Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Apr 15.
Published in final edited form as: Am J Hum Genet. 2026 Apr 2;113(4):828–841. doi: 10.1016/j.ajhg.2026.03.008

Multi-ancestry transcriptome prediction with functionally informed variants in TOPMed MESA improves performance of transcriptome-wide association studies

Xiaowei Hu 1,*, Daniel S Araujo 2,3, Chachrit Khunsriraksakul 4, Lida Wang 5, Quan Sun 6,7,8,9, Jia Wen 10, Lingbo Zhou 6, Lynette Ekunwe 11, Leslie A Lange 12, Ethan M Lange 12, Stephen B Montgomery 13, Alexander P Reiner 14, Francois Aguet 15, Kristin G Ardlie 15, Tuuli Lappalainen 16, Christopher R Gignoux 17, Esteban G Burchard 18, Kent D Taylor 19, Xiuqing Guo 19, Jerome I Rotter 19, Stephen S Rich 1, Elaine Cornell 20, Peter Durda 20, Russell P Tracy 20, Yongmei Liu 21, W Craig Johnson 22, George P Papanicolaou 23, Minoli A Perera 24, Michael H Cho 25, Dajiang J Liu 4, Laura M Raffield 10, Yun Li 6,10; TOPMed Multi-Omics Working Group, Heather E Wheeler 2, Hae Kyung Im 26, Ani Manichaikul 1,*
PMCID: PMC13078710  NIHMSID: NIHMS2161406  PMID: 41932314

Abstract

Reliable reference transcriptome prediction models are key to accurate multi-ancestry transcriptome-wide association study (TWAS). We propose three methods leveraging functionally informed variants (FIVs) for transcriptome prediction models to improve multi-ancestry TWAS. We trained models on 1,287 multi-ancestry participants from the Trans-Omics for Precision Medicine (TOPMed) program Multi-Ethnic Study of Atherosclerosis (MESA) with RNA-seq data from peripheral blood mononuclear cells (PBMCs). We validated models’ prediction accuracy on two external independent datasets, Geuvadis and Jackson Heart Study. To test robustness of our methods for TWAS, we integrated models with three multi-ancestry GWASs from blood cell, lipid, and pulmonary function traits, respectively. Our methods presented similar prediction accuracy while using a smaller and functionally informed set of variants compared to the benchmark method, Elastic Net (EN). Overall, our methods achieved higher power and accuracy (with average improved accuracy of 24% over EN) for TWAS. However, no single proposed method outperformed all GWAS traits. To further improve TWAS performance, we propose an omnibus approach that aggregates TWAS summary statistics from our methods. The omnibus approach yielded the highest number of Bonferroni-significant TWAS genes for all GWAS traits, and it further improved TWAS power and accuracy for blood cell traits. Additionally, the omnibus approach detected some trait-relevant important genes that the EN missed. Our study demonstrates the value of including FIVs in multi-ancestry transcriptome prediction models for improving TWAS performance. Further, the observed TWAS improvement depends on the GWAS trait’s relevance to the PBMCs used to build our transcriptome prediction models.

Introduction

Although large-scale genome-wide association studies (GWASs) have identified numerous genetic variants associated with complex human diseases or traits, the inference from identified genetic variants to candidate genes remains challenging. Over 90% of GWAS variants are in non-protein-coding regions and many candidate genes at GWAS loci are not the closest.1,2 Transcriptome-wide association study (TWAS), a widely used approach that integrates GWAS with transcriptome data,3 has potential to identify candidate genes and has gained a broad range of applications to human genetics.1,4–7 Three steps are involved in performing TWAS analysis. First, reference transcriptome prediction models are trained in genotype data and tissue or cell-type-specific transcriptome data. Second, genetically regulated gene expression (GReX) is imputed by integrating individual-level genotypes of GWAS with reference transcriptome prediction models. Lastly, the association test between imputed GReX and GWAS trait is performed to identify trait-associated genes. TWAS can also be used with GWAS summary statistics via tools like S-PrediXcan.3 Hence, reliable reference transcriptome prediction models are critical for accurate TWAS analysis.

Recently, multi-ancestry GWASs have been emerging and adding more biological insights into non-European ancestries and signals shared across ancestries.8–10 At present, reference transcriptome prediction models for TWAS have been constructed primarily based on individuals of European ancestry. However, studies have shown that the integration of GWAS with reference transcriptome prediction models from mismatched ancestries had suboptimal power.11,12 Further, the interpretation of ancestry-mismatched TWAS results may be difficult as causal variants underlying GWAS and expression quantitative trait locus (eQTL) hits may differ by ancestry. Reliable multi-ancestry transcriptome prediction models are needed for multi-ancestry TWAS.

The genomic variants overlapping with key annotations (e.g., fine-mapping, epigenetic processes, and 3D genomics-informed regions), defined as functionally informed variants (FIVs) in this study, are shown to be more likely to influence gene expression.13 Recent studies have demonstrated the power of incorporating FIVs into transcriptome prediction models to improve both prediction model accuracy and downstream TWAS performance.14–16 However, the existing transcriptome models with FIVs were built on individuals of European ancestry. There is a lack of ancestry diversity for transcriptome prediction models with FIVs.

Motivated by previous studies, we hypothesized that multi-ancestry transcriptome prediction models using FIV annotation could improve both prediction accuracy and multi-ancestry TWAS performance. Here, we investigated prediction accuracy of multi-ancestry transcriptome prediction models with FIVs (hereinafter referred to as FIV-based methods) and explored the benefits of methods on multi-ancestry TWAS. We constructed transcriptome prediction models on 1,287 multi-ancestry participants from the NHLBI Trans-Omics for Precision Medicine (TOPMed) program Multi-Ethnic Study of Atherosclerosis (MESA) with RNA-seq data from peripheral blood mononuclear cells (PBMCs). The developed transcriptome prediction models were integrated with three large-scale multi-ancestry GWAS of blood cell traits,8 lipid traits,9 and pulmonary function traits,10 respectively, to evaluate models’ performance for multi-ancestry TWAS. Further, inspired by the increased power of aggregated TWAS from different models or tissues in the previous studies,16–18 we developed an omnibus approach for our FIV-based methods by aggregating their TWAS p-values.

Material and Methods

Overview of approach

An overview of study design is shown in Figure 1. We applied four different prediction methods to construct transcriptome prediction models based on multi-ancestry participants from the NHLBI Trans-Omics for Precision Medicine (TOPMed) Multi-Ethnic Study of Atherosclerosis (MESA). The prediction methods examined included three proposed methods that leverage functionally informed variants (FIVs) (hereinafter referred to as FIV-based methods): Elastic Net with Fine-Mapped variants (EN-FM), Prediction Using Models Informed by Chromatin conformation and Epigenomics (PUMICE),16 and PUMICE with Fine-Mapped variants (PUMICE-FM), as well as one method examined for the purpose of benchmarking, Elastic Net (EN), in the PrediXcan19 framework. We examined prediction accuracy of transcriptome prediction models in the MESA training data, as well as two external independent validation data sets from Geuvadis20 consortium and the Jackson Heart Study21 (JHS). Finally, we applied transcriptome prediction models to three large-scale multi-ancestry GWASs to evaluate each model’s performance for multi-ancestry TWAS. In addition to assessing TWAS performance for the four methods, we also evaluated an omnibus approach for TWAS that aggregates TWAS summary statistics from our FIV-based methods. In total, there were five sets of TWAS results for each GWAS trait, and we assessed both power and accuracy for each TWAS.

Figure 1.

Figure 1

Open in a new tab

Study design. The two pie charts present ancestry distribution of MESA and Geuvadis participants respectively. TOPMed, Trans-Omics for Precision Medicine program; MESA, Multi-Ethnic Study of Atherosclerosis; PBMC, peripheral blood mononuclear cells; EN, Elastic Net; EN-FM, Elastic Net with Fine-Mapped variants; PUMICE, Prediction Using Models Informed by Chromatin conformation and Epigenomics; PUMICE-FM, PUMICE with Fine-Mapped variants; Omnibus, aggregated model of EN-FM, PUMICE, and PUMICE-FM; LCLs, lymphoblastoid cell lines; JHS, Jackson Heart Study; GWAS, genome-wide association study; NHW, Non-Hispanic White; AFA, African American; HIS, Hispanic/Latinos; CHN, Chinese; CEU, Utah residents with Northern and Western Europeans; FIN, Finnish in Finland; GBR, British in England and Scotland; TSI, Toscani in Italy; YRI, Yoruba in Ibadan, Nigeria.

