TorchGWAS : GPU-accelerated GWAS for thousands of quantitative phenotypes Xingzhong Zhao1, Ziqian Xie1, Islam, Sheikh Muhammad Saiful 1, Tian Xia1, Chen, Cheng1, Degui Zhi1
1. Department of Bioinformatics and Systems Medicine, McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, Texas, United States of America
Abstract Motivation: Modern bioinformatics workflows, particularly in imaging and representation learning, can generate thousands to tens of thousands of quantitative phenotypes from a single cohort. In such settings, running genomewide association analyses trait by trait rapidly becomes a computational bottleneck. While established GWAS tools are highly effective for individual traits, they are not optimized for phenotype-rich screening workflows in which the same genotype matrix is reused across a large phenotype panel. Results: We present TorchGWAS, a framework for high-throughput association testing of large phenotype panels through hardware acceleration. The current public release provides stable Python and command-line workflows for linear GWAS and multivariate phenotype screening, supports NumPy, PLINK, and BGEN genotype inputs, aligns phenotype and covariate tables by sample identifier, and performs covariate adjustment internally. In a benchmark with 8.9 million markers and 23,000 samples, fastGWA required approximately 100 second per phenotype on an AMD EPYC 7763 64-core CPU, whereas TorchGWAS completed 2,048 phenotypes in 10 minute and 20,480 phenotypes in 20 minutes on a single NVIDIA A100 GPU, corresponding to an approximately 300- to 1700-fold increase in phenotype throughput. TorchGWAS therefore makes large-scale GWAS screening practical in phenotype-rich settings where thousands of quantitative traits must be evaluated efficiently. Availability and implementation: TorchGWAS is implemented in Python and distributed as a documented source repository at https://github.com/ZhiGroup/TorchGWAS. The current release provides a command-line interface, packaged source code, tutorials, benchmark scripts, and example workflows. Keywords: GWAS, GPU computing, imaging genetics, deep phenotyping
1 Introduction Genome-wide association studies have become a standard approach for identifying genetic influences on complex traits(Visscher, Wray et al. 2017). With cohorts grow larger and phenotyping becomes increasingly high dimensional, the computational burden of GWAS is shaped not only by the number of variants and samples, but also by the number of phenotypes under study. This is especially apparent in resources such as UK Biobank and in modern machine-learning pipelines, where large numbers of candidate traits can be derived from the same individuals(Bycroft, Freeman et al. 2018). In these settings, the same genotype matrix may be scanned repeatedly across an expanding phenotype panel, creating a substantial computational burden for downstream analyses. Here, we present TorchGWAS, a high-throughput framework for association analysis of large quantitative phenotype panels. The software was motivated by brain MRI representation-learning workflows in which large numbers of candidate phenotypes are repeatedly screened for genetic association. TorchGWAS is intended to complement, rather than replace, established mixed-model tools by serving as an efficient first-pass screening engine in phenotype-rich studies. Existing GPU-accelerated association tools address related, but distinct, analytical settings from TorchGWAS. tensorQTL (Taylor-Weiner, Aguet et al. 2019) is the closest methodological comparator, but it was developed primarily for QTL mapping and expects genomically anchored phenotype input, making it less suitable for imagingderived traits and learned representations that are typically analyzed in a genome-wide manner. fastGWA(Jiang, Zheng et al. 2019) and SAIGE-GPU (Rodriguez, Kim et al. 2026) focus instead on efficient mixed-model analysis of individual traits or binary trait families, whereas permGWAS(John, Ankenbrand et al. 2022) and rMVP(Yin, Zhang et al. 2021) improve permutation-based inference and CPU-level parallelization, respectively. Together,
these tools address important computational bottlenecks in GWAS, but they do not specifically target the phenotyperich setting now common in imaging and representation-learning studies. TorchGWAS was developed for phenotype-rich studies involving large panels of quantitative traits measured in the same cohort. Rather than optimizing one phenotype at a time, it is designed to reuse phenotype preprocessing across the full panel, thereby improving efficiency as the number of traits increases. It also accepts PLINK, BGEN and NumPy genotype inputs directly, which facilitates large-scale application. TorchGWAS should therefore be viewed as complementary to existing mixed-model methods, serving as an efficient first-pass framework for dense genomewide screening across large quantitative phenotype panels. 1. Inputs
