Transcript-aware rare genetic variant association analyses of cardiopulmonary traits in participants from the All of Us Research Program Authors: Jingwen Zhang, So-hyeon Hong, Xin Wang, Sean J. Jurgens, Ching-Ti Liu, Josée Dupuis, Patrick T. Ellinor, George T. O’Connor, Quanshun Mei, Seung Hoan Choi Published: 2026 ABSTRACT Gene-based rare variant analyses often lack statistical power and may overlook transcript-specific effects. Here, we present a transcript-aware aggregation framework. In simulation studies, the framework maintains appropriate false-positive rates and shows competitive power relative to standard single-transcript analyses, approaching the performance of the ideal case of knowing the most informative transcript in advance. We then apply the approach to 129 cardiopulmonary traits in over 240,000 whole-genome-sequenced All of Us participants. By leveraging transcript-specific annotations, we identify 11 novel associations and recover 47 reported associations, including potentially pleiotropic genes linked to plasma lipid traits (PPARG) and body habitus (TCF12). Notably, for TTN, a gene known for its transcript-specific effects in cardiomyopathy, our framework strengthens the association signal and pinpoints the N2B isoform, which shows a stronger association with cardiomyopathy than other transcripts. These findings highlight the value of a transcript-aware framework for improving rare variant association studies. Introduction Advancements in high-throughput sequencing technologies have increased the availability of large-scale genomic datasets. The availability of whole-genome sequencing (WGS) and whole-exome sequencing (WES) datasets, such as those from the Trans Omics for Precision Medicine (TOPMed) Program1 of the National Heart, Lung and Blood Institute (NHLBI), the UK Biobank WGS/WES Program2, and the All of Us Research Program3, provides great opportunities for discovery. These datasets enable the identification of rare coding variants with large effects, such as predicted loss-of-function (LoF) and predicted deleterious missense (pDM) variants. Among these programs, the All of Us Research Program, initiated by the National Institutes of Health (NIH), has been one of the largest research resources for studying human diseases. The diversity in the All of Us population provides valuable insights to enhance our understanding of groups underrepresented in biomedical research4. In addition, the genomic data from All of Us serve as a reliable resource for discovering causal genes associated with various diseases5,6. Single-variant testing has been widely used to identify associations for common variants. However, this approach becomes less effective for rare variant association analysis because of reduced power due to the low frequency of minor alleles in rare variants5,7. To overcome this limitation, gene-based aggregation testing methods have been developed, including the burden test8,9 and the sequence kernel association test (SKAT)10,11. These methods often aggregate LoF or missense variants that disrupt protein production12–14, which are typically based on canonical transcripts or the most severe consequence of known protein-coding transcripts15. However, single-transcript-based approaches may fail to capture the functional contributions of alternative splicing or non-canonical transcripts16,17. Ignoring transcript-specific effects can reduce statistical power by excluding functional variations in non-canonical or tissue-specific transcripts, or by including biologically irrelevant variants. A notable example is TTN, a gene implicated in cardiomyopathy and atrial fibrillation, which produces multiple functionally distinct transcripts. The burden of LoF alleles in TTN substantially increases the risk of atrial fibrillation, with a higher prevalence of carriers observed in early-onset patients12. After analyzing rare variants in cardiac-specific transcripts, the statistical significance of TTN-disease associations improved substantially18. This highlights the potential and need for a transcript-aware approach in gene-trait association analysis. In this study, we developed a transcript-aware rare variant aggregation framework to facilitate the discovery of disease-associated genes. We conducted extensive simulations to evaluate the performance of the proposed framework across different scenarios and causal variant configurations in rare variant analysis settings. We then applied the proposed framework to perform rare variant analyses using variants in exonic regions from the WGS data of over 240,000 All of Us participants. Our study focuses on cardiopulmonary traits, given the substantial and growing burden of cardiopulmonary disease worldwide19–21. An overview of the real data application design and analysis workflow is summarized in Fig. 1.Fig. 1Study flowchart.*The exact sample size for the study population depends on the corresponding phenotypes. The proposed approach is outlined in the dashed box. Statistical analyses were conducted for each transcript from each gene. The resulting test statistics were then combined using the Cauchy combination test to obtain a single P value per gene. Abbreviations: WGS, whole-genome sequencing; LoF, loss-of-function. Study flowchart. *The exact sample size for the study population depends on the corresponding phenotypes. The proposed approach is outlined in the dashed box. Statistical analyses were conducted for each transcript from each gene. The resulting test statistics were then combined using the Cauchy combination test to obtain a single P value per gene. Abbreviations: WGS, whole-genome sequencing; LoF, loss-of-function. Results and discussion Simulation study Simulation results comparing four approaches (canonical-only, pseudo-only, transcript 3 [T3]-only, and the proposed transcript-aware [T-aware] method) across various causal configurations were summarized in Fig. 2 and Supplementary Fig. 1. At α = 0.001, all methods showed well-controlled type I error. The T-aware method kept the false positive rate close to the nominal level, while the T3-only approach was slightly lower than nominal (Fig. 2a). At α = 0.01 (Supplementary Fig. 1a), type I error rates were close to nominal for all methods, with overlapping confidence intervals (CIs). Similar patterns were observed at α = 0.05 (Supplementary Fig. 1b), with the T3-only approach slightly more conservative and T-aware slightly above the nominal level. To further assess these behaviors, we performed additional simulations under a higher outcome prevalence of 2% (Supplementary Fig. 