Training data for multi-ancestry transcriptome prediction models

The TOPMed MESA 1,287 multi-ancestry participants with TOPMed Freeze 8 whole-genome sequencing (WGS) data and RNA-seq data from peripheral blood mononuclear cells (PBMCs) were used to build transcriptome prediction models. The TOPMed MESA data are under controlled access in dbGaP through study accession number phs001416.v4.p1. Data in dbGap can be downloaded by controlled access with an approved application submitted through dbGap website https://www.ncbi.nlm.nih.gov/gap.

The Multi-Ethnic Study of Atherosclerosis:

MESA is a longitudinal study of subclinical cardiovascular disease and risk factors that predict progression to clinically overt cardiovascular disease or progression of subclinical disease.22 Between 2000 and 2002, MESA recruited 6,814 men and women 45 to 84 years of age from Forsyth County, North Carolina; New York City; Baltimore; St. Paul, Minnesota; Chicago; and Los Angeles. Exclusion criteria were clinical cardiovascular disease, weight exceeding 136 kg (300 lb.), pregnancy, and impediment to long-term participation. Approximately 38 percent of the recruited participants are white, 28 percent African American, 22 percent Hispanic/Latino, and 12 percent Asian, predominantly of Chinese descent. All research was carried out in accordance with relevant guidelines and regulations. All participants provided informed consent, and the protocols of MESA were approved by the IRBs of collaborating institutions and the NHLBI and informed consent was obtained from all participants. Research involving human research participants was performed in accordance with the Declaration of Helsinki.

RNA sequencing:

RNA-seq was performed for samples obtained from selected MESA participants at Exam 1 (corresponding to 2000–2002) and Exam 5 (2010–2011). RNA-seq was performed at the Broad Institute Genomics Platform as part of the TOPMed MESA Multi-Omics project based on the following criteria: (1) restrict to those already included in the TOPMed WGS effort,23 (2) preserve the race/ancestry distribution of participants in the parent MESA cohort, (3) maximize the amount of overlapping ‘omics data (with the other ‘omics included in the TOPMed MESA Multi-Omics pilot requiring availability of plasma samples for proteomics/metabolomics, RNA from monocytes or T cells for RNA-seq, and whole blood for DNA methylation profiling).

Whole genome sequencing:

Genotypes were obtained from WGS data in TOPMed Freeze 8 samples. ~30X WGS was performed at the Broad Institute of MIT and Harvard, and the TOPMed Informatics Research Center performed joint genotype calling using all samples in Freeze 8 on the GRCh38 assembly. Details regarding DNA sample extraction, WGS data sequencing method and quality control are available at https://topmed.nhlbi.nih.gov/topmed-whole-genome-sequencing-methods-freeze-8.

Preparation of RNA-seq data for transcriptome prediction modeling:

Quality-control of the WGS data and preparation of gene expression levels for prediction modeling were as described previously.24 The residuals of expression levels were generated from a linear regression model of rank-based inverse normalization of expression levels adjusted for age, sex, the first 10 genotype and 10 expression principal components (PCs) for individuals with only one-time point data (Exam 1 or 5). The number of PCs is the same as the original paper24 that performed transcriptome prediction modeling with the same gene expression data. For individuals with two Exams, the average expression data from two Exams was used the same way as described previously.24 We used the residuals of expression levels as dependent variables later for transcriptome prediction model development. Based on participant self-reported race/ethnicity and comparison with global proportion of each ancestry that was computed in a previous study of MESA participants,25 41%, 26%, 25%, and 8% of the participants were categorized as Non-Hispanic White (NHW), African American (AFA), Hispanic/Latinos (HIS), and Chinese (CHN) respectively among TOPMed MESA 1,287 participants (Figure 1).

Transcriptome prediction methods

We developed three prediction methods with FIVs to build transcriptome prediction models, and we used EN in the PrediXcan19 framework as a benchmark for comparison. For all four methods, we built transcriptome prediction models between residuals of expression levels and cis-SNPs within the 1Mb upstream region of transcription start site (TSS) and 1 Mb downstream region of transcription end site (TES). The prediction performance was estimated using a nested cross-validation approach as described in earlier works.16,19 As suggested in previous works,15,16,19,24 a model was considered significant and was kept if the average of cross-validated Pearson correlation coefficients> 0.1 and the estimated p-value<0.05.

EN: Elastic Net.

For each gene, EN includes all variants within cis-region for variable selection without incorporating variants’ priors informed by their biological functions. We used the R package glmnet26 to implement EN modeling with a mixing parameter α=0.5.

EN-FM: Elastic Net with Fine-Mapped variants.

Motivated by the goal of using the most likely causal variants, we applied SuSiE27 to 1,287 TOPMed MESA multi-ancestry participants to obtain fine-mapped variants for each gene by using MESA in-sample linkage disequilibrium (LD) and SuSiE’s default set-up of parameters, including maximum number of non-zero effects in the SuSiE regression model L=10 and minimum absolute correlation allowed in a credible set=0.5. The genomic variants included in our fine-mapping analysis were cis-SNPs within the 1Mb upstream region of TSS and 1 Mb downstream region of TES for each gene. Since fine-mapped variants with higher posterior inclusion probability (PIP) are the variants most likely to be causal for a gene expression trait, we used (1-PIP) as a penalty for each variant included in EN modeling as described in Barbeira et al. 2020.15 As a result, the variants with higher PIP received smaller penalties and were upweighted in prediction modeling. The usage of (1-PIP) as a penalty is the main difference between EN-FM and EN modeling. We implemented EN-FM modeling in R/glmnet26 with a mixing parameter α=0.5.

PUMICE: Prediction Using Models Informed by Chromatin conformation and Epigenomics.

PUMICE16 separates genomic variants into essential and non-essential predictors depending on whether they overlap with user-specified functional annotation. Additionally, PUMICE allows different choices for regions that harbor cis-regulatory variants as another tuning. We used epigenomic annotation tracks as functional annotation and 3D genomes as cis-regulatory regions. Given that neither epigenomic annotation nor 3D genomic data were available for PBMCs, we used the available data from EBV-transformed lymphocytes as a proxy. The epigenomic annotation data, including H3K27ac, H3K4me3, DNase hypersensitive, and CTCF, were obtained from Encyclopedia of DNA Elements (ENCODE)28,29 via accession number ENCFF028SGJ and under dataset ENCSR480YCS (https://www.encodeproject.org/annotations/ENCSR480YCS/). The 3D genomic data were obtained from GSE6352530 (loop, TAD, and domain), and GSE8618931 (pcHi-C) for lymphoblastoid cell line (GM12878). Both epigenomic annotation and 3D genomic data resources were built on GRCh37/hg19, we lifted them over to GRCh38/hg38 to align with genomic build of TOPMed MESA genotypes by using UCSC LiftOver tool.32 We followed code provided in the PUMICE Github (https://github.com/ckhunsr1/PUMICE) to run PUMICE with default settings.

PUMICE-FM: PUMICE with Fine-Mapped variants.

Prior investigation of PUMICE method found that a 250 kb window surrounding the gene start and end site was the most frequent choice among the best performing models.16 Accordingly, we applied the 250 kb window surrounding the TSS and TES as cis-regulatory regions and used fine-mapping as functional annotation to separate genomic variants into essential and non-essential groups for PUMICE-FM. The method for fine-mapping is provided above in the section of EN-FM modeling.

Prediction accuracy validation analysis

To evaluate prediction accuracy of all four transcriptome prediction methods, we used two independent external data sets from Geuvadis and JHS to validate prediction performance.

Geuvadis:

The data included 449 multi-ancestry participants with WGS and RNA-seq from lymphoblastoid cell lines (LCLs) in the Geuvadis20 consortium. Quality-control of the WGS data and preparation of gene expression levels for prediction modeling were described previously.24 The residuals of expression levels were generated from a linear regression model of rank-based inverse normalization of expression levels adjusted for the first 10 genotype and 10 expression principal components (PCs). The number of PCs is the same as the original paper24 that performed transcriptome prediction modeling with the same gene expression data. Among 449 Geuvadis participants, 80% and 20% of the participants were categorized as Europeans (Utah residents with Northern and Western European [CEU]; Finnish in Finland [FIN]; British in England and Scotland [GBR]; and Toscani in Italy [TSI]) and Africans (Yoruba in Ibadan, Nigeria [YRI]) respectively (Figure 1). The Geuvadis data can be accessed at https://www.internationalgenome.org/data-portal/data-collection/geuvadis.

Jackson Heart Study (JHS):

The JHS data included 1,012 African American participants with TOPMed Freeze 8 WGS and RNA-seq from PBMCs.21 Quality-control of the WGS and RNA-seq data were described previously.33 Using a linear regression model for the outcome of rank-based inverse normalized expression levels with covariate adjustment for age, sex, the first 10 genotype PCs, and 70 probabilistic estimations of expression residuals (PEER) factors, we obtained residuals of expression levels for use in evaluation of prediction accuracy for the MESA models. The genotype PCs and PEER factors were calculated as described previously,33 and we used the same number of PCs and PEER factors as suggested in the original paper33 that performed eQTL mapping. The JHS RNA-seq and WGS data are under controlled access in dbGaP through study accession number phs000286.v6.p2 and phs000964.v5.p1, respectively.

