in a large expression QTL meta-analysis. Taking advantage of a recent ‘saturated’ map of height associations, we then show that height-associated loci predicted by a model trained on molecular-QTL data replicated comparably, following Bonferroni correction, to those that were genome-wide significant in UK Biobank (88% compared to 91%). On a set of 64 disease outcomes in UK Biobank, the same model identified 143 independent novel disease associations, with at least one additional association for 64% (41/64) of the disease outcomes examined. Excluding associations involving the MHC region, we achieve a total uplift of over 8% (128/1,548). We successfully replicated 38% (39/103) of the novel disease associations in an independent sample, with suggestive evidence for six additional associations from GWAS Catalog. Replicated associations included for instance that between rs10774625 (nearest gene: SH2B3/ATXN2) and coeliac disease, and that between rs12350420 (nearest gene: MVB12B) and glaucoma. For many GWAS, attaining such an enhancement by simply increasing sample size may be prohibitively expensive, or impossible depending on disease prevalence.

全基因组关联研究（Genome-wide association studies, GWAS）迄今已成功揭示诸多遗传变异与疾病性状间的关联。然而，由于全基因组分析伴随的多重检验负担，鉴定出具有显著关联的基因座仍是一项颇具挑战的任务。为此，我们借助分子性状——即DNA CpG位点甲基化状态与RNA表达——的遗传关联来缓解这一难题。我们采用此前在Pinterest部署的、基于图卷积神经网络的推荐系统PinSage，对全基因组范围内二者的共关联关系进行编码。借助该框架，我们证实：仅基于甲基化数量性状基因座（methylation quantitative trait locus, QTL）数据训练的模型，可复现大型表达QTL元分析中鉴定出的超半数（554209/1021052）潜在单核苷酸多态性（Single Nucleotide Polymorphism, SNP）-RNA关联。借助近期发布的身高关联“饱和图谱”，我们进一步证明：基于分子QTL数据训练的模型所预测的身高关联基因座，经过Bonferroni校正（Bonferroni correction）后，其可重复性与英国生物银行（UK Biobank）中达到全基因组显著性的基因座相当（分别为88%与91%）。针对英国生物银行中的64种疾病结局，该模型鉴定出143个独立的全新疾病关联，其中64%（41/64）的被研究疾病结局至少存在1个此类新关联。若排除涉及主要组织相容性复合体（Major Histocompatibility Complex, MHC）区域的关联，我们整体的提升幅度可达8%以上（128/1548）。我们在独立样本中成功复现了38%（39/103）的全新疾病关联，另有6个关联在GWAS目录（GWAS Catalog）中显示出提示性证据。复现的关联例如包括rs10774625（最近基因：SH2B3/ATXN2）与乳糜泻（coeliac disease）之间的关联，以及rs12350420（最近基因：MVB12B）与青光眼（glaucoma）之间的关联。对于诸多GWAS而言，仅通过增加样本量来实现此类提升可能成本过高，或是根据疾病患病率而言根本无法实现。