针对代表性不足群体的基因组预测迁移学习
内容总结：
谷歌研究揭示：跨种群遗传风险预测中“大数据并非越大越好”
谷歌研究院软件工程师Joey Poomarin Phloyphisut和高级软件工程师Cory McLean于2026年9月3日发布了一项关于改进跨种群遗传风险预测方法的研究成果。该研究系统评估了利用欧洲人群数据辅助其他种群遗传风险预测的效果，发现了一个出人意料的规律：当目标种群样本量较小时，借助欧洲人群数据确实能提升预测准确性，但随着目标种群自身样本量的增加，这种外部数据的辅助作用反而会降低预测精度。
研究背景：遗传风险评分面临“欧洲中心化”困境
多基因风险评分（PRS）通过分析数百万个遗传变异来预测疾病风险，但目前该技术在临床决策中的应用有限。主要原因在于，历史上的全基因组关联研究（GWAS） overwhelmingly 以欧洲人群为对象，导致所建立的预测模型在应用于非欧洲人群时准确性严重下降。这种差异源于不同种群间遗传结构、群体结构和变异等位基因频率的不同。
由于在数十万人中开展全新的GWAS研究对大多数医疗系统而言成本过高，研究人员探索了一种替代方案：利用现有的大规模欧洲人群GWAS数据，并结合目标种群的小规模GWAS数据进行“迁移学习”。
研究方法：基于英日两大生物库的系统对比
该研究利用英国生物库（UKB，含数十万欧洲人群样本）和日本生物库（BBJ，含近20万日本人群样本）两套大规模基因型和表型数据，系统评估了跨种群PRS预测性能。研究人员选取了八个临床相关性状进行评估：体质指数（BMI）、收缩压、舒张压、红细胞计数、白细胞计数、高密度脂蛋白胆固醇（HDL）、低密度脂蛋白胆固醇（LDL）和血糖。
实验采用三种方法进行对比分析，所有模型均在相同的BBJ留出样本集上评估表现。当目标样本量极小时（如5000例），将UKB欧洲数据纳入训练可提供显著的统计学增益；但令人意外的是，随着目标种群样本量增至15000例以上，仅使用目标种群自身数据的模型表现反而超过了合并外部数据的模型。
关键发现：性状差异决定数据融合的“临界点”
这一令人惊讶的发现适用于所有被检测的性状，但“拐点”出现的位置因性状的遗传特征而异。研究人员通过计算跨种群的遗传相关性来衡量“共享遗传”的程度。对于遗传保守性状（如BMI），即两个种群间遗传相关性较高的性状，借助UKB欧洲数据训练的优势可以维持到更大的目标样本量（25000至40000例以上）；相比之下，脂质水平（HDL、LDL）和血糖等高度种群特异性状，其最佳UKB样本量远小于最大值，且这些性状从UKB数据中获益有限，因为其遗传架构与欧洲人群差异更大。
先进方法评估：元分析与PRS-CSx各有千秋
为进一步探索种群特异性遗传变异的捕捉，研究人员还测试了两种进阶方法：跨种群荟萃分析和PRS-CSx多祖先模型。结果发现，对于遗传保守性状，荟萃分析的影响较小，主要受限于BBJ样本量较小时统计效力不足；但对于HDL、LDL和血糖等种群特异性状，荟萃分析明显优于单一种群的分析方法。而PRS-CSx虽然理论上对不同种群特异性的适应性更强，但需要较大的样本量才能发挥优势——在目标样本量低于25000例时表现逊于弹性网络模型，而当样本量接近10万例时则达到或超越了最佳模型。
研究结论与启示
该研究系统评估表明，在将多基因风险评分应用于代表性不足的祖先种群时，更大的外部种群数据集并非always有益。研究最终强调，要在不同人群中实现最优预测性能，既需要扩大本地多样化生物库的建设，也需要根据目标性状的遗传特性和样本量谨慎选择建模策略。这一发现为精准医学指导下的跨种群遗传研究提供了重要参考依据。
该研究致谢了日本生物库及其合作机构RIKEN和东京大学医学科学研究所的支持。
中文翻译：
2026年9月3日
Joey Poomarin Phloyphisut，软件工程师，以及 Cory McLean，谷歌研究院资深员工软件工程师
我们评估了改善跨群体遗传风险预测的方法，并发现虽然从欧洲人群队列进行迁移学习能改善小型代表性不足群体的预测效果，但一旦目标队列样本量增大，这种方法反而会降低预测精度，尤其是对于具有群体特异性遗传结构的性状。
多基因风险评分（PRS）用于根据遗传变异预测疾病风险，通常涵盖数百到数百万个遗传变异的影响。然而，目前其在临床决策中的应用率较低，部分原因是历史上的全基因组关联研究（GWAS）绝大多数以欧洲人群为对象，导致将其应用于非欧洲人群时准确率严重下降。这些准确性差异源于跨群体在遗传结构、群体结构和变异等位基因频率方面的差异。
此外，在数十万人中开展全新的GWAS研究，对于大多数医疗系统而言成本过高。从现有的以欧洲为中心的GWAS进行迁移学习，并将大规模结果与目标群体特异性的GWAS数据相结合，提供了一种潜在的解决方案。
为此，在本篇博客文章中，我们描述了一项研究，该研究在目标非欧洲人群中评估PRS性能，同时变化目标人群和欧洲人群的样本量——这些欧洲人群数据用于构建涵盖八项临床性状的预测模型。具体而言，我们评估了在英国生物银行（UKB）中从数十万欧洲个体获得的PRS，迁移至日本生物银行（BBJ）样本的可迁移性——BBJ是一个包含近20万日本个体、具有深度表型数据的队列。我们的主要目标是提供系统性的、基于实证的指南，说明应如何开展跨群体GWAS和PRS模型训练，以优化目标人群中的预测性能。