To validate prediction accuracy, we computed the Pearson correlation coefficient between the observed residuals of expression and the estimated genetically regulated expression component (GReX)19 for each gene. The estimated GReX was calculated by multiplying the genotype dosage of each variant by its respective weight that was estimated from training models and then summing across all variants for each gene.

TWAS

To examine performance of transcriptome prediction models in multi-ancestry TWAS, we applied S-PrediXcan3 to integrate multi-ancestry GWAS summary statistics with transcriptome prediction models to get TWAS results. We followed code provided in the S-PrediXcan Github (https://github.com/hakyimlab/MetaXcan/wiki/S-PrediXcan-Command-Line-Tutorial). The TOPMed MESA individual-level genotype data was used to compute covariance matrix that is required by running S-PrediXcan pipeline. We selected three large-scale multi-ancestry GWAS for investigation:(1) GWAS of eight blood cell traits,8 (2) GWAS of five lipid traits,9 and (3) GWAS of four pulmonary function traits,10 to test robustness of the transcriptome prediction methods for downstream TWAS analyses. Among the blood cell traits, we conducted TWAS for six white blood cell traits: basophil (BASO), eosinophil (EOS), lymphocyte (LYM), monocyte (MONO), neutrophil (NEU), and white blood cell count (WBC); and two platelet traits, mean platelet volume (MPV) and platelet count (PLT). For lipid traits, we performed TWAS for five traits: high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), log-transformed triglycerides (logTG), non-high-density lipoprotein cholesterol (nonHDL-C), and total cholesterol (TC). For pulmonary function traits, we generated TWAS for four traits: forced expiratory volume in 1 second (FEV1), forced vital capacity (FVC), ratio of FEV1 to FVC (FEV1/FVC), and peak expiratory flow rate (PEF). All GWAS summary statistics are available in the GWAS Catalog34 (https://www.ebi.ac.uk/gwas/home) except multi-ancestry GWAS for blood cell traits. We obtained multi-ancestry GWAS summary statistics of blood cell traits via lead contact of paper8. For the other traits, the GWAS Catalog study accession numbers for multi-ancestry GWAS are GCST90239649 - GCST90239673 for lipid traits and GCST90244092 - GCST90244095 for pulmonary function traits. The GWAS Catalog study accession numbers for European (EUR)-specific GWAS are GCST90002292- GCST90002357 for blood cell traits, GCST90239652 - GCST90239676 for lipid traits and GCST90292609 - GCST90292612 for pulmonary function traits.

Curated genes for TWAS evaluation

For each GWAS trait, we prepared a curated list of known or candidate genes for evaluation of TWAS performance in identifying trait-relevant genes. The curated gene lists were identified from previous gene prioritization efforts for the same GWAS. For blood cell traits, curated genes were defined as those in ‘Category 1’ from the VAMPIRE;35 these genes were selected based on consistent evidence from epigenomic/sequence constraint features, expression/protein/splicing QTL, and 3D genome conformation. For lipid and pulmonary function traits, we defined curated genes that were prioritized by the criteria that did not use transcriptome data information. More specifically, we defined curated genes as those genes prioritized in Kanoni et al.36 according to at least one of the following criteria: genes associated with Mendelian lipid disorders, human ortholog of Mouse Knockout genes associated with lipid phenotypes, genes with lipid-associated protein-altering variants, and genes nearest to sentinel GWAS variants. For pulmonary function traits, we constructed the list of curated genes as those prioritized in Shrine et al.10, implicated by at least one of the following criteria: genes associated with Mendelian respiratory diseases, human ortholog of Mouse Knockout genes associated with respiratory phenotypes, genes with significant rare (minor allele frequency MAF<1%) exonic associations (P<5×10−6) near (±500 kb) sentinel GWAS variants using both single-variant and gene-based collapsing tests, and genes nearest to sentinel GWAS variants. As stated in the above three studies,10,35,36 the annotation resources used to identify the curated genes were mainly from individuals of European ancestry.

Omnibus approach of FIV-based methods

Inspired by the increased power of aggregated TWAS from different models or tissues in the previous studies,16–18 we developed an omnibus approach for our FIV-based methods by aggregating TWAS p-values from EN-FM, PUMICE, and PUMICE-FM using the Aggregated Cauchy Association Test (ACAT).37 If a gene had a corresponding prediction model from at least two of our FIV-based methods, then we applied ACAT to the p-values from each corresponding method to get aggregated p-values. The corresponding z-scores were then derived from the aggregated p-values. If a gene was only predicted by one of our FIV-based methods, we then used the available TWAS summary statistics for the gene’s omnibus TWAS results.

TWAS evaluation

TWAS results for each of four proposed TWAS methods (EN-FM, PUMICE, PUMICE-FM, and omnibus) and the method for benchmarking (EN) were evaluated for each GWAS trait. We began by checking the Quantile-Quantile (QQ) plots with corresponding λ, a measure to quantify the inflation in the test statistics of TWAS. We then used R package BACON38 with 50,000 iterations and 10,000 burnin period to correct for TWAS inflation when λ >1.15. Next, as a proxy for evaluation of TWAS power, we compared median Chi-square test statistics of all TWAS genes that were included in the curated gene lists. The TWAS genes refer to the genes with non-missing TWAS summary statistics (i.e., z-score and p-value). The Chi-square test statistic for a TWAS gene is computed as the square of inflation-corrected TWAS’s z-score of the gene. We performed a Mann-Whitney U statistical test to test if the observed median Chi-square test statistic from each proposed method was significantly greater than that from TWAS from EN prediction models. In addition to evaluating improved power at nominal significance level, we also applied a Bonferroni correction cut-off 0.0125 (0.05/4) on four pairwise comparisons (i.e., EN-FM vs. EN, PUMICE vs. EN, PUMICE-FM vs. EN, and Omnibus vs. EN) for each trait to define a stringent significance.

Finally, we evaluated the accuracy of Bonferroni-significant TWAS genes that were included in curated gene lists using the F1 score. The F1 score, a harmonic mean of the precision and recall, measures the accuracy of significant TWAS genes by accounting for true positives, false positives, and false negatives. F1 score=2*precision*recall/(precision+recall). The higher F1 score indicates better performance. We treated curated genes as actual class and Bonferroni-significant TWAS genes as predicted class. Hence, each TWAS gene has a binary indicator (yes or no) for both actual and prediction classes. We then measured true positives, false positives, and false negatives to calculate the F1 score for each prediction method.

Comparison between multi-ancestry and single-ancestry TWAS

To further investigate the TWAS performance built on European (EUR)-specific prediction models and EUR-specific GWAS to evaluate the benefits of our proposed multi-ancestry prediction models on TWAS over single-ancestry models, we conducted additional analyses by 1) building our proposed three prediction models (i.e., EN-FM, PUMICE, and PUMICE-FM) on the same MESA non-Hispanic White participants as used in our MESA multi-ancestry models, hereinafter referred to as EUR models; 2) integrating these EUR models with EUR-specific GWASs of blood cell traits,8 lipid traits,9 and pulmonary function traits10 respectively to get their corresponding TWAS, hereinafter referred to as EUR_EUR; 3) integrating the EUR models with three multi-ancestry GWASs respectively to get their corresponding TWAS, hereinafter referred to as EUR_Multi; 4) integrating our multi-ancestry models with EUR-specific GWASs respectively to get their corresponding TWAS, hereinafter referred to as Multi_EUR; and 5) applying omnibus approach to EUR_EUR, EUR_Multi, and Multi_EUR TWASs to get their corresponding omnibus TWAS respectively for each of 17 GWAS traits. By following the labeling rule, our original TWASs built on multi-ancestry prediction models and multi-ancestry GWASs are referred to as Multi_Multi.

Results

Similar prediction accuracy while using a smaller set of variants by FIV-based methods

After removing non-significant gene expression prediction models (Methods), there were similar numbers of genes remaining for EN (11,897), EN-FM (11,033), PUMICE (11,601), and PUMICE-FM (11,000) (Figure 2A). EN had the largest median model size (median number of SNPs in the models of 49), while the EN-FM produced the smallest median model size (median number of SNPs in the models of 3), which was expected based upon fine-mapping of diverse ancestry participants (Figure 2A). To make a fair comparison of prediction accuracy evaluated by Pearson correlation coefficient between the observed residuals of expression and the estimated GReX across the four different methods, we considered the intersection of genes with prediction models available from all four methods for both discovery (MESA) and validation (Geuvadis and JHS) analyses. Based on the overlapped set of 8,659 genes across the four methods trained by MESA multi-ancestry participants, the median Pearson correlation coefficients are similar (Figure 2B) among the four methods, and further most genes had similar prediction accuracy (Figure S1). Likewise, no noticeable difference in prediction accuracy among the four methods was observed in validation analysis evaluated on the overlapped set of 9,533 and 9,098 genes for Geuvadis multi-ancestry participants and JHS African Americans, respectively (Figure S2). However, the JHS had much higher prediction accuracy than Geuvadis (median Pearson correlation coefficient=0.2 from JHS vs 0.1 from Geuvadis) (Figure S2), which indicates the importance of cell-type relevance in the replication of transcriptome prediction.

