XPXP: improving polygenic prediction by cross-population and cross-phenotype analysis

Methodological opportunities in genomic data analysis to advance health equity

The causes and consequences of inequities in genomic research and medicine are complex and widespread. However, it is widely acknowledged that underrepresentation of diverse populations in human genetics research risks exacerbating existing health disparities. Efforts to improve diversity are ongoing, but an often-overlooked source of inequity is the choice of analytical methods used to process, analyse and interpret genomic data. This choice can influence all areas of genomic research, from genome-wide association studies and polygenic score development to variant prioritization and functional genomics. New statistical and machine learning techniques to understand, quantify and correct for the impact of biases in genomic data are emerging within the wider genomic research and genomic medicine ecosystems. At this crucial time point, it is important to clarify where improvements in methods and practices can, or cannot, have a role in improving equity in genomics. Here, we review existing approaches to promote equity and fairness in statistical analysis for genomics, and propose future methodological developments that are likely to yield the most impact for equity.

JointPRS: A data-adaptive framework for multi-population genetic risk prediction incorporating genetic correlation

Genetic risk prediction for non-European populations is hindered by limited Genome-Wide Association Study (GWAS) sample sizes and small tuning datasets. We propose JointPRS, a data-adaptive framework that leverages genetic correlations across multiple populations using GWAS summary statistics. It achieves accurate predictions without individual-level tuning data and remains effective in the presence of a small tuning set thanks to its data-adaptive approach. Through extensive simulations and real data applications to 22 quantitative and four binary traits in five continental populations evaluated using the UK Biobank (UKBB) and All of Us (AoU), JointPRS consistently outperforms six state-of-the-art methods across three data scenarios: no tuning data, same-cohort tuning and testing, and cross-cohort tuning and testing. Notably, in the Admixed American population, JointPRS improves lipid trait prediction in AoU by 6.46%-172.00% compared to the other existing methods.

Unsupervised Ensemble Learning for Efficient Integration of Pre-trained Polygenic Risk Scores

The growing availability of pre-trained polygenic risk score (PRS) models has enabled their integration into real-world applications, reducing the need for extensive data labeling, training, and calibration. However, selecting the most suitable PRS model for a specific target population remains challenging, due to issues such as limited transferability, data heterogeneity, and the scarcity of observed phenotype in real-world settings. Ensemble learning offers a promising avenue to enhance the predictive accuracy of genetic risk assessments, but most existing methods often rely on observed phenotype data or additional genome-wide association studies (GWAS) from the target population to optimize ensemble weights, limiting their utility in real-time implementation. Here, we present the UN supervised en Semble PRS ( UNSemblePRS ), an unsupervised ensemble learning framework, that combines pre-trained PRS models without requiring phenotype data or summaries from the target population. Unlike traditional supervised approaches, UNSemblePRS aggregates models based on prediction concordance across a curated subset of candidate PRS models. We evaluated UNSemblePRS using both continuous and binary traits in the All of Us database, demonstrating its scalability and robust performance across diverse populations. These results underscore UNSemblePRS as an accessible tool for integrating PRS models into real-world contexts, offering broad applicability as the availability of PRS models continues to expand.

Advancements and limitations in polygenic risk score methods for genomic prediction: a scoping review