2). Consistent with the results under 0.2% prevalence, all methods controlled type I error well at α = 0.001 and α = 0.01, with T3-only still slightly conservative and T-aware slightly above nominal at α = 0.05. The pattern of type I error estimates at α = 0.05 may reflect both Monte Carlo variation from the finite number of replicates and the difficulty of calibration under case-control imbalance in rare variant analyses.Fig. 2Type I error and power under different simulation settings.a Estimated type I error rates for each method at α = 0.001 in causal configuration 1, with 95% confidence intervals (CIs). b Power of each method in causal configuration 2 as the percentage of causal variants assigned to the canonical transcript increases (30%, 60%, 80%, and 100%) at α = 0.001, with 95% CIs. Transcript 3 is the most outcome-relevant transcript. c Power of each method in configuration 3 as the total number of causal variants increases (10, 20, and 30) at α = 0.001, with 95% CIs. Transcript 3 does not contain causal variants. d Power of each method in configuration 4 as the total number of causal variants increases at α = 0.001, with 95% CIs. In all panels, T3 represents the test using only transcript 3. Canonical represents the test using only the canonical transcript. Pseudo represents the test using only the pseudo transcript. T-aware is the proposed approach. Points from panels a to d represent empirical type I error rates (panel a) or power (panels b–d). Error bars represent 95% CIs. All results are summarized from independent simulations with n = 200,000 individuals and 20,000 replicates per scenario. Source data are provided as a Source Data file. Type I error and power under different simulation settings. a Estimated type I error rates for each method at α = 0.001 in causal configuration 1, with 95% confidence intervals (CIs). b Power of each method in causal configuration 2 as the percentage of causal variants assigned to the canonical transcript increases (30%, 60%, 80%, and 100%) at α = 0.001, with 95% CIs. Transcript 3 is the most outcome-relevant transcript. c Power of each method in configuration 3 as the total number of causal variants increases (10, 20, and 30) at α = 0.001, with 95% CIs. Transcript 3 does not contain causal variants. d Power of each method in configuration 4 as the total number of causal variants increases at α = 0.001, with 95% CIs. In all panels, T3 represents the test using only transcript 3. Canonical represents the test using only the canonical transcript. Pseudo represents the test using only the pseudo transcript. T-aware is the proposed approach. Points from panels a to d represent empirical type I error rates (panel a) or power (panels b–d). Error bars represent 95% CIs. All results are summarized from independent simulations with n = 200,000 individuals and 20,000 replicates per scenario. Source data are provided as a Source Data file. When causal variants were concentrated in the non-canonical transcript T3 (causal configuration 2), the T-aware method was more powerful than the canonical-only and pseudo-only approaches across all scenarios, with non-overlapping CIs for the estimated power (Fig. 2b and Supplementary Fig. 1c, d). The difference in power was larger when the canonical transcript contained fewer causal variants. Notably, even when the canonical transcript contained all causal variants, the canonical-only approach was still less powerful because it included many additional non-causal variants. As expected, the highest power was achieved when the most outcome-relevant transcript (T3) was known in advance and analyzed alone. When all causal variants were concentrated in the canonical transcript, and T3 served as a noise transcript (causal configuration 3), the canonical-only approach outperformed T-aware (Fig. 2c and Supplementary Fig. 1e, f), because the canonical transcript was the most informative one under this configuration. However, T-aware remained competitive and achieved the second-highest power, because this configuration can be treated as a special case of configuration 2, where a canonical transcript was the most outcome-relevant transcript instead. In causal configuration 4, where causal variants were randomly distributed across transcripts, the benefit of leveraging multiple transcripts became more apparent. As expected, the pseudo-only approach showed the highest power most of the time, while T-aware had the second-highest power, and the difference between them depended on the total number of causal variants (Fig. 2d and Supplementary Fig. 1g, h). Overall, these results showed that power depended on multiple factors, including the causal configuration under our simulation design, while T-aware remained competitive with the ideal single-transcript approach across settings. More importantly, because the truly outcome-relevant transcript and the distribution of causal variants are generally unknown in reality, T-aware may serve as a practical surrogate for the ideal approach. Outside the simulation setting, we examined the behavior of the final combined P value from the proposed framework under an extreme scenario where only one transcript was significant, and all others were non-significant (Supplementary Methods). We found that adding more non-significant transcripts gradually reduced the strength of the combined signal. However, the magnitude of this decrease depended heavily on both the smallest P value in the set and the magnitudes of the remaining P values (Supplementary Fig. 3 and Supplementary Note 1). With an extremely low P value, the combined result remained significant even after many additional transcripts were included, suggesting that the framework is reasonably robust to dilution by non-significant transcripts. Although this extreme scenario is unlikely to occur in practice, it provides useful insight into the behavior of the proposed framework. Transcript-aware analysis highlighted TTN and pinpointed the N2B isoform We analyzed 118 independent, clinically relevant binary phenotypes related to circulatory and respiratory conditions, as well as 11 inverse-rank-normalized quantitative traits from All of Us participants. After correcting for multiple testing using the Benjamini–Hochberg procedure, we observed 58 significant gene-phenotype associations across 129 traits with a q-value < 0.05 (Figs. 3, 4, Table 1, and Supplementary Data 1, 2). The proposed framework successfully reproduced 47 previously reported rare-variant associations, confirming proper calibration and consistency with prior studies. Notably, TTN was associated