Figure 2.

Figure 2

Open in a new tab

Transcriptome prediction model size and prediction accuracy. A) Boxplots of transcriptome prediction model size of genes in MESA multi-ancestry training data. The solid line inside the box indicates the median; box edges represent 25th and 75th quartiles; whiskers extend to the minimum and maximum values within 1.5 times the interquartile range (IQR); individual points indicate data outside 1.5 times the IQR. B) Violin plots of prediction accuracy of 8,659 common genes across four methods of MESA multi-ancestry training data. The prediction accuracy is evaluated by Pearson correlation coefficient between the observed residuals of expression and the estimated genetically regulated expression component (GReX). The embedded box plots indicate the median (horizontal line) and IQR (the rectangular box); whiskers extend to the minimum and maximum values within 1.5 times the IQR; the width of the violin plots represent the density of Pearson correlation coefficient at a specific value. C) Pie chart representing the distribution of number of our proposed methods achieving higher prediction accuracy than EN, examined across 8,659 genes with prediction models for all methods. Zero indicates EN had higher prediction accuracy. D) Stacked bar chart of proportion of shared SNPs between EN and proposed models. MESA, Multi-Ethnic Study of Atherosclerosis; EN, Elastic Net; EN-FM, Elastic Net with Fine-Mapped variants; PUMICE, Prediction Using Models Informed by Chromatin conformation and Epigenomics; PUMICE-FM, PUMICE with Fine-Mapped variants.

Although the four methods presented similar prediction accuracy, 84% of 8,659 common genes showed higher prediction accuracy in at least one of our FIV-based methods compared to EN (Figure 2C). More specifically, 43% of genes exhibited higher prediction accuracy in all three methods with FIVs (Figure 2C). Among our three FIV-based methods, EN-FM had the most overlap of SNPs used to predict the same gene with EN while PUMICE had the least intersection with EN. We observed that 82% of 8,659 common genes from EN-FM had at least 80% shared SNPs with EN, however, the model size of EN-FM was much smaller than EN, which indicates the advantage of using fine-mapped variants for predicting gene expression (Figure 2D). On the other hand, we found that only 7% of common 8,659 genes from PUMICE shared at least 80% SNPs with EN, which indicates the set of SNPs included in two methods was quite different (Figure 2D). Furthermore, we specifically examined the linkage disequilibrium (LD) of SNPs included in models for genes in which the proposed FIV-based method produced the same prediction accuracy as EN (i.e., Pearson Correlation Coefficient, rounded to the nearest hundredth). There were 1648, 1804, and 1762 such genes from EN-FM, PUMICE, and PUMICE-FM respectively. Overall, most SNPs included in the three FIV-based methods were uncorrelated (i.e., LD≤0.2) with SNPs included in the EN (Figure S3). However, the proportion of highly correlated (i.e., LD≥0.8) SNPs increases as the proportion of shared SNPs increases, for example, 14% and 3% of PUMICE SNPs are highly correlated with EN SNPs in the scenario of genes that share 80–100% and 0–50% of model SNPs, respectively (Figure S3B). Moreover, we found that the SNPs included in our FIV-based methods were closer to both TSS and TES compared with SNPs included in EN (Figure S4).

Overall, our three FIV-based methods had fewer variants included in the prediction models, while achieving similar gene expression prediction performance compared to the benchmark method, EN.

Overall higher TWAS power from FIV-based methods

To assess transcriptome prediction methods on multi-ancestry TWAS performance, we integrated three large-scale multi-ancestry GWASs with transcriptome prediction models to obtain trait-specific TWAS (Methods). We first checked genomic inflation of TWAS and found a genomic inflation measure, λ, ranging from 1.19 to 1.54 for blood cell traits (Figures S5–12), from 1.28 to 1.47 for lipid traits (Figures S13–17), and from 1.18 to 1.32 for pulmonary function traits (Figures S18–21). The inflated TWAS could be attributed in part to large proportions of true positive results due to powerful GWAS and the polygenicity of GWAS traits, as pointed out in a previous study of TWAS inflation.39 However, given the large levels of genomic inflation, we applied genomic control (GC) correction on results from each of the TWAS (Methods). The λ for GC-corrected TWAS ranged from 1.06 to 1.16 (Figures S5–21). Our subsequent evaluation of TWAS results was based on GC-corrected TWAS. Overall, EN produced the largest number of TWAS genes, while EN-FM had the smallest list of genes combined across all traits (Table S1). To evaluate overall TWAS power, we overlapped all TWAS genes with curated genes and compared models’ median Chi-square test statistic (Methods). The Chi-square test statistic for a TWAS gene is computed as the square of GC-corrected TWAS’s z-score of the gene. The total number of curated genes for each GWAS trait is shown in Table S2, and a full list of curated genes is provided in Table S3–5 respectively for blood cell, lipid, and pulmonary function traits. The number of TWAS genes overlapping with curated genes is shown in Table S6. In general, EN overlapped most with curated genes for blood cell and pulmonary function traits, while the PUMICE had more overlapped curated genes for lipid traits (Table S6). Although EN identified more curated genes, the median Chi-square test statistics from at least one of our FIV-based methods were higher than those from EN across all traits except BASO (Table S7). Further, our FIV-based methods had nominally significantly higher (P<0.05) median Chi-square for three GWAS traits, however, EN did not significantly outperform in any traits. More specifically, PUMICE produced higher median Chi-square values for PLT (P=0.0395), LDL-C (P=0.0419), and TC (P=0.0264), and PUMICE-FM had significantly higher median Chis-square for PLT (P=0.0220) at nominal significance level (Figure 3, Table S7).

Figure 3.

Figure 3

Open in a new tab

Power of transcriptome-wide association study evaluated by distribution of Chi-square values. The horizontal line inside the box indicates the median; box edges represent 25th and 75th quartiles; whiskers extend to the minimum and maximum values within 1.5 times the interquartile range (IQR); individual points indicate data outside 1.5 times the IQR. * P-value<0.05; ** P-value<0.005 from Mann-Whitney U test without correcting for multiple testing. The genes that were included in the evaluation were overlap of GC-corrected TWAS genes with curated genes for each trait. For visualization purposes, the Chi-square values over 200 have been excluded from the plots but not excluded from statistical tests. The number of genes included in the statistical test for each method and each trait is provided in Table S6. There were 15, 24, and 21 genes excluded from the plot for PLT, LDL-C, and TC respectively. The number shown on top of each box plot is the median chi-square value. Chi-square value, the square of GC-corrected TWAS’s z-score of a gene; GC, genomic control; TWAS, transcriptome-wide association study; p-value, from Mann-Whitney U test to test if the median Chi-square value from proposed method was significantly greater than that from EN TWAS; PLT, platelet count; LDL-C, low-density lipoprotein cholesterol; TC, total cholesterol; EN, Elastic Net; EN-FM, Elastic Net with Fine-Mapped variants; PUMICE, Prediction Using Models Informed by Chromatin conformation and Epigenomics; PUMICE-FM, PUMICE with Fine-Mapped variants; Omnibus, aggregated model of EN-FM, PUMICE, and PUMICE-FM.

Higher accuracy of significant TWAS genes from FIV-based methods

To evaluate accuracy of statistically significant TWAS genes, we overlapped Bonferroni-significant TWAS genes with curated genes (hereinafter referred to as Bonferroni-curated TWAS genes) and then calculated the F1 scores for Bonferroni-curated TWAS genes to assess their accuracy. The Bonferroni threshold was based on the rounded number of union of TWAS genes across the four methods and across the 17 GWAS traits, i.e., 0.05/13000=3.85×10−6. The number of Bonferroni-significant TWAS genes and number of Bonferroni-curated TWAS genes are shown in Figure S22 and Figure S23 respectively for each GWAS trait. Our FIV-based methods not only produced the same or more Bonferroni-curated TWAS genes in 13 out of 17 GWAS traits compared with EN (Figure S23) but also had higher F1 scores for all tested GWAS traits except BASO and WBC (Figure 4). The improvement of accuracy ranges 5–15% for blood cell traits, 4–18% for lipid traits, and 12–133% for pulmonary function traits respectively (Figure 4). Prominent improvements were observed for PEF (F1=0.14 from PUMICE vs 0.06 from EN) and FVC (F1=0.07 from EN-FM vs 0.03 from EN, Figure 4). In sensitivity analysis, we further calculated F1 scores using overlapped genes between FDR-significant TWAS genes (FDR<0.05) and curated genes. Our FIV-based methods still outperform EN for most GWAS traits (Figure S24).