两个大规模、深度基因分型和表型数据集（UKB和BBJ）的存在，使得我们可以通过数据集消融实验，系统性地评估PRS性能随样本量的变化。我们选择了在两个人群中均有测量的八项临床相关性状进行评估：体质指数（BMI）、收缩压、舒张压、红细胞计数、白细胞计数、高密度脂蛋白胆固醇（HDL）、低密度脂蛋白胆固醇（LDL）和血糖。在UKB中，测量遗传变异所解释的性状方差比例（单核苷酸多态性遗传度）范围为0.07至0.28。
我们使用了三种方法来评估PRS性能的可迁移性：
在所有实验中，PRS模型均在相同的保留BBJ样本集上进行评估。
对于每项实验，我们使用皮尔逊相关系数量化模型性能。正如预期，当目标人群数据有限或不存在时，欧洲人群的发现数据提供了有用的基线。然而，当目标人群样本量达到1.5万或以上时，目标人群特异性的模型表现更优，因为与外部欧洲数据共同训练似乎限制了通过更大样本量所获得的目标人群精度提升。
这一令人惊讶的观察结果在我们检查的所有表型中均成立。当目标样本量极小时（例如5,000个样本），在训练期间合并欧洲UKB数据能提供宝贵的统计增益。随着用于训练PRS模型的目标样本量增加，跨人群合并数据的收益逐渐减少。
虽然这一趋势在所有表型中均一致，但欧洲数据使用效益开始递减的交叉点高度依赖于所研究的具体性状。我们可以使用同一性状在不同人群间的遗传相关性来量化“共享遗传”的程度。我们观察到，“保守”性状，即两个人群间遗传相关性较高的性状，在目标人群特异性训练数据达到同等水平之前，能保留合并UKB欧洲训练数据的收益直到更大的目标样本量（25,000-40,000+个样本）。
与此形成鲜明对比的是，血脂水平（HDL、LDL）和血糖表现出小得多的BBJ样本量阈值——超过该阈值后，最优UKB样本量小于最大值，且最优样本量也更小。这些高度群体特异性的性状从UKB数据中获益较少，因为这些数据与目标人群的分布差异更大。
上述实验将PRS模型的输入变异限制为在UKB欧洲人群中发现的变异，排除了BBJ样本独有的任何性状相关变异。为了扩展分析以捕捉这些变异，我们另外创建了两种预测方法。
首先，我们对每个BBJ样本量进行了GWAS分析。第一种新方法使用完整的UKB GWAS数据和特定样本量的BBJ GWAS数据进行跨人群荟萃分析以识别候选变异，然后在这些变异上拟合弹性网络模型。第二种方法使用PRS-CSx来合并两组GWAS汇总统计量。
通过追踪荟萃分析和PRS-CSx在不同发现样本量下的净性能增益，我们观察到不同BBJ样本量下的性能差异。
对于保守性状，荟萃分析的影响较小，主要原因是样本量小得多的BBJ GWAS统计功效降低。然而，对于群体特异性性状，如HDL和LDL，以及程度稍低的血糖，荟萃分析的性能明显优于单人群发现方法。这些增益主要归因于修改了弹性网络的变异输入：当使用10,000个或更少的BBJ样本时，在训练中纳入UKB欧洲样本可略微改善预测效果。但在更大的BBJ样本量下，这种改善并未出现。
由于PRS-CSx动态加权群体特异性模型，其性能在理论上对保守性状与群体特异性性状之间的差异不太敏感。然而，我们观察到该模型比弹性网络模型需要更多数据才能表现良好。当目标样本量低于25,000时，除BMI外，PRS-CSx在所有表型上的表现均不如最强的相应弹性网络模型。当样本量接近100,000时，PRS-CSx在所有表型上达到或超过了最佳模型的表现，血糖除外。
对跨群体基因组预测的系统性评估揭示，在将多基因风险评分（PRS）应用于代表性不足的族群时，更大的外部人群数据集并不总是有益的。具体而言，虽然从英国生物银行的大型欧洲人群队列进行迁移学习在目标人群样本量较小（日本生物银行中低于15,000个样本）时提供了统计增益，但随着BBJ样本量的增长，它实际上反而降低了预测精度。这一性能交叉点因性状而异：遗传上保守的性状（如BMI）在更大的样本量下仍能保留外部数据合并的收益，而群体特异性性状（如血脂和血糖）在较少样本量时收益就已减弱。重要的是，像PRS-CSx这样的先进多族裔方法需要大量目标人群样本才能超越简单方法，而跨人群荟萃分析在较小样本量下对群体特异性性状提供了稳健的收益。最终，这些发现强调，优化不同人群中的预测性能既需要扩大本地多样化生物银行的建设，也需要根据性状遗传度和样本量精心选择建模策略。
我们衷心感谢日本生物银行以及我们在日本理化学研究所和东京大学医学科学研究所的合作者使这项研究得以实现，同时感谢其他谷歌合作者：Babak Behsaz、Andrew Carroll、Farhad Hormozdiari和Taedong Yun。我们还要感谢Hiroki Kayama和Joe Ledsam的机构支持，以及Michael Brenner和Katherine Chou的领导支持。
英文来源：
September 3, 2026
Joey Poomarin Phloyphisut, Software Engineer, and Cory McLean, Senior Staff Software Engineer, Google Research
We evaluate ways to improve cross-population genetic risk prediction and find that while transfer learning from European cohorts improves prediction in small underrepresented populations, it degrades accuracy once target cohort sample sizes grow, especially for traits with population-specific genetic architectures.