with multiple traits, including cardiomyopathy (Combined P value[Pcombined] = 1.4 × 10−27), atrial fibrillation and flutter (Pcombined = 2.0 × 10−11), congestive heart failure (Pcombined = 1.2 × 10−12), heart rate (Pcombined = 5.3 × 10−15), and systolic blood pressure (SBP; Pcombined = 9.5 × 10−7), with additional significant associations detected (Supplementary Data 1, 2, Supplementary Note 1).Fig. 3Manhattan plots of exome-wide rare variant analysis.a Manhattan plot for 118 binary phenotypes. The y-axis shows –log10 of the combined P value (Pcombined) from the proposed transcript-aware framework. For clarity, names of binary phenotypes are replaced by the corresponding phecodes on the x-axis. b Manhattan plot for continuous traits using Pcombined from the proposed framework. The x-axis displays the names of the corresponding continuous traits. In both panels, Pcombined was calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. Plotted P values are unadjusted, and multiple testing correction was applied to define the genome-wide significance threshold. The red line represents the genome-wide significance threshold (P value ≈≈ 1.5 × 10−6) under a Benjamini–Hochberg false discovery rate (FDR) of 5% across all tests. Abbreviations: BMI, body mass index; LDL, low-density lipoprotein cholesterol in serum or plasma; CRP, C-reactive protein in serum or plasma; DBP, diastolic blood pressure; HDL, high-density lipoprotein cholesterol in serum or plasma; SBP, systolic blood pressure; WHR, waist-to-hip ratio. Source data are provided as a Source Data file.Fig. 4Significant associations identified under the transcript-aware testing framework.a Significant associations by transcript type for binary traits. b Significant associations by transcript type for continuous traits. In both panels, the x-axis displays -log10(P value) for the corresponding associations. To better illustrate the differences in results, P values from analyses using different transcript types are presented in different shapes and colors: red diamond, results from the proposed method; blue square, canonical transcript; black star, pseudo transcript; green triangle, other transcript types. For the proposed method, P values were calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. For canonical, pseudo, and other transcripts, P values were obtained from two-sided rare variant burden tests with SPA adjustment. Plotted P values are unadjusted, and multiple testing correction was applied to define the genome-wide significance threshold. The dashed vertical line represents the genome-wide significance threshold (P value ≈≈ 1.5 × 10−6) under a Benjamini–Hochberg false discovery rate (FDR) of 5% across all tests. Abbreviations: NOS, not otherwise specified; BMI, body mass index; LDL, low-density lipoprotein cholesterol in serum or plasma; HDL, high-density lipoprotein cholesterol in serum or plasma; SBP, systolic blood pressure; WHR, waist-to-hip ratio. Source data are provided as a Source Data file.Table 1Novel associations identified from the proposed testing frameworkTraitPhecode/Concept idPcombined (Ntranscript)GeneTop transcriptTranscript IDEffect/OR (95%CI)Carrier/CasesCarrier/ControlsPtranscriptBinary phenotypesAtherosclerosis of renal artery440.11.2E-06 (2)CYSTM1ENST0000064407835.87 (13.05,92.08)≤20/34723/217,5691.2E-06Encounter for long-term (current) use of anticoagulants, antithrombotics, aspirin4571.0E-06 (4)CCNE2C-ENST00000308108, ENST000005205096.36 (3.34,11.5)≤20/1210058/186,1775.5E-07Occlusion and stenosis of precerebral arteries433.11.7E-07 (9)LDLRC-ENST00000558518, ENST000002524444.78 (2.88,7.49)≤20/3625353/223,2464.0E-08Stricture of artery447.13.9E-07 (4)POLNC-ENST00000511885, ENST000003828655.43 (3.24,8.57)≤20/501720/217,5693.0E-07Transcript IDEffect/OR (95%CI)Carrier/All samplesPtranscriptQuantitative phenotypesBMI9031248.3E-09 (4)YLPM1Pseudo0.52 (0.35,0.69)97/237,1772.9E-091.2E-06 (11)TCF12ENST000005596090.52 (0.32,0.72)93/237,1773.1E-07WHR–3.2E-07 (11)TCF12Pseudo0.47 (0.3,0.64)98/218,6076.1E-081.1E-06 (2)MIB1C-ENST00000261537, Pseudo0.15 (0.09,0.21)760/218,6071.1E-06Heart rate9031261.1E-08 (5)NAA15Pseudo0.69 (0.46,0.92)69/237,8093.0E-091.3E-08 (8)PTENENST000007000210.85 (0.57,1.14)45/237,8094.0E-09HDL30070701.0E-06 (2)SORCS3C-ENST00000369701, Pseudo−0.76 (−1.06,−0.45)28/96,9081.0E-06Novel associations with a false discovery rate (FDR) q-value < 0.05 (P ≈≈ 1.5 × 10−6) among 58 identified gene-trait associations after multiple testing correction. Values for Pcombined represent the combined test results across all transcripts using the proposed framework, and Ntranscript is the number of transcripts available for analysis (including pseudo; details on Ntranscript are provided in Supplementary Note 1). Pcombined was calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. For the top transcript (the transcript with the smallest P value), we present the transcript ID, odds ratios (ORs) or effects in standard deviation units, 95% confidence intervals (CIs), the number of carriers among cases and controls, and transcript-specific P values (Ptranscript). Ptranscript are P values from two-sided rare variant burden tests with SPA adjustment. For continuous phenotypes, the number of carriers of the top transcript and the sample size are presented. The ORs and 95% CIs of the top transcripts were obtained from Firth’s regression models. The prefix ‘C’ is added to transcript IDs to distinguish canonical transcripts from non-canonical ones. If more than one transcript has the smallest P value, all corresponding transcript IDs are included. To comply with the All of Us privacy policy, exact carrier counts below 20 are not reported. Concept ID is not available for WHR because WHR is calculated from waist circumference and hip circumference. Abbreviations: BMI, body mass index; WHR, waist-to-hip ratio; HDL, high-density lipoprotein cholesterol in serum or plasma. Manhattan plots of exome-wide rare variant analysis. a Manhattan plot for 118 binary phenotypes. The y-axis shows –log10 of the combined P value (Pcombined) from the proposed transcript-aware framework. For clarity, names of binary phenotypes are replaced by the corresponding phecodes on the x-axis. b Manhattan plot for continuous traits using Pcombined from the proposed framework. The x-axis displays the names of the corresponding continuous traits. In both panels, Pcombined was calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. Plotted P values are unadjusted, and multiple testing correction was applied to define the genome-wide significance threshold. The red line