Figure 4.

Figure 4

Open in a new tab

Accuracy of Bonferroni-significant transcriptome-wide association study genes evaluated by F1 score. A) F1 scores of eight blood cell traits; B) F1 scores of five lipid traits; C) F1 scores of four pulmonary function traits. The genes included in the evaluation were overlapping Bonferroni-significant TWAS genes with curated genes for each trait. TWAS, transcriptome-wide association study; F1 score=2×precision×recall÷(precision+recall); BASO, Basophil; EOS, Eosinophil; LYM, Lymphocyte; MONO, Monocyte; NEU, Neutrophil; WBC, white blood cell count; MPV, mean platelet volume; PLT, platelet count; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; logTG, log-transformed triglycerides; nonHDL-C, non-high-density lipoprotein cholesterol; TC, total cholesterol; FEV1, forced expiratory volume in 1 second; FVC, forced vital capacity; FEV1/FVC, ratio of FEV1 to FVC; PEF, peak expiratory flow rate; EN, Elastic Net; EN-FM, Elastic Net with Fine-Mapped variants; PUMICE, Prediction Using Models Informed by Chromatin conformation and Epigenomics; PUMICE-FM, PUMICE with Fine-Mapped variants; Omnibus, aggregated model of EN-FM, PUMICE, and PUMICE-FM.

Further improved TWAS power by Omnibus approach

Although our FIV-based methods leveraging FIVs showed higher accuracy than the benchmark EN method for TWAS, no single proposed method uniformly outperformed for all tested GWAS traits. Inspired by the increased power of aggregated TWAS from different models or tissues in the previous studies, we sought to aggregate TWAS summary statistics from three FIV-based methods using an omnibus approach (Methods). We followed the same evaluation procedure to assess omnibus TWAS. We began by evaluating genomic inflation and then applied GC correction to omnibus TWAS. The GC-corrected λ ranged from 1.08 to 1.19 for all GWAS traits (Table S8). Our subsequent evaluation of omnibus TWAS results was based on GC-corrected TWAS as well. The omnibus approach produced the largest number of TWAS genes (Table S1, ranging from 12,619 to 12,672), and produced the most overlapped genes with curated gene lists for all GWAS traits (Table S6), compared to EN. For overall TWAS power, the omnibus approach had the greatest median of Chi-square values for all GWAS traits except for BASO, LDL-C, FEV1, and FVC (Table S7). Furthermore, the omnibus approach yielded nominally significantly higher (P<0.05) TWAS power than EN in five out of eight blood cell traits and in two out five lipid traits (Table S7). The most statistically significantly improved power was found for TC (P=0.0012), which passed Bonferroni significance (i.e., P<0.0125, Methods) (Figure 3, Table S7). However, the omnibus approach did not achieve significantly greater power than EN for any of the four pulmonary function traits (Table S7). In the evaluation of accuracy for Bonferroni-significant TWAS genes, although the omnibus approach only produced higher F1 scores for six out of eight blood cell traits (Figure 4), the omnibus approach had the largest number of Bonferroni-significant TWAS genes for all GWAS traits (Figure S22) and produced the most Bonferroni-curated TWAS genes for all GWAS traits except for BASO (Figure S23).

Further, we compared Bonferroni-curated TWAS genes identified between EN and omnibus. The common Bonferroni-curated TWAS genes between EN and omnibus had the same TWAS effect direction for all GWAS traits (Table S9–11). We then summarized the number of Bonferroni-curated TWAS genes uniquely identified by either method in Table S12 and provided TWAS summary statistics of these genes in Table S13–15. The omnibus approach uniquely identified more Bonferroni-curated TWAS genes than EN for all GWAS traits except for BASO, with a count ratio (omnibus/EN) ranging from 1.6 (FEV1/FVC) to 12 (EOS) (Table S12). Besides, the omnibus approach uniquely identified 2 and 7 Bonferroni-curated TWAS genes for FVC and PEF respectively, while the EN did not uniquely identify any genes for these two traits (Table S12). We further checked some important curated genes that were uniquely identified by the omnibus approach. We list three examples here for the demonstration of improvement from our omnibus approach. The first example is TRAF1 for LYM. TRAF1 (Tumor necrosis factor Receptor-Associated Factor 1) is a protein coding gene and is located on chromosome 9. TRAF1 has been reported to be a protein playing a role in the survival, proliferation, and cytokine production of lymphocytes.40 Additionally, the increased expression of TRAF1 has been shown to activate lymphocytes.41,42 Our omnibus approach identified it as a Bonferroni-curated TWAS gene (LYM TWAS P =1.15×10−13 and Z-score=7.42, Table S13). However, the EN method produced a much weaker signal for this gene (LYM TWAS P =0.04 and Z-score=2.11, Table S13). The second example is APOA1 for HDL-C. APOA1 (Apolipoprotein A1), located on chromosome 11, is a protein coding gene. APOA1 is a vital HDL-C protein that makes up the majority of HDL-C and helps in the production of HDL particles.43–45 This gene was a Bonferroni-significant TWAS gene from our omnibus approach (HDL-C TWAS P =5.09×10−36 and Z-score=12.53, Table S14), however, it was not included in EN model. Likewise, TGFBR3, a protein coding gene, was the most significant Bonferroni-curated gene uniquely identified by our omnibus approach for FEV1/FVC ratio but was missed by the EN model (Table S15). TGFBR3 (Transforming Growth Factor Beta Receptor 3), located on chromosome 1, is one of TGF-beta superfamily members. The TGF-beta superfamily proteins play a key role in the regulation of extracellular matrix composition and alveolar epithelial cell and fibroblast function in the lung.46 The potential key role of TGFBR3 in the pathogenesis of COPD susceptibility has been suggested in several studies.46–48

Multi-ancestry prediction models improved TWAS performance over single-ancestry models

As the three multi-ancestry GWASs are dominated by European (EUR) ancestry (i.e., about 75%, 80%, and 81% of GWAS participants is EUR for blood cell traits, lipid traits, and pulmonary function traits respectively), we further investigated the TWAS performance built on EUR-specific prediction models and EUR-specific GWASs to evaluate the benefits of our proposed multi-ancestry prediction models over single-ancestry models. As we have demonstrated the improvement of TWAS from the omnibus approach, the evaluation of new TWASs focused on their omnibus results on the performance of power, accuracy, number of Bonferroni-significant TWAS genes, and number of Bonferroni-Curated TWAS genes. We found that the multi-ancestry prediction models produced higher TWAS power than EUR models for blood cell traits on both multi-ancestry GWASs (i.e., Multi_Multi vs EUR_Multi) and EUR-specific GWASs (i.e., Multi_EUR vs EUR_EUR) except for BASO and WBC (Figures S25–26). More specifically, the EOS and LYM had significantly higher power from Multi_EUR than EUR_EUR (Nominal significance P=0.02 for EOS and Bonferroni significance P=7.77×10−4 for LYM), and significantly higher power from Multi_Multi than EUR_Multi (Nominal significance P=0.04 for EOS and Bonferroni significance P=2.33×10−3 for LYM) (Figure S25). However, the multi-ancestry prediction models produced similar TWAS power with EUR models for lipid traits (Figure S27), and it had higher TWAS power only for PEF among four pulmonary function traits (Figure S28) on both multi-ancestry and EUR-specific GWASs. For the accuracy of Bonferroni-significant TWAS genes, the multi-ancestry prediction models produced similar or higher TWAS accuracy compared with EUR models for all GWAS traits except for BASO and FVC (Figure S29). The outperformance of multi-ancestry prediction models for all 17 traits except FVC was also observed for the counts of Bonferroni-significant and Bonferroni-curated TWAS genes (Figures S30–31).

Our investigation supports the overall advantage of using multi-ancestry prediction models over EUR-specific models for both multi-ancestry and EUR-specific GWASs of blood cell traits, lipid traits, and pulmonary function traits, which indicates the capability of our proposed multi-ancestry prediction models on reducing impact on TWAS discovery due to mild ancestry mismatch between TWAS reference model and GWAS. Furthermore, we found that the multi-ancestry prediction models performed similarly for 14 out of 17 GWAS traits for comparisons between multi-ancestry and EUR-specific GWASs (i.e., Multi_Multi vs Multi_EUR) in the evaluation of TWAS power and accuracy. More specifically, both TWAS power and accuracy were higher from Multi_Multi TWAS for NEU, FEV1, and PEF compared with their Multi_EUR TWASs, however, the Multi_EUR TWASs were higher in power and accuracy for BASO, WBC, and MPV (Figures S25–29). These results indicate that using our proposed multi-ancestry prediction models on multi-ancestry GWASs yields performance similar or better than the EUR-specific GWASs for TWAS discovery of blood cell traits, lipid traits, and pulmonary function traits.