This scoping review aims to identify and evaluate the landscape of Polygenic Risk Score (PRS)-based methods for genomic prediction from 2013 to 2023, highlighting their advancements, key concepts, and existing gaps in knowledge, research, and technology. Over the past decade, various PRS-based methods have emerged, each employing different statistical frameworks aimed at enhancing prediction accuracy, processing speed and memory efficiency. Despite notable advancements, challenges persist, including unrealistic assumptions regarding sample sizes and the polygenicity of traits necessary for accurate predictions, as well as limitations in exploring hyper-parameter spaces and considering environmental interactions. We included studies focusing on PRS-based methods for risk prediction that underwent methodological evaluations using valid approaches and released computational tools/software. Additionally, we restricted our selection to studies involving human participants that were published in English language. This review followed the standard protocol recommended by Joanna Briggs Institute Reviewer's Manual, systematically searching Ovid MEDLINE, Ovid Embase, Scopus and Web of Science databases. Additionally, searches included grey literature sources like pre-print servers such as bioRxiv, and articles recommended by experts to ensure comprehensive and diverse coverage of relevant records. This study identified 34 studies detailing 37 genomic prediction methods, the majority of which rely on linkage disequilibrium (LD) information and necessitate hyper-parameter tuning. Nine methods integrate functional/gene annotation, while 12 are suitable for cross-ancestry genomic prediction, with only one considering gene-environment (GxE) interaction. While some methods require individual-level data, most leverage summary statistics, offering flexibility. Despite progress, challenges remain. These include computational complexity and the need for large sample sizes for high prediction accuracy. Furthermore, recent methods exhibit varying effectiveness across traits, with absolute accuracies often falling short of clinical utility. Transferability across ancestries varies, influenced by trait heritability and diversity of training data, while handling admixed populations remains challenging. Additionally, the absence of standard error measurements for individual PRSs, crucial in clinical settings, underscores a critical gap. Another issue is the lack of customizable graphical visualization tools among current software packages. While genomic prediction methods have advanced significantly, there is still room for improvement. Addressing current challenges and embracing future research directions will lead to the development of more universally applicable, robust, and clinically relevant genomic prediction tools.

Polygenic prediction for underrepresented populations through transfer learning by utilizing genetic similarity shared with European populations

Because current genome-wide association studies are primarily conducted in individuals of European ancestry and information disparities exist among different populations, the polygenic score derived from Europeans thus exhibits poor transferability. Borrowing the idea of transfer learning, which enables the utilization of knowledge acquired from auxiliary samples to enhance learning capability in target samples, we propose transPGS, a novel polygenic score method, for genetic prediction in underrepresented populations by leveraging genetic similarity shared between the European and non-European populations while explaining the trans-ethnic difference in linkage disequilibrium (LD) and effect sizes. We demonstrate the usefulness and robustness of transPGS in elevated prediction accuracy via individual-level and summary-level simulations and apply it to seven continuous phenotypes and three diseases in the African, Chinese, and East Asian populations of the UK Biobank and Genetic Epidemiology Research Study on Adult Health and Aging cohorts. We further reveal that distinct LD and minor allele frequency patterns across ancestral groups are responsible for the dissatisfactory portability of PGS.

JointPRS: A Data-Adaptive Framework for Multi-Population Genetic Risk Prediction Incorporating Genetic Correlation

Genetic prediction accuracy for non-European populations is hindered by the limited sample size of Genome-wide association studies (GWAS) data in these populations. Additionally, it is challenging to tune model parameters with a small tuning dataset for methods that require tuning data, which is often the case for non-European samples. To address these challenges, we propose JointPRS, a novel, data-adaptive framework that simultaneously models multiple populations using GWAS summary statistics. JointPRS incorporates genetic correlation structures into the prediction framework, enabling accurate performance even without individual-level tuning data. Additionally, it uniquely employs a data-adaptive approach, providing a robust solution when only a small tuning dataset is available. Through extensive simulations and real data applications to 22 quantitative traits and four binary traits in five continental populations (European (EUR); East Asian (EAS); African (AFR); South Asian (SAS); and Admixed American (AMR)) evaluated using the UK Biobank (UKBB) and All of Us (AoU), we demonstrate that JointPRS outperforms six other state-of-art methods across three different data scenarios (no tuning data, tuning and testing data from the same cohort, and tuning and testing data from different cohorts) for most traits in non-European populations, while maintaining model simplicity and computational efficiency.