represents the genome-wide significance threshold (P value ≈≈ 1.5 × 10−6) under a Benjamini–Hochberg false discovery rate (FDR) of 5% across all tests. Abbreviations: BMI, body mass index; LDL, low-density lipoprotein cholesterol in serum or plasma; CRP, C-reactive protein in serum or plasma; DBP, diastolic blood pressure; HDL, high-density lipoprotein cholesterol in serum or plasma; SBP, systolic blood pressure; WHR, waist-to-hip ratio. Source data are provided as a Source Data file. Significant associations identified under the transcript-aware testing framework. a Significant associations by transcript type for binary traits. b Significant associations by transcript type for continuous traits. In both panels, the x-axis displays -log10(P value) for the corresponding associations. To better illustrate the differences in results, P values from analyses using different transcript types are presented in different shapes and colors: red diamond, results from the proposed method; blue square, canonical transcript; black star, pseudo transcript; green triangle, other transcript types. For the proposed method, P values were calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. For canonical, pseudo, and other transcripts, P values were obtained from two-sided rare variant burden tests with SPA adjustment. Plotted P values are unadjusted, and multiple testing correction was applied to define the genome-wide significance threshold. The dashed vertical line represents the genome-wide significance threshold (P value ≈≈ 1.5 × 10−6) under a Benjamini–Hochberg false discovery rate (FDR) of 5% across all tests. Abbreviations: NOS, not otherwise specified; BMI, body mass index; LDL, low-density lipoprotein cholesterol in serum or plasma; HDL, high-density lipoprotein cholesterol in serum or plasma; SBP, systolic blood pressure; WHR, waist-to-hip ratio. Source data are provided as a Source Data file. Table 1Novel associations identified from the proposed testing frameworkTraitPhecode/Concept idPcombined (Ntranscript)GeneTop transcriptTranscript IDEffect/OR (95%CI)Carrier/CasesCarrier/ControlsPtranscriptBinary phenotypesAtherosclerosis of renal artery440.11.2E-06 (2)CYSTM1ENST0000064407835.87 (13.05,92.08)≤20/34723/217,5691.2E-06Encounter for long-term (current) use of anticoagulants, antithrombotics, aspirin4571.0E-06 (4)CCNE2C-ENST00000308108, ENST000005205096.36 (3.34,11.5)≤20/1210058/186,1775.5E-07Occlusion and stenosis of precerebral arteries433.11.7E-07 (9)LDLRC-ENST00000558518, ENST000002524444.78 (2.88,7.49)≤20/3625353/223,2464.0E-08Stricture of artery447.13.9E-07 (4)POLNC-ENST00000511885, ENST000003828655.43 (3.24,8.57)≤20/501720/217,5693.0E-07Transcript IDEffect/OR (95%CI)Carrier/All samplesPtranscriptQuantitative phenotypesBMI9031248.3E-09 (4)YLPM1Pseudo0.52 (0.35,0.69)97/237,1772.9E-091.2E-06 (11)TCF12ENST000005596090.52 (0.32,0.72)93/237,1773.1E-07WHR–3.2E-07 (11)TCF12Pseudo0.47 (0.3,0.64)98/218,6076.1E-081.1E-06 (2)MIB1C-ENST00000261537, Pseudo0.15 (0.09,0.21)760/218,6071.1E-06Heart rate9031261.1E-08 (5)NAA15Pseudo0.69 (0.46,0.92)69/237,8093.0E-091.3E-08 (8)PTENENST000007000210.85 (0.57,1.14)45/237,8094.0E-09HDL30070701.0E-06 (2)SORCS3C-ENST00000369701, Pseudo−0.76 (−1.06,−0.45)28/96,9081.0E-06Novel associations with a false discovery rate (FDR) q-value < 0.05 (P ≈≈ 1.5 × 10−6) among 58 identified gene-trait associations after multiple testing correction. Values for Pcombined represent the combined test results across all transcripts using the proposed framework, and Ntranscript is the number of transcripts available for analysis (including pseudo; details on Ntranscript are provided in Supplementary Note 1). Pcombined was calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. For the top transcript (the transcript with the smallest P value), we present the transcript ID, odds ratios (ORs) or effects in standard deviation units, 95% confidence intervals (CIs), the number of carriers among cases and controls, and transcript-specific P values (Ptranscript). Ptranscript are P values from two-sided rare variant burden tests with SPA adjustment. For continuous phenotypes, the number of carriers of the top transcript and the sample size are presented. The ORs and 95% CIs of the top transcripts were obtained from Firth’s regression models. The prefix ‘C’ is added to transcript IDs to distinguish canonical transcripts from non-canonical ones. If more than one transcript has the smallest P value, all corresponding transcript IDs are included. To comply with the All of Us privacy policy, exact carrier counts below 20 are not reported. Concept ID is not available for WHR because WHR is calculated from waist circumference and hip circumference. Abbreviations: BMI, body mass index; WHR, waist-to-hip ratio; HDL, high-density lipoprotein cholesterol in serum or plasma. Novel associations identified from the proposed testing framework Novel associations with a false discovery rate (FDR) q-value < 0.05 (P ≈≈ 1.5 × 10−6) among 58 identified gene-trait associations after multiple testing correction. Values for Pcombined represent the combined test results across all transcripts using the proposed framework, and Ntranscript is the number of transcripts available for analysis (including pseudo; details on Ntranscript are provided in Supplementary Note 1). Pcombined was calculated using the Cauchy combination test based on two-sided rare variant burden test P values with saddle-point approximation (SPA) adjustment. For the top transcript (the transcript with the smallest P value), we present the transcript ID, odds ratios (ORs) or effects in standard deviation units, 95% confidence intervals (CIs), the number of carriers among cases and controls, and transcript-specific P values (Ptranscript). Ptranscript are P values from two-sided rare variant burden tests with SPA adjustment. For continuous phenotypes, the number of carriers of the top transcript and the sample size are presented. The ORs and 95% CIs of the top transcripts were obtained from Firth’s regression models. The prefix ‘C’ is added to transcript IDs to distinguish canonical transcripts from non-canonical ones. If more than one transcript has the smallest P value, all corresponding transcript IDs are included. To comply with the All of Us privacy policy, exact carrier counts below 20 are not reported. Concept ID is not available for WHR because WHR is calculated from waist circumference and hip circumference. Abbreviations: BMI, body mass index; WHR, waist-to-hip ratio; HDL, high-density lipoprotein cholesterol in serum or plasma. Interestingly, the strongest