Discussion

We proposed three FIV-based methods that upweight genomic variants overlapped with key annotations (i.e., fine-mapping, epigenetic processes, and 3D genomics informed regions) to improve both prediction accuracy of multi-ancestry transcriptome prediction models and power of multi-ancestry TWAS. Building on 1,287 multi-ancestry participants from the TOPMed MESA with TOPMed Freeze 8 WGS data and RNA-seq data from PBMCs, our proposed methods presented similar prediction accuracy in both MESA discovery analysis and validation analysis with two external data sets (Geuvadis and JHS) while using a smaller and functionally informed set of variants compared to the benchmark method, EN. Considering the imperfect match of both ancestry composition and cell-type relevance between MESA discovery data and two external validation data sets, our validation results support the generalizability of our methods across conditions. The higher prediction accuracy observed in JHS compared to Geuvadis may support the importance of cell-type relevance in the prediction accuracy of transcriptome models. Further, our FIV-based methods demonstrated improved multi-ancestry TWAS power and accuracy for a variety of GWAS traits. While the three multi-ancestry GWASs are dominated by individuals of European ancestry, TWAS using multi-ancestry models still demonstrated overall advantages in performance compared to TWAS using EUR-specific models. These findings support the potential of broad applicability of our proposed multi-ancestry prediction models for TWAS based on multi-ancestry GWAS. We hypothesize the benefits of using our multi-ancestry gene expression prediction models for TWAS discovery could be attributed in part to the broader ancestral representation of these models which may alleviate negative impacts of ancestry mismatch observed in prior studies that focused on single-ancestry prediction models.11,12,24 We also showed that the observed improvement in TWAS performance depends on the GWAS trait’s relevance to the tissue or cell-type(s) used to build the models.

The smaller model size from our FIV-based methods is noticeable, especially for EN-FM. EN-FM has a median model size of 3, however, it achieved similar prediction accuracy to EN which has a much larger median model size of 49. This confirms the importance of including and upweighting FIVs in the prediction models. The accuracy of multi-ancestry fine-mapping in our study was affected by sample size, which may affect model performance of EN-FM and PUMICE-FM. As observed in the previous studies of multi-ancestry fine-mapping, differences in LD patterns and association effect sizes across ancestries may contribute to improved fine-mapping.8,9,49 Although our transcriptome prediction models were built on diverse ancestry populations, the overall sample size (i.e., 1,287 multi-ancestry participants) is not large enough to produce powerful multi-ancestry eQTLs and thus affects the accuracy of fine-mapping. Additionally, the original PUMICE16 paper concluded that PUMICE would achieve better performance from using tissue or cell-matched functional annotation data. Due to the unavailability of neither epigenetic processes nor 3D genomic data for PBMCs, we used available data from EBV-transformed lymphocytes as a proxy to construct PUMICE model, which probably affected model performance.

To test the robustness of our FIV-based methods for TWAS, we integrated our models with three large-scale multi-ancestry GWASs from eight blood cell traits,8 five lipid traits,9 and four pulmonary function traits,10 respectively. Our FIV-based methods presented higher power and accuracy for multi-ancestry TWAS, even though no single proposed method uniformly outperformed all tested GWAS traits. The improved power and accuracy from our FIV-based methods on multi-ancestry TWAS might be attributed to the inclusion of FIVs and the upweighting of these variants as well. EN assigns penalties to variants and estimates their weights mathematically without knowing prior importance of variants. Some biologically important variants might be excluded or underweighted from the EN model due to their small coefficients, and the small coefficients might be caused by small MAF. While our FIV-based methods presented improved TWAS performance, the variation in performance across three TWASs was likely due to the GWAS trait’s relevance to PBMCs. Our FIV-based methods produced significantly higher TWAS power than EN for one blood cell trait and two lipid traits, but not for any of four pulmonary function traits. However, our FIV-based methods had higher accuracy of Bonferroni-significant TWAS genes for all tested GWAS traits except for two blood cell traits, BASO and WBC.

Realizing that different transcriptome prediction methods might benefit genes differently due to gene’s underlying architecture (e.g., sparse vs polygenic), and no single method works uniformly well for all genes, we hypothesized that it could be beneficial to further improve TWAS by an omnibus approach that aggregates TWAS summary statistics from our three FIV-based methods. Our proposed omnibus approach produced significantly higher TWAS power for five out of eight blood cell traits, compared to the improvement in only one blood cell trait by a single FIV-based method. Further, among the curated genes, while EN uniquely identified some of them at Bonferroni significance, the omnibus approach identified many more. This result indicates the improved capability of our proposed omnibus approach in identifying trait-relevant genes. As demonstrated in other studies that aggregated TWAS from different methods or tissues increased power,16–18 the omnibus approach could be a trend for future TWAS method that leverages multiple transcriptome prediction methods with different assumptions of underlying genetic architectures. While we observed overall improvements in TWAS performance based on the proposed multi-ancestry prediction methods, we should be aware that our current approach does not allow the identification of significant ancestry-specific TWAS genes as our methods do not have ancestry-shared and ancestry-specific components to differentiate them. We believe this topic will be an important topic for future TWAS research.

Some limitations of our study must be noted. First, the power of our developed transcriptome prediction models is limited due to sample size of multi-ancestry participants and accuracy of FIVs. We believe that our model performance can be substantially improved by increasing both sample size and number of participants from different ancestry groups. With increased sample size, we believe that the accuracy of multi-ancestry fine-mapping would be improved as well, which would directly affect performance of models with fine-mapped variants. The epigenetic and 3D genomic data used to construct PUMICE are not from PBMCs but from a proxy cell-type, EBV-transformed lymphocytes, which would affect model performance. Second, our assessment of TWAS power and accuracy depends on the curated genes that were identified by previous gene prioritization efforts10,35,36 for the same published GWASs that we used to carry out TWAS analyses. We should caution that these curated genes may not be ground truth for GWAS traits. Further, the annotation resources used to identify the curated genes were mainly from individuals of European ancestry, even though the GWAS efforts were carried out as multi-ancestry studies. Accordingly, we should be aware that those curated genes are only the approximations of the “true” causal genes for the corresponding GWAS traits. Third, PBMCs are not the most relevant cell-type for all GWAS traits we examined in the study. Accordingly, we should be cautious in the interpretation of our multi-ancestry TWAS results for traits that have less relevance to PBMC transcriptomes. Future work developing multi-ancestry transcriptome models using disease-relevant tissues may prove useful in identifying biologically relevant genes from multi-ancestry TWAS. We should note that the proposed methods including FIVs in the multi-ancestry transcriptome prediction models can be extended to other tissues straightforwardly with both diverse ancestry transcriptomes and functional annotation data from disease-relevant tissues. Last, although our study shows improvements in TWAS performance by using multi-ancestry versus EUR-specific prediction models in the analysis of EUR-specific GWAS, these findings may arise from increased sample sizes, resolution and variant representation afforded by the multi-ancestry prediction approach, as well as the high representation of European ancestry in the MESA cohort. We should caution that ancestry mismatch between TWAS reference models and GWAS may yield reduced power for TWAS discoveries in some cases, as pointed out by several previous studies of multi-ancestry TWAS.11,12, 24

In conclusion, our study demonstrates the value of including FIVs in the multi-ancestry transcriptome prediction models to improve multi-ancestry TWAS. Additionally, we show that the improvement of TWAS performance depends on GWAS trait’s relevance to tissue or cell-type used to build the models. Our proposed methods leveraging FIVs for gene expression prediction in the context of TWAS can be extended naturally to other tissues and other omics data as well.

Supplementary Material

1
2

The Supplemental information file includes 31 supplemental figures.

We proposed methods leveraging functionally informed variants for multi-ancestry transcriptome prediction and demonstrated improvements on multi-ancestry TWAS performance. Our proposed omnibus approach that aggregates TWAS results further improves TWAS power. The observed TWAS improvement depends on the GWAS trait’s relevance to peripheral blood mononuclear cells used to build our models.