Optimizing clinico-genomic disease prediction across ancestries: a machine learning strategy with Pareto improvement

BACKGROUND

Accurate prediction of an individual's predisposition to diseases is vital for preventive medicine and early intervention. Various statistical and machine learning models have been developed for disease prediction using clinico-genomic data. However, the accuracy of clinico-genomic prediction of diseases may vary significantly across ancestry groups due to their unequal representation in clinical genomic datasets.

METHODS

We introduced a deep transfer learning approach to improve the performance of clinico-genomic prediction models for data-disadvantaged ancestry groups. We conducted machine learning experiments on multi-ancestral genomic datasets of lung cancer, prostate cancer, and Alzheimer's disease, as well as on synthetic datasets with built-in data inequality and distribution shifts across ancestry groups.

RESULTS

Deep transfer learning significantly improved disease prediction accuracy for data-disadvantaged populations in our multi-ancestral machine learning experiments. In contrast, transfer learning based on linear frameworks did not achieve comparable improvements for these data-disadvantaged populations.

CONCLUSIONS

This study shows that deep transfer learning can enhance fairness in multi-ancestral machine learning by improving prediction accuracy for data-disadvantaged populations without compromising prediction accuracy for other populations, thus providing a Pareto improvement towards equitable clinico-genomic prediction of diseases.

Evaluating and improving health equity and fairness of polygenic scores

Polygenic scores (PGSs) are quantitative metrics for predicting phenotypic values, such as human height or disease status. Some PGS methods require only summary statistics of a relevant genome-wide association study (GWAS) for their score. One such method is Lassosum, which inherits the model selection advantages of Lasso to select a meaningful subset of the GWAS single-nucleotide polymorphisms as predictors from their association statistics. However, even efficient scores like Lassosum, when derived from European-based GWASs, are poor predictors of phenotype for subjects of non-European ancestry; that is, they have limited portability to other ancestries. To increase the portability of Lassosum, when GWAS information and estimates of linkage disequilibrium are available for both ancestries, we propose Joint-Lassosum (JLS). In the simulation settings we explore, JLS provides more accurate PGSs compared to other methods, especially when measured in terms of fairness. In analyses of UK Biobank data, JLS was computationally more efficient but slightly less accurate than a Bayesian comparator, SDPRX. Like all PGS methods, JLS requires selection of predictors, which are determined by data-driven tuning parameters. We describe a new approach to selecting tuning parameters and note its relevance for model selection for any PGS. We also draw connections to the literature on algorithmic fairness and discuss how JLS can help mitigate fairness-related harms that might result from the use of PGSs in clinical settings. While no PGS method is likely to be universally portable, due to the diversity of human populations and unequal information content of GWASs for different ancestries, JLS is an effective approach for enhancing portability and reducing predictive bias.

Improving polygenic risk prediction in admixed populations by explicitly modeling ancestral-differential effects via GAUDI

Polygenic risk scores (PRS) have shown successes in clinics, but most PRS methods focus only on participants with distinct primary continental ancestry without accommodating recently-admixed individuals with mosaic continental ancestry backgrounds for different segments of their genomes. Here, we develop GAUDI, a novel penalized-regression-based method specifically designed for admixed individuals. GAUDI explicitly models ancestry-differential effects while borrowing information across segments with shared ancestry in admixed genomes. We demonstrate marked advantages of GAUDI over other methods through comprehensive simulation and real data analyses for traits with associated variants exhibiting ancestral-differential effects. Leveraging data from the Women's Health Initiative study, we show that GAUDI improves PRS prediction of white blood cell count and C-reactive protein in African Americans by > 64% compared to alternative methods, and even outperforms PRS-CSx with large European GWAS for some scenarios. We believe GAUDI will be a valuable tool to mitigate disparities in PRS performance in admixed individuals.