signals for the binary traits were localized to the N2B cardiac isoform (ENST00000460472). The TTN gene encodes titin, a large muscle protein with two main cardiac isoforms: N2B and N2BA (ENST00000591111). N2B is the shorter and stiffer isoform, and its transcript is fully nested within the longer N2BA transcript22. This N2B isoform is highly expressed in the adult human heart and contributes to myocardial passive stiffness through shortened elastic elements, particularly the PEVK region. Disruption of this isoform impairs sarcomere mechanics and has been implicated in the development of dilated cardiomyopathy23. In our study, results from N2B showed stronger signals compared to N2BA. Specifically in cardiomyopathy, the N2B isoform exhibited a stronger association signal, with a larger odds ratio (OR 3.06, 95% CI [2.58, 3.59]) and a much smaller P value (Ptranscript = 2.3 × 10−28) than N2BA (OR 2.38, 95% CI [2.03, 2.77], Ptranscript = 7.5 × 10−21), while the CIs partially overlapped. This distinction likely reflects the fact that LoF variants in N2B affect both isoforms, whereas variants specific to N2BA do not impact N2B expression. Developmentally, N2BA dominates in fetal hearts, but in adults, the N2BA:N2B ratio shifts to ~30:70, favoring N2B23. These findings further emphasize the importance of transcript-specific interpretation in cardiomyopathy24, and thus imply a significant impact that transcript-aware analysis can have in uncovering biologically meaningful gene-phenotype associations. Transcript-aware analysis recovered a biologically plausible PPARG signal with potential pleiotropic effects For PPARG, a non-canonical transcript (ENST00000652522) showed a strong association with triglycerides (β = 1.10, 95% CI [0.72, 1.48], Ptranscript = 1.8 × 10−8). This is consistent with prior findings near PPARG from a multi-ancestry genome-wide association (GWAS) meta-analysis of lipid traits25. Previous studies have also linked PPARG and the broader PPAR pathway to triglyceride biology and metabolic regulation26–28. In addition to triglycerides, PPARG was associated with high-density lipoprotein cholesterol (HDL) in both this study and a previous study29, reinforcing its role in lipid metabolism. PPARG encodes peroxisome proliferator-activated receptor gamma, a protein that modulates lipid metabolism. The leave-one-variant-out (LOVO) analysis revealed that a shared pDM variant, rs1354592503 (risk allele G), had the largest influence on both the triglycerides and HDL associations, suggesting a potential pleiotropic effect (Supplementary Fig. 4). This missense variant is listed in ClinVar30 as pathogenic/likely pathogenic and has been reported in PPARG-related familial partial lipodystrophy31,32, a disorder characterized by impaired adipose storage and metabolic complications that commonly include hypertriglyceridemia. Together, these results support the biological plausibility of the PPARG signal identified by the proposed framework. Novel associations from exome-wide rare variant analysis We found 11 gene-phenotype associations not previously reported in the context of rare variant association analyses (Table 1). Three of them were artery-related, including CYSTM1 with atherosclerosis of renal artery (Pcombined = 1.2 × 10−6), LDLR with occlusion and stenosis of precerebral arteries (Pcombined = 1.7 × 10−7), and POLN with stricture of artery (Pcombined = 3.9 × 10−7). Common variants of CYSTM1 and LDLR are associated with low-density lipoprotein cholesterol (LDL)33,34, and elevated LDL levels increase the risk of arterial diseases35. Additionally, CCNE2 was identified in association with current long-term use of anticoagulants, antithrombotics, and aspirin (Pcombined = 1.0 × 10−6), serving as a medication-based proxy for underlying cardiac or vascular disease. For heart rate, novel associations were identified for NAA15 (Pcombined = 1.1 × 10−8) and PTEN (Pcombined = 1.3 × 10−8). The strongest signals arose from NAA15 pseudo transcript (β = 0.69, 95% CI [0.46,0.92], Ptranscript = 3.0 × 10−9) and PTEN non-canonical transcript ENST00000700021 (β = 0.85, 95% CI [0.57,1.14], Ptranscript = 4.0 × 10−9), respectively. A prior study suggested that variants in NAA15 contributed to a genetic predisposition to hypertrophic cardiomyopathy36, and experimental studies showed that loss of PTEN attenuated the development of heart failure37. Novel associations for TCF12 suggested potential pleiotropic effects. Rare variants in TCF12 were significantly associated with both body mass index (BMI, Pcombined = 1.2 × 10−6) and waist-to-hip ratio (WHR, Pcombined = 3.2 × 10−7), traits that were closely related to obesity and metabolic health. Previously, a common variant in TCF12 was reported to be significantly associated with BMI38, and common variants near TCF12 were also linked to BMI39 and WHR40. The LOVO analysis (Supplementary Fig. 4) showed that the pDM variants rs2061221326 (risk allele A) and rs754118933 (risk allele T) drove the BMI and WHR associations, respectively. Functionally, TCF12 is involved in the process of adipocyte differentiation41 and the WNT signaling pathway42, with additional links to type 2 diabetes susceptibility43,44. These findings indicate that rare variants in TCF12 may contribute to metabolic disease risk. The LOVO analysis indicated that most significant associations resulted from the aggregation of information across rare variants (Supplementary Fig. 4). Quantile-quantile plots demonstrated well-controlled type I error rates under the proposed framework (Supplementary Note 1, Supplementary Figs. 5–9). Analyses using rare synonymous variants as negative controls showed no inflation. Additionally, ancestry-stratified analyses revealed consistent directional effects and comparable effect sizes across European (EUR)-like and non-EUR-like ancestry groups (Supplementary Fig. 10). For significant gene–phenotype associations, carrier penetrance varied across binary traits, with PKD1 showing 32% penetrance for hypertension, while other associations showed penetrance below 20% (Supplementary Fig. 11). To replicate new findings, we evaluated burden test results for pLoF variants from the UK Biobank (UKB) presented in Genebass45. Of the eight gene-phenotype pairs available in Genebass, two were successfully replicated. The remaining six could not be replicated due to either the absence of results for pDM variants or the limited ancestral diversity of the UKB, which is predominantly of EUR ancestry (Supplementary Note 1, Supplementary Table 1). Transcript-specific contributions improved rare-variant