Biomedical researchers analyzing complex psychiatric disorders increasingly depend on large-scale genome-wide association studies that stratify summary statistics across clinical subgroups, ancestries, and biological sexes. To resolve the computational complexity of harmonizing and evaluating these disparate datasets, researcher Zhong-shan Cheng at the Center for Applied Bioinformatics within St. Jude Children’s Research Hospital has developed MultiGWAS-Explorer. Detailed in Frontiers in Genetics, the software framework establishes an open-source, locally executable pipeline engineered to standardize stratified summary statistics, execute formal between-group effect comparisons, prioritize genomic loci using linkage disequilibrium reference panels, and generate both genome-wide and locus-specific visualizations. While between-stratum effect difference testing represents established methodology, MultiGWAS-Explorer addresses practical bottlenecks by uniting fragmented tools into a single, version-controlled environment executable via command-line scripts or an optional artificial intelligence interface.
The polygenic architecture of psychiatric conditions presents substantial hurdles because disease risk is distributed across hundreds of loci with modest individual effect sizes. Consortia, notably the Psychiatric Genomics Consortium, have generated expansive genome-wide association summary datasets across multiple diagnostic categories and refined demographic strata. These resources enable investigators to evaluate whether genetic signals exhibit uniform properties across related phenotypes or display marked heterogeneity across sex and ancestry backgrounds. Schizophrenia provides a compelling test case for this comparative paradigm, exhibiting epidemiological differences between males and females regarding age of onset, presenting symptom profiles, therapeutic responsiveness, and adverse-effect burdens. Furthermore, multi-ancestry analyses demonstrate that incorporating diverse ancestral backgrounds enhances the resolution of brain-related genetic signals, requiring software that preserves local chromosomal context across heterogeneous population groups.
Historically, investigators conducting comparative genome-wide association evaluations have relied on disparate software programs designed for isolated workflow segments. Tools such as GWAMA perform multi-study meta-analyses while calculating heterogeneity metrics, whereas EasyStrata evaluates stratified summary outputs through between-stratum effect difference tests and basic plotting. For local visualization and functional interpretation, geneticists frequently turn to separate utilities, including LocusZoom, FUMA, LocusCompare2, and the UCSC Genome Browser. Despite their utility, assembling them requires researchers to write bespoke scripts to reconcile differing allele conventions, execute repeated cross-stratum calculations, implement linkage disequilibrium clumping, and format publication-grade plots. MultiGWAS-Explorer eliminates this procedural fragmentation by offering an integrated, local pipeline that standardizes variant formatting, tracks quality filters, and executes standardized comparisons without requiring external web servers or introducing novel statistical algorithms.
To demonstrate the computational capabilities of the workflow, the author analyzed sex- and ancestry-stratified schizophrenia summary statistics released by the Psychiatric Genomics Consortium. The source datasets comprised 23,031 female cases and 45,626 female controls alongside 39,967 male cases and 41,159 male controls in the pooled sample. Within the European ancestry stratum, records detailed 17,760 female cases, 36,903 female controls, 33,097 male cases, and 35,190 male controls, while the Asian stratum incorporated 5,321 female cases, 8,823 female controls, 6,973 male cases, and 6,210 male controls. Because participant counts vary across genetic variants, the software tracked variant-specific sample sizes independently. In total, the demonstration coordinated six input summary datasets, organizing them into three primary female-versus-male comparisons across pooled, European, and Asian cohorts to evaluate cross-stratum effect differences systematically.
During data harmonization, MultiGWAS-Explorer standardizes variant records by aligning chromosomes, positions, alleles, effect sizes, standard errors, association p-values, imputation scores, and sample sizes across paired datasets. In the schizophrenia demonstration, the scan evaluated 22,638,007 pair-specific variant keys, confirming that both strata were present for 22,292,345 keys. Quality control protocols filtered out 3,386,012 strand-ambiguous adenine-thymine and cytosine-guanine variants to eliminate potential strand misalignment. The pipeline also excluded variant pairs where the aligned effect-allele frequency differed by more than 0.20 between strata, leaving 18,906,333 paired variants eligible for subsequent statistical analysis. Evaluation of genomic inflation demonstrated well-calibrated summary distributions, with differential genomic control lambda values reaching 1.020 for the pooled comparison, 1.021 for the European comparison, and 0.997 for the Asian comparison, against stratum-specific values ranging from 1.065 to 1.366.