Principles and methods for transferring polygenic risk scores across global populations

Polygenic risk scores (PRSs) summarize the genetic predisposition of a complex human trait or disease and may become a valuable tool for advancing precision medicine. However, PRSs that are developed in populations of predominantly European genetic ancestries can increase health disparities due to poor predictive performance in individuals of diverse and complex genetic ancestries. We describe genetic and modifiable risk factors that limit the transferability of PRSs across populations and review the strengths and weaknesses of existing PRS construction methods for diverse ancestries. Developing PRSs that benefit global populations in research and clinical settings provides an opportunity for innovation and is essential for health equity.

XMAP: Cross-population fine-mapping by leveraging genetic diversity and accounting for confounding bias

Fine-mapping prioritizes risk variants identified by genome-wide association studies (GWASs), serving as a critical step to uncover biological mechanisms underlying complex traits. However, several major challenges still remain for existing fine-mapping methods. First, the strong linkage disequilibrium among variants can limit the statistical power and resolution of fine-mapping. Second, it is computationally expensive to simultaneously search for multiple causal variants. Third, the confounding bias hidden in GWAS summary statistics can produce spurious signals. To address these challenges, we develop a statistical method for cross-population fine-mapping (XMAP) by leveraging genetic diversity and accounting for confounding bias. By using cross-population GWAS summary statistics from global biobanks and genomic consortia, we show that XMAP can achieve greater statistical power, better control of false positive rate, and substantially higher computational efficiency for identifying multiple causal signals, compared to existing methods. Importantly, we show that the output of XMAP can be integrated with single-cell datasets, which greatly improves the interpretation of putative causal variants in their cellular context at single-cell resolution.

mtPGS: Leverage multiple correlated traits for accurate polygenic score construction

Accurate polygenic scores (PGSs) facilitate the genetic prediction of complex traits and aid in the development of personalized medicine. Here, we develop a statistical method called multi-trait assisted PGS (mtPGS), which can construct accurate PGSs for a target trait of interest by leveraging multiple traits relevant to the target trait. Specifically, mtPGS borrows SNP effect size similarity information between the target trait and its relevant traits to improve the effect size estimation on the target trait, thus achieving accurate PGSs. In the process, mtPGS flexibly models the shared genetic architecture between the target and the relevant traits to achieve robust performance, while explicitly accounting for the environmental covariance among them to accommodate different study designs with various sample overlap patterns. In addition, mtPGS uses only summary statistics as input and relies on a deterministic algorithm with several algebraic techniques for scalable computation. We evaluate the performance of mtPGS through comprehensive simulations and applications to 25 traits in the UK Biobank, where in the real data mtPGS achieves an average of 0.90%-52.91% accuracy gain compared to the state-of-the-art PGS methods. Overall, mtPGS represents an accurate, fast, and robust solution for PGS construction in biobank-scale datasets.

A new method for multiancestry polygenic prediction improves performance across diverse populations

Polygenic risk scores (PRSs) increasingly predict complex traits; however, suboptimal performance in non-European populations raise concerns about clinical applications and health inequities. We developed CT-SLEB, a powerful and scalable method to calculate PRSs, using ancestry-specific genome-wide association study summary statistics from multiancestry training samples, integrating clumping and thresholding, empirical Bayes and superlearning. We evaluated CT-SLEB and nine alternative methods with large-scale simulated genome-wide association studies (~19 million common variants) and datasets from 23andMe, Inc., the Global Lipids Genetics Consortium, All of Us and UK Biobank, involving 5.1 million individuals of diverse ancestry, with 1.18 million individuals from four non-European populations across 13 complex traits. Results demonstrated that CT-SLEB significantly improves PRS performance in non-European populations compared with simple alternatives, with comparable or superior performance to a recent, computationally intensive method. Moreover, our simulation studies offered insights into sample size requirements and SNP density effects on multiancestry risk prediction.