association detection A key observation from our study was that rare variant tests limited to a single canonical transcript did not always yield the strongest statistical significance (Fig. 4, Supplementary Data 1, 2). Among the significant associations, the transcript-aware approach produced smaller P values than the pseudo-only and canonical-only analyses in 46.55% and 39.66% of cases (excluding ties), respectively, and was the most significant across all three approaches in 29.31% of cases. For about half of the significant associations (excluding ties), the strongest ones were non-canonical transcripts, which may reflect organ- or tissue-specific relevance. The proposed framework effectively captured transcript-specific signals, improving their statistical significance. Specifically, a non-canonical transcript showed a P value of 2.3 × 10−28 for the TTN-cardiomyopathy association, resulting in a combined P value of 1.4 × 10−27, compared with 2.9 × 10−20 from the canonical transcript. This substantial improvement in the statistical evidence demonstrated the value of non-canonical transcripts that may be more relevant to the outcome. Similarly, for TTN and atrial fibrillation and flutter, incorporating non-canonical transcripts reduced the P value by approximately 98.5% compared to canonical-only analysis. All significant associations involving TTN were driven by non-canonical transcripts, underscoring their critical role in disease mechanisms46. For PPARG and triglycerides, significance was only achieved when transcript-specific contributions were considered (Pcombined = 1.1 × 10−7 compared to Pcanonical = 2.9 × 10−6 and Ppseudo = 5.6 × 10−6). Similarly, the PROC-deep vein thrombosis (DVT) association did not reach significance using the canonical transcript alone (Pcanonical = 3.1 × 10−6), but became significant after aggregation (Pcombined = 7.1 × 10−7). Furthermore, for PPARG and HDL, transcript-aware aggregation improved the significance by 80.2% (Pcombined = 8.7 × 10−8 compared to Pcanonical = 4.4 × 10−7). These findings highlighted the importance of incorporating transcript-specific information in rare variant analyses to better capture disease-relevant signals. A major advantage of the proposed framework is its ability to integrate transcript-specific contributions without manual transcript selection. While manually selecting tissue-, organ-, or disease-relevant transcripts can be powerful, it is often time-consuming and subjective, with results varying based on the selection criteria47. By incorporating all known protein-coding transcripts within a statistical framework, our approach minimizes the risk of overlooking functional variants. This is particularly important for complex human genes like TTN, with 364 exons and capable of generating over a million splice variants48. Non-canonical transcripts of TTN likely reflect alternative splicing patterns, which contribute significantly to its role in cardiac diseases49. In our study, we observed distinct and stronger associations from the disease-relevant N2B transcript of TTN. Notably, despite well-controlled type I error across all tested associations, no instances were observed where the P value for the canonical transcript was significant while the combined P value was not, demonstrating the robustness of this framework against noise (as also shown in Supplementary Fig. 3). This robustness was consistent with our simulation results, where the transcript-aware method maintained well-controlled type I error at stringent significance thresholds. Simulations also suggested that the framework may serve as a good surrogate for the ideal but unknown single-transcript approach, with comparable power across different scenarios. In addition to these advantages, the diversity of the All of Us cohort significantly enhanced the generalizability and robustness of our findings, with nearly 50% of participants from non-EUR-like ancestries. To assess the robustness of our results, we performed ancestry-stratified analyses (EUR-like vs. non-EUR-like), which showed consistent effect estimates and overlapping 95% CIs (Supplementary Fig. 10), indicating that the observed associations were not driven by population structure. Furthermore, we examined the variants within the most significant transcript of PROC for DVT, which was non-canonical. Some of the variants were classified as (likely) pathogenic for thrombophilia due to protein C deficiency, while others were absent from the ClinVar database30 (Supplementary Table 2), suggesting potential undiscovered pathogenic variants requiring further investigation. These results emphasize the critical role of diversity combined with the proposed approach in uncovering rare variant associations. We recognize some limitations of this study. According to our simulation results, power depends on multiple factors, including the number of causal variants and how they are distributed across transcripts. Other factors, such as transcript length, allele frequency, and sample size, can also influence power. Therefore, more comprehensive simulations that reflect additional complexities of rare variant architecture may provide clearer guidance for future method refinement. For the proposed framework, although incorporating the pseudo transcript increases sensitivity by integrating variants from canonical and non-canonical transcripts, it may also include variants whose functional impact is uncertain or weaker than predicted. This could introduce noise and dilute the association signal in some settings, particularly when the pseudo transcript aggregates variants across transcripts that are not equally relevant to the phenotype. Predicted deleteriousness annotations are useful for interpretation but are not always accurate, and experimental validation remains essential to confirm the biological relevance of these variants50. For real data applications, we recognize that including phenotypes beyond cardiopulmonary traits would further demonstrate the broader utility of the transcript-aware framework. In addition, while some of our findings were supported by comparisons with existing literature and UKB summary statistics45, large-scale, ancestrally diverse cohorts and harmonized analytic approaches are still needed to enable more comprehensive, comparable, and generalizable genetic analyses in the future. In conclusion, we introduced a transcript-aware gene-based aggregation framework for rare variant association analyses, demonstrating improved sensitivity and precision in detecting associations. In simulations, the transcript-aware