2. Preprocessing
3. Linear GWAS
GENOTYPE
• •
•
Chunked genotype scan
NumPy matrix PLINK bed/bim/fam BGEN + sample
• •
•
Marker-wise correlation / t-statistic
•
Trait-specific P values
•
Compressed results table
• PHENOTYPES
•
Infer genotype sample order Align phenotype/covariate tables by IID Drop zero-variance columns Compute covariate QR basis internally Residualize and standardize phenotypes Record QC and run metadata
Matrix or tabular file Quantitative traits LINEAR OUTPUTS COVARIATES Matrix or tabular file Aligned by IID
PREP OUTPUTS
results.tsv.gz run.json qc.json beta, se, t, -log10 P
phenotype_processed.npy covariate_q.npy, qc.json, prep.json
Fig.1 Overview of the TorchGWAS
2 Methods 2.1 Software design TorchGWAS is implemented in Python and organized as a packaged command-line workflow. The software is based on a simple observation: for quantitative phenotypes measured on the same subjects, the core workload of association testing can be reorganized as repeated matrix operations between batches of genotypes and a phenotype matrix. Rather than loading one phenotype at a time and rerunning an association pipeline independently, TorchGWAS loads a phenotype panel, residualizes and standardizes it once, and then reuses this matrix across the genome scan. 2.2 Association kernel For an input phenotype matrix 𝑌 ∈ ℝ!×# with P phenotypes over N samples TorchGWAS first mean-centers each phenotype and removes covariate effects using an orthonormal basis 𝑄 spanning the covariate space. In the current implementation, phenotype residualization is performed as 𝑌$%& = (𝐼−𝑄𝑄' )(𝑌−𝑌¯) (1) where 𝑌¯ denotes the column-wise phenotype means and 𝑄 is assumed to have orthonormal columns. The phenotype matrix is then standardized column-wise to unit variance, yielding 𝑌,. Genotypes are processed in batches of 𝑀 markers. For each batch, each marker vector is standardized across samples， and giving a genotype matrix 𝐺, ∈ ℝ(×! . TorchGWAS then computes the genotype-phenotype correlation matrix as 𝑅 = 𝐺, 𝑌,/𝑁 (2) where 𝑅 ∈ ℝ(×# . These correlation coefficients are converted to t-statistics using 𝑇 = 𝑅3
!)*
+), !
(3)
Thus, each genotype batch produces an 𝑀 × 𝑃 matrix of association statistics across all phenotypes simultaneously. This matrix formulation allows phenotype preprocessing and genotype traversal to be shared across large phenotype panels, substantially improving throughput in phenotype-rich settings. This formulation is equivalent to a correlation-based linear association test for standardized quantitative traits and standardized genotypes. The resulting statistics can therefore be converted directly to P values, used to rank associated loci, or incorporated into downstream phenotype-level prioritization. 2.4 Scope of the current version The present version of TorchGWAS is intended for high-throughput quantitative-trait association under a linear model formulation. Its principal strength is throughput and workflow stability when many phenotypes share the same cohort and genotype matrix. Two caveats are important. First, the current release focuses on linear and multivariate quantitative-trait screening and does not yet provide the broader modelling options available in mature mixed-model toolkits. Second, although PLINK and BGEN entry points are supported, large-scale end-to-end performance still depends on backend reader efficiency, storage bandwidth, and the execution environment. In phenotype-rich settings, however, the standardized workflow remains attractive because preprocessing and genome scans are amortized across many traits.
3 Results 3.1 Benchmark results We evaluated TorchGWAS using complementary lines of evidence. First, we compared its performance against archived fastGWAS results generated from a dataset comprising 8.9 million markers and approximately 23,000 samples. In those prototype benchmarks, fastGWA required approximately 100 s per phenotype on an AMD EPYC 7763 64-core processor, whereas TorchGWAS processed 2,048 phenotypes in about 940 s and 20,480 phenotypes in about 1108 s on a single NVIDIA A100 GPU Second, results from TorchGWAS were in near-perfect agreement with archived PLINK outputs for representative traits (Pearson correlation = 0.999). The slight departure from exact concordance likely reflects small implementation-level differences between software pipelines, including floatingpoint precision and preprocessing order, rather than substantive differences in association estimates.
30
20 105
Seconds
− log10(P) from PLINK
fastGWA torchGWAS
106
25
15
104
10
5 103 0 0
5
10 15 20 − log10(P) from torchGWAS
25
30
2048
20480 Number of phenotypes
Figure 2. Validation and benchmarking of TorchGWAS. Left: Comparison of SNP association significance estimated by TorchGWAS and PLINK, showing near-perfect agreement for representative archived traits. Right: Runtime comparison of TorchGWAS and fastGWA for 2,048 and 20,480 phenotypes. Runtimes are shown on a log scale.