Evaluating and Improving Health Equity and Fairness of Polygenic Scores

Polygenic scores (PGS) are quantitative metrics for predicting phenotypic values, such as human height or disease status. Some PGS methods require only summary statistics of a relevant genome-wide association study (GWAS) for their score. One such method is Lassosum, which inherits the model selection advantages of Lasso to select a meaningful subset of the GWAS single nucleotide polymorphisms as predictors from their association statistics. However, even efficient scores like Lassosum, when derived from European-based GWAS, are poor predictors of phenotype for subjects of non-European ancestry; that is, they have limited portability to other ancestries. To increase the portability of Lassosum, when GWAS information and estimates of linkage disequilibrium are available for both ancestries, we propose Joint-Lassosum. In the simulation settings we explore, Joint-Lassosum provides more accurate PGS compared with other methods, especially when measured in terms of fairness. Like all PGS methods, Joint-Lassosum requires selection of predictors, which are determined by data-driven tuning parameters. We describe a new approach to selecting tuning parameters and note its relevance for model selection for any PGS. We also draw connections to the literature on algorithmic fairness and discuss how Joint-Lassosum can help mitigate fairness-related harms that might result from the use of PGS scores in clinical settings. While no PGS method is likely to be universally portable, due to the diversity of human populations and unequal information content of GWAS for different ancestries, Joint-Lassosum is an effective approach for enhancing portability and reducing predictive bias.

Addressing the Challenge of Biomedical Data Inequality: An Artificial Intelligence Perspective

Artificial intelligence (AI) and other data-driven technologies hold great promise to transform healthcare and confer the predictive power essential to precision medicine. However, the existing biomedical data, which are a vital resource and foundation for developing medical AI models, do not reflect the diversity of the human population. The low representation in biomedical data has become a significant health risk for non-European populations, and the growing application of AI opens a new pathway for this health risk to manifest and amplify. Here we review the current status of biomedical data inequality and present a conceptual framework for understanding its impacts on machine learning. We also discuss the recent advances in algorithmic interventions for mitigating health disparities arising from biomedical data inequality. Finally, we briefly discuss the newly identified disparity in data quality among ethnic groups and its potential impacts on machine learning.

Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics

Polygenic risk scores (PRS) calculated from genome-wide association studies (GWAS) of Europeans are known to have substantially reduced predictive accuracy in non-European populations, limiting their clinical utility and raising concerns about health disparities across ancestral populations. Here, we introduce a statistical framework named X-Wing to improve predictive performance in ancestrally diverse populations. X-Wing quantifies local genetic correlations for complex traits between populations, employs an annotation-dependent estimation procedure to amplify correlated genetic effects between populations, and combines multiple population-specific PRS into a unified score with GWAS summary statistics alone as input. Through extensive benchmarking, we demonstrate that X-Wing pinpoints portable genetic effects and substantially improves PRS performance in non-European populations, showing 14.1%-119.1% relative gain in predictive R2 compared to state-of-the-art methods based on GWAS summary statistics. Overall, X-Wing addresses critical limitations in existing approaches and may have broad applications in cross-population polygenic risk prediction.

Leveraging the local genetic structure for trans-ancestry association mapping

Over the past two decades, genome-wide association studies (GWASs) have successfully advanced our understanding of the genetic basis of complex traits. Despite the fruitful discovery of GWASs, most GWAS samples are collected from European populations, and these GWASs are often criticized for their lack of ancestry diversity. Trans-ancestry association mapping (TRAM) offers an exciting opportunity to fill the gap of disparities in genetic studies between non-Europeans and Europeans. Here, we propose a statistical method, LOG-TRAM, to leverage the local genetic architecture for TRAM. By using biobank-scale datasets, we showed that LOG-TRAM can greatly improve the statistical power of identifying risk variants in under-represented populations while producing well-calibrated p values. We applied LOG-TRAM to the GWAS summary statistics of various complex traits/diseases from BioBank Japan, UK Biobank, and African populations. We obtained substantial gains in power and achieved effective correction of confounding biases in TRAM. Finally, we showed that LOG-TRAM can be successfully applied to identify ancestry-specific loci and the LOG-TRAM output can be further used for construction of more accurate polygenic risk scores in under-represented populations.