Polygenic risk scores (PRSs) are used to predict disease risk from genetic variants. They usually incorporate the influence of hundreds to millions of genetic variants. However, their adoption for clinical decision making is currently low, partly because historical genome-wide association studies (GWASs) have overwhelmingly evaluated European cohorts, resulting in severe accuracy drops when applied to non-European populations. These accuracy differences arise because of cross-population differences in genetic architectures, population structure, and variant allele frequencies.
Moreover, conducting de novo GWASs across hundreds of thousands of individuals is cost-prohibitive for most healthcare systems. Transfer learning from existing European-centric GWAS, augmenting these large-scale results with target-population-specific GWAS, provides a potential solution.
To that end, in this blog post we describe a study evaluating PRS performance in a target non-European population while varying the size of both the target population and European populations used to create the predictive model across eight clinical traits. Specifically, we evaluate transferability of PRS ascertained in hundreds of thousands of European individuals within the UK Biobank (UKB) to samples within Biobank Japan (BBJ), a deeply-phenotyped cohort of nearly 200 thousand Japanese individuals. Our primary objective is to provide systematic, empirical guidelines on how cross-population GWAS and PRS model training should be performed to optimize predictive performance in a target population.
The presence of two large-scale, deeply genotyped and phenotyped datasets (UKB and BBJ) enables dataset ablation experiments to systematically evaluate PRS performance as a function of sample size. We selected eight clinically relevant traits measured in both populations for evaluation: body mass index (BMI), systolic blood pressure, diastolic blood pressure, red blood cell count, white blood cell count, high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), and blood glucose. In UKB, the estimated fraction of trait variance explained by the measured genetic variants (single nucleotide polymorphism heritability) ranged from 0.07–0.28.