Acknowledgments

This work is supported by R01-HL153248 (AM and XH) and R01-ES036042 (DJL). LMR is funded by R01AG075884. YL is funded by U01HG011720, R01HL146500, and R01HL165061. HEW and DSA are funded by R15HG009569. The authors acknowledge Research Computing at The University of Virginia (https://rc.virginia.edu) for providing computational resources and technical support that have contributed to the results reported within this manuscript. We gratefully acknowledge the cohorts and participants who provided biological samples and data for the Trans-Omics in Precision Medicine (TOPMed) program supported by the National Heart, Lung, and Blood Institute (NHLBI). Study-specific acknowledgements are given in supplemental information. A full list of investigators for the NHLBI TOPMed Consortium is provided at https://topmed.nhlbi.nih.gov/topmed-banner-authorship.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Declaration of interests

SSR and LMR are consultants to the NHLBI TOPMed Administrative Coordinating Center (Westat). MHC has received grant funding from Bayer and consulting fees from Apogee Therapeutics and BMS. Other authors declare no conflict of interests.

Data and code availability

Our developed multi-ancestry TOPMed MESA transcriptome prediction models and code are available at https://zenodo.org/records/18644222.

References

  • 1.Zhu Z, Zhang F, Hu H, Bakshi A, Robinson MR, Powell JE, Montgomery GW, Goddard ME, Wray NR, Visscher PM, et al. (2016). Integration of summary data from GWAS and eQTL studies predicts complex trait gene targets. Nat. Genet 48, 481–487. 10.1038/ng.3538. [DOI] [PubMed] [Google Scholar]
  • 2.Mountjoy E, Schmidt EM, Carmona M, Schwartzentruber J, Peat G, Miranda A, Fumis L, Hayhurst J, Buniello A, Karim MA, et al. (2021). An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nat. Genet 53, 1527–1533. 10.1038/s41588-021-00945-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Barbeira AN, Dickinson SP, Bonazzola R, Zheng J, Wheeler HE, Torres JM, Torstenson ES, Shah KP, Garcia T, Edwards TL, et al. (2018). Exploring the phenotypic consequences of tissue specific gene expression variation inferred from GWAS summary statistics. Nat. Commun 9, 1825. 10.1038/s41467-018-03621-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Gusev A, Mancuso N, Won H, Kousi M, Finucane HK, Reshef Y, Song L, Safi A, McCarroll S, Neale BM, et al. (2018). Transcriptome-wide association study of schizophrenia and chromatin activity yields mechanistic disease insights. Nat. Genet 50, 538–548. 10.1038/s41588-018-0092-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Sun Y, Zhu J, Zhou D, Canchi S, Wu C, Cox NJ, Rissman RA, Gamazon ER, and Wu L (2021). A transcriptome-wide association study of Alzheimer’s disease using prediction models of relevant tissues identifies novel candidate susceptibility genes. Genome Med 13, 141. 10.1186/s13073-021-00959-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Al-Barghouthi BM, Rosenow WT, Du K-P, Heo J, Maynard R, Mesner L, Calabrese G, Nakasone A, Senwar B, Gerstenfeld L, et al. (2022). Transcriptome-wide association study and eQTL colocalization identify potentially causal genes responsible for human bone mineral density GWAS associations. eLife. 10.7554/eLife.77285. [DOI] [Google Scholar]
  • 7.Hu X, Kim JS, Saferali A, Huang Y, Ma S-F, Bingham GC, Bonham CA, Flores C, Castaldi P, Hersh CP, et al. (2024). Transcriptome-Wide Association Study of Idiopathic Pulmonary Fibrosis Survival Identifies PTPN9 and SNRPB2. Am. J. Respir. Crit. Care Med 210, 683–686. 10.1164/rccm.202310-1741LE. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Chen M-H, Raffield LM, Mousas A, Sakaue S, Huffman JE, Moscati A, Trivedi B, Jiang T, Akbari P, Vuckovic D, et al. (2020). Trans-ethnic and Ancestry-Specific Blood-Cell Genetics in 746,667 Individuals from 5 Global Populations. Cell 182, 1198–1213.e14. 10.1016/j.cell.2020.06.045. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Graham SE, Clarke SL, Wu K-HH, Kanoni S, Zajac GJM, Ramdas S, Surakka I, Ntalla I, Vedantam S, Winkler TW, et al. (2021). The power of genetic diversity in genome-wide association studies of lipids. Nature 600, 675–679. 10.1038/s41586-021-04064-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Shrine N, Izquierdo AG, Chen J, Packer R, Hall RJ, Guyatt AL, Batini C, Thompson RJ, Pavuluri C, Malik V, et al. (2023). Multi-ancestry genome-wide association analyses improve resolution of genes and pathways influencing lung function and chronic obstructive pulmonary disease risk. Nat. Genet 55, 410–422. 10.1038/s41588-023-01314-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Bhattacharya A, García-Closas M, Olshan AF, Perou CM, Troester MA, and Love MI (2020). A framework for transcriptome-wide association studies in breast cancer in diverse study populations. Genome Biol 21, 42. 10.1186/s13059-020-1942-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.A B, Jb H, D Z, W Z, J Z, M K, B P, Er G, and Nj C (2022). Best practices for multi-ancestry, meta-analytic transcriptome-wide association studies: Lessons from the Global Biobank Meta-analysis Initiative. Cell Genomics 2. 10.1016/j.xgen.2022.100180. [DOI] [Google Scholar]
  • 13.Gaffney DJ, Veyrieras J-B, Degner JF, Pique-Regi R, Pai AA, Crawford GE, Stephens M, Gilad Y, and Pritchard JK (2012). Dissecting the regulatory architecture of gene expression QTLs. Genome Biol 13, R7. 10.1186/gb-2012-13-1-r7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Zhang W, Voloudakis G, Rajagopal VM, Readhead B, Dudley JT, Schadt EE, Björkegren JLM, Kim Y, Fullard JF, Hoffman GE, et al. (2019). Integrative transcriptome imputation reveals tissue-specific and shared biological mechanisms mediating susceptibility to complex traits. Nat. Commun 10, 3834. 10.1038/s41467-019-11874-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Barbeira AN, Melia OJ, Liang Y, Bonazzola R, Wang G, Wheeler HE, Aguet F, Ardlie KG, Wen X, and Im HK (2020). Fine-mapping and QTL tissue-sharing information improves the reliability of causal gene identification. Genet. Epidemiol 44, 854–867. 10.1002/gepi.22346. [DOI] [Google Scholar]
  • 16.Khunsriraksakul C, McGuire D, Sauteraud R, Chen F, Yang L, Wang L, Hughey J, Eckert S, Dylan Weissenkampen J, Shenoy G, et al. (2022). Integrating 3D genomic and epigenomic data to enhance target gene discovery and drug repurposing in transcriptome-wide association studies. Nat. Commun 13, 3258. 10.1038/s41467-022-30956-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Dai Q, Zhou G, Zhao H, Võsa U, Franke L, Battle A, Teumer A, Lehtimäki T, Raitakari OT, Esko T, et al. (2023). OTTERS: a powerful TWAS framework leveraging summary-level reference data. Nat. Commun 14, 1271. 10.1038/s41467-023-36862-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Hu T, Parrish RL, Dai Q, Buchman AS, Tasaki S, Bennett DA, Seyfried NT, Epstein MP, and Yang J (2024). Omnibus proteome-wide association study identifies 43 risk genes for Alzheimer disease dementia. Am. J. Hum. Genet 111, 1848–1863. 10.1016/j.ajhg.2024.07.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Gamazon ER, Wheeler HE, Shah KP, Mozaffari SV, Aquino-Michaels K, Carroll RJ, Eyler AE, Denny JC, Nicolae DL, Cox NJ, et al. (2015). A gene-based association method for mapping traits using reference transcriptome data. Nat. Genet 47, 1091–1098. 10.1038/ng.3367. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Lappalainen T, Sammeth M, Friedländer MR, ‘t Hoen PAC, Monlong J, Rivas MA, Gonzàlez-Porta M, Kurbatova N, Griebel T, Ferreira PG, et al. (2013). Transcriptome and genome sequencing uncovers functional variation in humans. Nature 501, 506–511. 10.1038/nature12531. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Wilson JG, Rotimi CN, Ekunwe L, Royal CDM, Crump ME, Wyatt SB, Steffes MW, Adeyemo A, Zhou J, Taylor HA, et al. (2005). Study design for genetic analysis in the Jackson Heart Study. Ethn. Dis 15, S6–30–37. [Google Scholar]