method maintained well-controlled type I error rates at the stringent significance threshold and provided higher or competitive power than the canonical-only and pseudo-only approaches. Applying this approach to the diverse All of Us cohort, we identified 11 novel associations not previously reported, including one potentially pleiotropic gene, TCF12. For complex genes such as TTN12,51, leveraging transcript-specific information enhanced association signals by identifying the functionally impactful N2B isoform, resulting in stronger and more biologically relevant associations. Interestingly, 23 out of 24 significant binary trait associations were linked to circulatory conditions rather than respiratory conditions, which may be more strongly influenced by environmental factors. Future work incorporating tissue-specific transcriptomic data52 and expanding the phenotypic scope will further strengthen the biological interpretation and generalizability of the transcript-aware framework. In summary, our study underscores the potential of transcript-aware rare variant analyses in diverse populations to improve association detection, incorporate transcript-level information, and capture a broader spectrum of rare variant contributions to complex traits and diseases. Methods This study used de-identified data from the NIH All of Us Research Program accessed through the All of Us Researcher Workbench. The All of Us Research Program has received approval from the All of Us Research Program Institutional Review Board, and all participants provided informed consent. Simulation We conducted simulations with n = 200,000 individuals for a hypothetical gene with 300 rare and ultra-rare variants under the general rare variant analysis settings (Supplementary Methods). Of these variants, 20% were rare (minor allele frequency [MAF] ≈≈ 0.005–0.05%) and 80% were ultra-rare (MAF ≈≈ 0.0005–0.005%). For simplicity, we assumed no linkage disequilibrium among variants, which is reasonable for rare and ultra-rare variants. Variants were mapped to four transcripts (T1–T4), where T1 was the canonical (longest) transcript. We considered four general causal configurations to evaluate the proposed framework, including one null setting for type I error and three settings for power. In configuration 1 (scenario 1), none of the transcripts contained causal variants, and this setting was used to evaluate the type I error rate. In configuration 2 (scenarios 2–5), the total number of causal variants was fixed to 30, with all causal variants increasing disease risk (all have the same effect direction). T3 was treated as the most outcome-relevant transcript because it was set to contain all causal variants. Power was assessed by varying the proportion of causal variants assigned to the canonical transcript while T3 always contained all causal variants. These proportions were 30%, 60%, 80%, and 100%, corresponding to 9, 18, 24, and 30 causal variants in the canonical transcript (Supplementary Fig. 12). In configuration 3 (scenarios 6-8), the canonical transcript was the most outcome-relevant transcript that always contained all causal variants, and T3 was treated as noise with no causal variants. In configuration 4 (scenarios 9-11), causal variants were randomly located across transcripts. Configurations 3 and 4 were evaluated under varying numbers of causal variants (10, 20, and 30; Supplementary Fig. 13). For each of the eleven scenarios, we repeated the simulations 20,000 times. The outcome was binary with a prevalence of 0.2%, roughly corresponding to a disease that occurs about 1 in 500 individuals, such as hypertrophic cardiomyopathy53(Supplementary Methods). We compared four approaches: (1) canonical-only, (2) pseudo-only, (3) transcript 3 (T3)-only, and (4) our proposed transcript-aware method, all using the burden test. Here, canonical-only represented a standard approach that used a single transcript per gene without pre-selection. The “pseudo” transcript was constructed by combining the most severe consequence of each variant from the canonical and non-canonical transcripts within a gene15. In configuration 2, T3-only represented an idealized case in which the most informative transcript was known in advance. For each method and scenario, we evaluated the type I error rate and power at significance levels α = 0.001, 0.01, and 0.05. As an additional set of simulations, we repeated scenario 1 (the null setting) with an outcome prevalence of 2%. Furthermore, to better understand the behavior of the combined P value, we plotted the final aggregated P value from the proposed framework using hypothetical P value vectors outside the simulation setting (details in Supplementary Methods). Study population and phenotypes The All of Us Research Program is a longitudinal cohort study with available data from over 400,000 participants (Controlled Tier C2022Q4R9) across the United States, of which 80% of the core participants are individuals from communities underrepresented in biomedical research3. Participants were included in our study based on the availability of demographic and genomic data. In this study, we utilized participants’ age at enrollment, biological sex, predicted ancestries, and the first 16 principal components (PCs) computed by the All of Us Research program, as well as variants in protein-coding (exonic) regions derived from the WGS data (Supplementary Methods and Supplementary Tables 3, 4). Binary phenotypes were selected from two general categories, (1) circulatory system and (2) respiratory system, reflecting the importance of cardiovascular and pulmonary diseases as leading contributors to global morbidity and mortality19–21. These diseases impose a significant public health burden, with genetic factors playing an important role in their development and progression. Understanding their underlying genetic mechanisms can help identify at-risk individuals and develop targeted therapeutic interventions. The case status of each binary phenotype, ascertained by Electronic Health Record (EHR) data, was determined using the corresponding ICD-9 and ICD-10 codes. Previous studies have shown that recall accuracy could be affected by the number of phecodes used to define cases, with the maximum average recall accuracy achieved when cases were defined as having two phecodes54,55. Thus, we adopted this definition in our subsequent analysis. To enhance statistical robustness, phenotypes with fewer than 200 cases were excluded. A hierarchical clustering analysis using the distance matrix (1 - cosine similarity) was performed to identify