Three methods were used to evaluate the PRS performance transferability:
In all experiments, the PRS models were evaluated on the same held-out set of BBJ samples.
For each experiment, we quantified model performance using Pearson correlation. As expected, European discovery data provides a helpful baseline when target population data is limited or non-existent. However, the target-population-specific models outperform at sample sizes of 15k or more, as it appears that co-training with external European data limits the target population accuracy gains achieved through higher sampling.
This surprising observation held across all phenotypes we examined. When the target sample size is extremely small (e.g., 5,000 samples), pooling European UKB data during training provides a valuable statistical boost. As the target sample size used to train the PRS model increases, the benefit of out-of-population pooling diminishes.
While the trend holds across phenotypes, the crossover point at which the success of using European data diminishes depends heavily on the specific trait being studied. We can quantify the extent of “shared genetics” using genetic correlation of the same trait across populations. We observe that “conserved” traits, or those with higher genetic correlations between the two populations, retain the benefits of pooling UKB European training data up to much larger target sample sizes (25-40k+ samples) before target-population-specific training data reaches parity.
In sharp contrast, lipid levels (HDL, LDL) and blood glucose exhibit much smaller BBJ sample sizes beyond which optimal UKB sample sizes are smaller than maximum, and that optimal sample size is also smaller. These highly population-specific traits do not benefit as much from UKB data, since that data is farther out-of-distribution.
The above experiments restricted PRS model input variants to those discovered in the UKB European population, excluding any trait-associated variants unique to BBJ samples. To extend the analyses to capture these, we created two additional prediction methods.
First, we ran GWAS on each BBJ sample size. The first new method performed a cross-population meta-analysis using the full UKB GWAS and the sample-size-specific BBJ GWAS to identify candidate variants, and then fit an elastic net on those variants. The second method used PRS-CSx to combine the two sets of GWAS summary statistics.
By tracking the net performance gain of meta-analysis and PRS-CSx across varying discovery sample sizes, we observed differences in performance across BBJ sample sizes.
The influence of meta-analysis is low for conserved traits, largely due to reduced statistical power in the much smaller BBJ GWAS sample sizes. However, for population-specific traits like HDL and LDL, and to a lesser extent blood glucose, meta-analysis substantially outperforms single-population discovery. The gains are primarily due to modifying the elastic net variant input: including UKB European samples during training slightly improved prediction when using 10,000 or fewer BBJ samples. This improvement was not seen with larger BBJ samples.
Because PRS-CSx dynamically weights population-specific models, its performance is theoretically less sensitive to conserved vs population-specific traits. However, we observed that the model requires more data than elastic net models to perform well. For target sample sizes under 25k, PRS-CSx performs worse than the strongest corresponding elastic net model in all phenotypes except BMI. As sample sizes approached 100k, PRS-CSx matched or exceeded the best performing model across all phenotypes except blood glucose.
Systematic evaluation of cross-population genomic prediction reveals that larger out-of-population datasets are not always beneficial when applying polygenic risk scores (PRSs) to underrepresented ancestries. Specifically, while transfer learning from the large European cohort of UK Biobank provided a statistical boost at low target population sizes (under 15,000 samples in Biobank Japan), it actually degraded predictive accuracy as the BBJ sample size grew. This performance crossover is trait-dependent: genetically conserved traits (such as BMI) retain the benefits of external data pooling up to larger sample sizes, whereas population-specific traits (such as lipids and blood glucose) show diminished benefits with fewer samples. Importantly, advanced multi-ancestry methods like PRS-CSx require substantial target-population samples to outperform simpler approaches, while cross-population meta-analysis offers robust benefits for population-specific traits at smaller sizes. Ultimately, these findings emphasize that optimizing predictive performance in diverse populations will require both the expansion of local, diverse biobanks and careful selection of modeling strategy that is tailored to both trait heritability and sample size.
We sincerely thank Biobank Japan and our collaborators at RIKEN and The Institute of Medical Science The University of Tokyo for enabling this research, as well as additional Google collaborators: Babak Behsaz, Andrew Carroll, Farhad Hormozdiari, and Taedong Yun. Our thanks also go to Hiroki Kayama and Joe Ledsam for institutional support, and Michael Brenner and Katherine Chou for their leadership support.