  • 22.Bild DE, Bluemke DA, Burke GL, Detrano R, Diez Roux AV, Folsom AR, Greenland P, Jacob DR, Kronmal R, Liu K, et al. (2002). Multi-Ethnic Study of Atherosclerosis: objectives and design. Am. J. Epidemiol 156, 871–881. 10.1093/aje/kwf113. [DOI] [PubMed] [Google Scholar]
  • 23.Taliun D, Harris DN, Kessler MD, Carlson J, Szpiech ZA, Torres R, Taliun SAG, Corvelo A, Gogarten SM, Kang HM, et al. (2021). Sequencing of 53,831 diverse genomes from the NHLBI TOPMed Program. Nature 590, 290–299. 10.1038/s41586-021-03205-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Araujo DS, Nguyen C, Hu X, Mikhaylova AV, Gignoux C, Ardlie K, Taylor KD, Durda P, Liu Y, Papanicolaou G, et al. (2023). Multivariate adaptive shrinkage improves cross-population transcriptome prediction and association studies in underrepresented populations. Hum. Genet. Genomics Adv 4, 100216. 10.1016/j.xhgg.2023.100216. [DOI] [Google Scholar]
  • 25.Yang C, Hallmark B, Chai JC, O’Connor TD, Reynolds LM, Wood AC, Seeds M, Chen Y-DI, Steffen LM, Tsai MY, et al. (2021). Impact of Amerind ancestry and FADS genetic variation on omega-3 deficiency and cardiometabolic traits in Hispanic populations. Commun. Biol 4, 1–12. 10.1038/s42003-021-02431-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Friedman J, Hastie T, and Tibshirani R (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. J. Stat. Softw 33, 1–22. [PMC free article] [PubMed] [Google Scholar]
  • 27.Wang G, Sarkar A, Carbonetto P, and Stephens M (2020). A Simple New Approach to Variable Selection in Regression, with Application to Genetic Fine Mapping. J. R. Stat. Soc. Ser. B Stat. Methodol 82, 1273–1300. 10.1111/rssb.12388. [DOI] [Google Scholar]
  • 28.Consortium, T.E.P. (2011). A User’s Guide to the Encyclopedia of DNA Elements (ENCODE). PLOS Biol 9, e1001046. 10.1371/journal.pbio.1001046. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Dunham I, Kundaje A, Aldred SF, Collins PJ, Davis CA, Doyle F, Epstein CB, Frietze S, Harrow J, Kaul R, et al. (2012). An integrated encyclopedia of DNA elements in the human genome. Nature 489, 57–74. 10.1038/nature11247. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Rao SSP, Huntley MH, Durand NC, Stamenova EK, Bochkov ID, Robinson JT, Sanborn AL, Machol I, Omer AD, Lander ES, et al. (2014). A 3D map of the human genome at kilobase resolution reveals principles of chromatin looping. Cell 159, 1665–1680. 10.1016/j.cell.2014.11.021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Jung I, Schmitt A, Diao Y, Lee AJ, Liu T, Yang D, Tan C, Eom J, Chan M, Chee S, et al. (2019). A compendium of promoter-centered long-range chromatin interactions in the human genome. Nat. Genet 51, 1442–1449. 10.1038/s41588-019-0494-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Casper J, Speir ML, Raney BJ, Perez G, Nassar LR, Lee CM, Hinrichs AS, Gonzalez JN, Fischer C, Diekhans M, et al. (2026). The UCSC Genome Browser database: 2026 update. Nucleic Acids Res 54, D1331–D1335. 10.1093/nar/gkaf1250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Wen J, Sun Q, Huang L, Zhou L, Doyle MF, Ekunwe L, Durda P, Olson NC, Reiner AP, Li Y, et al. (2024). Gene expression and splicing QTL analysis of blood cells in African American participants from the Jackson Heart Study. Genetics 228, iyae098. 10.1093/genetics/iyae098. [DOI] [Google Scholar]
  • 34.Cerezo M, Sollis E, Ji Y, Lewis E, Abid A, Bircan KO, Hall P, Hayhurst J, John S, Mosaku A, et al. (2025). The NHGRI-EBI GWAS Catalog: standards for reusability, sustainability and diversity. Nucleic Acids Res 53, D998–D1005. 10.1093/nar/gkae1070. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Sun Q, Crowley CA, Huang L, Wen J, Chen J, Bao EL, Auer PL, Lettre G, Reiner AP, Sankaran VG, et al. (2022). From GWAS variant to function: A study of ~148,000 variants for blood cell traits. HGG Adv 3, 100063. 10.1016/j.xhgg.2021.100063. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Kanoni S, Graham SE, Wang Y, Surakka I, Ramdas S, Zhu X, Clarke SL, Bhatti KF, Vedantam S, Winkler TW, et al. (2022). Implicating genes, pleiotropy, and sexual dimorphism at blood lipid loci through multi-ancestry meta-analysis. Genome Biol 23, 268. 10.1186/s13059-022-02837-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Liu Y, Chen S, Li Z, Morrison AC, Boerwinkle E, and Lin X (2019). ACAT: A Fast and Powerful p Value Combination Method for Rare-Variant Analysis in Sequencing Studies. Am. J. Hum. Genet 104, 410–421. 10.1016/j.ajhg.2019.01.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.van Iterson M, van Zwet EW, Heijmans BT, and the BIOS Consortium (2017). Controlling bias and inflation in epigenome- and transcriptome-wide association studies using the empirical null distribution. Genome Biol 18, 19. 10.1186/s13059-016-1131-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Liang Y, Nyasimi F, and Im HK (2026). A gene-specific variance-control approach corrects polygenicity-driven inflation observed in transcriptome-wide association studies. Am. J. Hum. Genet 113, 276–290. 10.1016/j.ajhg.2025.12.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Edilova MI, Abdul-Sater AA, and Watts TH (2018). TRAF1 Signaling in Human Health and Disease. Front. Immunol 9. 10.3389/fimmu.2018.02969. [DOI] [Google Scholar]
  • 41.Tsitsikov EN, Laouini D, Dunn IF, Sannikova TY, Davidson L, Alt FW, and Geha RS (2001). TRAF1 Is a Negative Regulator of TNF Signaling: Enhanced TNF Signaling in TRAF1-Deficient Mice. Immunity 15, 647–657. 10.1016/S1074-7613(01)00207-2. [DOI] [PubMed] [Google Scholar]
  • 42.Oyoshi MK, Barthel R, and Tsitsikov EN (2007). TRAF1 regulates recruitment of lymphocytes and, to a lesser extent, neutrophils, myeloid dendritic cells and monocytes to the lung airways following lipopolysaccharide inhalation. Immunology 120, 303–314. 10.1111/j.1365-2567.2006.02499.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Florvall G, Basu S, and Larsson A (2006). Apolipoprotein A1 is a stronger prognostic marker than are HDL and LDL cholesterol for cardiovascular disease and mortality in elderly men. J. Gerontol. A. Biol. Sci. Med. Sci 61, 1262–1266. 10.1093/gerona/61.12.1262. [DOI] [PubMed] [Google Scholar]
  • 44.Rhee E-J, Byrne CD, and Sung K-C (2017). The HDL cholesterol/apolipoprotein AI ratio: an indicator of cardiovascular disease. Curr. Opin. Endocrinol. Diabetes Obes 24, 148–153. 10.1097/MED.0000000000000315. [DOI] [PubMed] [Google Scholar]
  • 45.Bhale AS, and Venkataraman K (2022). Leveraging knowledge of HDLs major protein ApoA1: Structure, function, mutations, and potential therapeutics. Biomed. Pharmacother 154, 113634. 10.1016/j.biopha.2022.113634. [DOI] [PubMed] [Google Scholar]
  • 46.Morty RE, Königshoff M, and Eickelberg O (2009). Transforming growth factor-beta signaling across ages: from distorted lung development to chronic obstructive pulmonary disease. Proc. Am. Thorac. Soc 6, 607–613. 10.1513/pats.200908-087RM. [DOI] [PubMed] [Google Scholar]
  • 47.Hersh CP, Hansel NN, Barnes KC, Lomas DA, Pillai SG, Coxson HO, Mathias RA, Rafaels NM, Wise RA, Connett JE, et al. (2009). Transforming growth factor-beta receptor-3 is associated with pulmonary emphysema. Am. J. Respir. Cell Mol. Biol 41, 324–331. 10.1165/rcmb.2008-0427OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Sunny SK, Zhang H, Relton CL, Ring S, Kadalayil L, Mzayek F, Ewart S, Holloway JW, and Arshad SH (2021). Sex-specific longitudinal association of DNA methylation with lung function. ERJ Open Res 7, 00127–02021. 10.1183/23120541.00127-2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Lake J, Warly Solsberg C, Kim JJ, Acosta-Uribe J, Makarious MB, Li Z, Levine K, Heutink P, Alvarado CX, Vitale D, et al. (2023). Multi-ancestry meta-analysis and fine-mapping in Alzheimer’s disease. Mol. Psychiatry 28, 3121–3132. 10.1038/s41380-023-02089-w. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS2161406-supplement-1.pdf (3.2MB, pdf)
2
NIHMS2161406-supplement-2.xlsx (385.3KB, xlsx)

Data Availability Statement

Our developed multi-ancestry TOPMed MESA transcriptome prediction models and code are available at https://zenodo.org/records/18644222.

RESOURCES