highly correlated or redundant phenotypes56. The number of clusters was determined by the highest average silhouette width (Supplementary Methods and Supplementary Fig. 14)57. After reviewing all phenotypes within each cluster, we selected one representative with better clinical interpretability per cluster (Supplementary Data 3, 4). For continuous traits, 11 commonly studied circulatory and respiratory traits or biomarkers were selected, including All of Us computed body mass index (BMI, kg/m2), waist-to-hip ratio (WHR), heart rate per minute, glucose concentration in serum or plasma (GLU, mg/dL), eosinophils per 100 leukocytes in blood (EOS, %), C-reactive protein in serum or plasma (CRP, mg/L), diastolic blood pressure (DBP, mmHg), systolic blood pressure (SBP, mmHg), triglyceride concentration in serum or plasma (mg/dL), high-density lipoprotein cholesterol in serum or plasma (HDL, mg/dL), and low-density lipoprotein cholesterol in serum or plasma (LDL, mg/dL). Invalid values (e.g., physiologically implausible or with missing units) were excluded during data curation (Supplementary Methods and Supplementary Table 5). For each individual, the most recent non-missing measurement was used. Quality controls and annotation In addition to quality control (QC) performed by the All of Us Genome Centers and Data and Research Center, we carried out additional QC measures at both the genotype and variant levels (Supplementary Methods)58. In brief, we included only genotypes that passed central QC procedures and had a Genotype Quality score above 20. Variants with a missing rate greater than 10% were excluded. For sample QC, flagged population outliers (Supplementary Methods), potential duplicates, samples with a call rate below 90%, and individuals lacking genetically determined sex information were eliminated. This refinement process resulted in a final analysis dataset of 242,881 participants. Variants in exon regions of protein-coding transcripts based on GENCODE v42 were annotated using Variant Effect Predictor (VEP) version 10959. Because our approach incorporates transcript-specific effects, we retained a consequence of a variant per transcript as well as the most severe consequence. LOFTEE60 was implemented to determine predicted high-confidence LoF (pLoF) variants, including stop-gained, splice site disrupting, and frameshift variants. Among these pLoF variants, we filtered out flagged variants provided by LOFTEE. Furthermore, missense variants were annotated using 34 in-silico prediction tools provided by dbNSFP4.461. The predicted deleterious missense (pDM) variants were defined as those in which 80% or more of the prediction tools agreed on their deleteriousness58. Statistical analyses The primary analysis focused on identifying associations between rare variants and cardiopulmonary phenotypes. Continuous traits were inverse-rank-normalized before conducting any statistical analyses. We aggregated rare (MAF < 0.1%) pLoF variants without any flags and pDM variants in a gene. The cumulative burden of these variants was tested for association with cardiopulmonary phenotypes. We fitted a logistic mixed-effects model for binary traits and a linear mixed-effects model for continuous traits. For both models, we adjusted the fixed effects for age at enrollment, biological sex, the first two PCs, and any additional associated PCs (Supplementary Methods). The random effects accounted for the empirical relationships among participants, using a sparse kinship matrix estimated by KING62. To address case-control imbalance, the saddle-point approximation (SPA) method was applied to genes with nominal associations63. All analyses were conducted utilizing an adapted version of GENESIS v2.18 in R15,64. To enable transcript-aware aggregation, we constructed transcript-specific masks from variant annotations to capture the functional consequence of each variant in individual transcripts. For each gene, the burden of rare variants was calculated for each protein-coding transcript. Additionally, the pseudo transcript defined in the Simulation section was constructed for each gene15. As a result, the total number of analyzed transcripts per gene included the protein-coding transcript(s) plus one pseudo transcript, assuming all protein-coding transcripts reached the minimal carrier count. Association tests were not conducted when there were fewer than 20 carriers. Because burden tests for rare variants require the score statistics and covariances of individual variants, and the pseudo transcript encompasses all variants from all protein-coding transcripts, we first computed these values for variants presented in the pseudo transcript. As all other transcripts were nested within the pseudo transcript, association tests for them were derived using precomputed individual variant results and their covariance, ensuring greater computational efficiency. Subsequently, our proposed testing framework combined the burden P values from canonical, non-canonical, and pseudo transcripts using the Cauchy combining approach65. In cases where the same list of variants was included in multiple transcripts, a single representative P value was used and combined with other P values. This process yielded one P value per gene, representing the result of a gene-based test (Fig. 1 and Supplementary Methods). The Benjamini–Hochberg (BH) procedure was implemented to calculate q-values for all genes across all phenotypes to assess the statistical significance while controlling the false discovery rate (FDR) for multiple testing. A gene was considered significantly associated with the phenotype if its q-value was less than 0.05. The systemic inflation in type I error rate was evaluated by examining genomic control lambdas at the 50th and 90th percentiles. For significant gene-phenotype associations, ancestry-specific ORs were calculated for individuals of EUR-like and non-EUR-like ancestry using the most significant transcript. These ORs were derived using Firth’s logistic regression with the same model setting as the main analysis66. Additionally, a leave-one-variant-out analysis was conducted to assess the impact of individual rare variants on gene-level significance. To further validate the type I error rates of the proposed approach, the same analysis process was repeated using rare synonymous variants for phenotypes with at least one significant gene-phenotype association. Details for all statistical analyses can be found in the Supplementary Methods section. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Source data Source Data Source Data