3.2 Scaling behavior and practical implications Runtime increased much more slowly than linearly with the number of phenotypes, because phenotype preprocessing and genotype traversal were shared across the phenotype matrix. As a result, TorchGWAS is well suited to phenotype-rich settings, where the main scaling challenge is the number of traits rather than the number of variants alone. Overall, the benchmark supports TorchGWAS as a reproducible high-throughput GWAS workflow for large collections of quantitative phenotypes.
4 Discussion TorchGWAS addresses an important gap in GWAS workflows for ultra-high dimensional phenotypes. Many widely used tools were developed for association testing on a single phenotype, or at most a limited set of traits, which is increasingly restrictive in settings such as imaging-derived phenotypes, representation learning, iterative feature construction, and voxel-level summaries. In these contexts, repeatedly scanning the same genotype resource across large phenotype panels is computationally inefficient. TorchGWAS therefore offers a reusable, well-documented workflow for high-throughput screening of quantitative traits, with stable inputs and reproducible outputs. Importantly, the current implementation already includes relatedness-aware sample exclusion during preprocessing, although full mixed-model correction is not yet implemented within the association engine itself. The current release should therefore be regarded as a screening-oriented framework rather than a complete replacement for established GWAS software. At present, it supports linear association testing only and does not yet provide generalized linear models, mixed-model inference. Performance also depends on the genotype backend, storage throughput, and hardware configuration; accordingly, the reported speedups should be interpreted in light of the benchmark environment used in this study. In addition, although the concordance analyses support numerical correctness, broader validation across additional datasets and analytical settings will be important in future releases. Despite these limitations, TorchGWAS provides a practical and reproducible solution to an increasingly important computational bottleneck in high-dimensional phenotype analysis.
Availability and implementation TorchGWAS is implemented in Python and is distributed as a documented source repository at https://github.com/ZhiGroup/TorchGWAS. The repository contains packaged source code, installation instructions, tutorials, benchmark scripts, toy examples, and local real-data workflow templates.
Funding This work was supported by U01AG070112.
Conflict of interest None declared.
References 1.
Bycroft, C., C. Freeman, D. Petkova, G. Band, L. T. Elliott, K. Sharp, A. Motyer, D. Vukcevic, O. Delaneau, J. O'Connell, A. Cortes, S. Welsh, A. Young, M. Effingham, G. McVean, S. Leslie, N. Allen, P. Donnelly and J. Marchini (2018). "The UK Biobank resource with deep phenotyping and genomic data." Nature 562(7726): 203209. 2. Jiang, L., Z. Zheng, T. Qi, K. E. Kemper, N. R. Wray, P. M. Visscher and J. Yang (2019). "A resource-efficient tool for mixed model association analysis of large-scale data." Nat Genet 51(12): 1749-1755. 3. John, M., M. J. Ankenbrand, C. Artmann, J. A. Freudenthal, A. Korte and D. G. Grimm (2022). "Efficient permutation-based genome-wide association studies for normal and skewed phenotypic distributions." Bioinformatics 38(Suppl_2): ii5-ii12. 4. Rodriguez, A., Y. Kim, T. N. Nandi, K. Keat, R. Kumar, M. Conery, R. Bhukar, M. Liu, J. Hessington, K. Maheshwari, V. A. M. V. Program, E. Begoli, G. Tourassi, S. Muralidhar, P. Natarajan, B. F. Voight, K. Cho, J. M. Gaziano, S. M. Damrauer, K. P. Liao, W. Zhou, J. E. Huffman, A. Verma and R. K. Madduri (2026). "SAIGEGPU: accelerating genome- and phenome-wide association studies using GPUs." Bioinformatics 42(3). 5. Taylor-Weiner, A., F. Aguet, N. J. Haradhvala, S. Gosai, S. Anand, J. Kim, K. Ardlie, E. M. Van Allen and G. Getz (2019). "Scaling computational genomics to millions of individuals with GPUs." Genome Biol 20(1): 228. 6. Visscher, P. M., N. R. Wray, Q. Zhang, P. Sklar, M. I. McCarthy, M. A. Brown and J. Yang (2017). "10 Years of GWAS Discovery: Biology, Function, and Translation." Am J Hum Genet 101(1): 5-22. 7. Yin, L., H. Zhang, Z. Tang, J. Xu, D. Yin, Z. Zhang, X. Yuan, M. Zhu, S. Zhao, X. Li and X. Liu (2021). "rMVP: A Memory-efficient, Visualization-enhanced, and Parallel-accelerated Tool for Genome-wide Association Study." Genomics Proteomics Bioinformatics 19(4): 619-628.