Differential locus analysis calculates effect differences between matched strata by subtracting the second effect size from the first. The standard error accounts for stratum variances alongside an inter-stratum covariance term, set to zero in the primary demonstration because male and female cohorts were treated as non-overlapping. The resulting test statistic produces a two-sided p-value based on the standard normal distribution, which the workflow uses for variant ranking. The pipeline applies a minor allele frequency filter requiring values greater than 0.01 by default, while tracking stratum-specific imputation information metrics. Prioritized candidate loci undergo collective linkage disequilibrium clumping using PLINK2, drawing on phased reference genotypes from the 1000 Genomes Project Phase 3 across European and East Asian superpopulations using an r-squared threshold of 0.10.
To confirm that MultiGWAS-Explorer produces results identical to established software, the author conducted an empirical benchmark using 100,000 paired variants. The workflow was compared directly against GWAMA version 2.2.2 and EasyStrata version 8.6 under identical analytical assumptions. Both external reference programs successfully processed all 100,000 variants without failure. Comparative assessment revealed absolute agreement across all evaluated significance thresholds, showing zero discrepancies between MultiGWAS-Explorer and the two reference tools at nominal significance of p less than 0.05, suggestive significance of p less than 1.00 times 10 to the negative fifth power, and genome-wide significance of p less than 5.00 times 10 to the negative eighth power. This benchmark confirmed that automated harmonization and calculation modules replicate standard calculations without introducing numerical artifacts, rounding drift, or algorithmic bias.
Beyond command-line execution, the pipeline features an optional natural-language interface built around the Model Context Protocol, enabling local artificial intelligence assistants to execute computational tasks from plain-text instructions. The author evaluated this interface using a benchmark bank of 100 requests, selecting 20 representative queries across five categories: environment verification, configuration generation, minor allele frequency safeguards, missing-specification handling, and contradictory request resolution. Evaluated in headless Codex sessions, the interface initially achieved successful execution on 17 of 20 requests. After refining specification routing logic, improving error handling, and clarifying three ambiguous prompts, the system achieved success on 19 of 20 requests. Assessable configuration files matched command-line references exactly, as verified by canonicalized SHA-256 hashes, with a median execution time of 32.3 seconds and an interquartile range of 26.1 to 48.4 seconds.
In the demonstration using Psychiatric Genomics Consortium schizophrenia datasets, no variant achieved genome-wide significance for between-sex differences at the conventional threshold of p less than 5.00 times 10 to the negative eighth power. The strongest exploratory sex-differential finding appeared at variant rs185665940, which yielded an effect difference of negative 0.520, a standard error of 0.103, and a p-value of 4.321 times 10 to the negative seventh power in pooled comparisons. However, rs185665940 exhibited imputation information scores of 0.788 in females and 0.782 in males. Because both values fall below the conventional 0.80 threshold, the variant represents an exploratory, quality-control-sensitive signal requiring replication. A second variant, rs10166057, exhibited a nominal differential association with an effect difference of 0.092 and a p-value of 0.004, but statistical power was estimated at only 0.46 percent.
The workflow also evaluated candidate variants through linkage disequilibrium clumping using reference panels from the 1000 Genomes Project. Applying an exploratory threshold of p less than 1.00 times 10 to the negative sixth power identified three candidates passing minor allele frequency filters, retaining rs185665940 and rs753634 as lead variants while pruning rs7755143 due to high correlation. Direct calculation between rs185665940 and rs10166057 yielded an r-squared value of 0.001 in European genotypes, while the value could not be estimated in East Asian samples because rs185665940 was monomorphic. The author emphasized that this minimal correlation does not demonstrate that both genetic signals persist after mutual conditioning. Determining true conditional independence requires formal conditional and joint analysis or individual-level regression modeling, which cannot be accomplished using summary statistics alone.
The genomic region encompassing rs185665940 includes candidate genes such as CYP26B1, EXOC6B, and DYSF, providing biological context for future investigation. CYP26B1 encodes an enzyme responsible for metabolizing retinoic acid during human prefrontal development, while EXOC6B encodes an exocyst complex component associated with neurodevelopmental phenotypes. However, physical proximity between a prioritized variant and a nearby gene does not establish causality. Furthermore, tissue-level expression data from the Genotype-Tissue Expression project were unadjusted and based on nominal p-values, serving solely as descriptive genomic context rather than proof of sex-differential expression quantitative trait loci or causal mechanisms. A complementary common-association analysis evaluated 11,701 candidate rows, resulting in 206 retained lead variants, 8,947 pruned variants, and 2,548 distance-based fallback assignments, demonstrating pipeline behavior rather than newly discovered disease loci.
The publication delineates several important methodological limitations that contextualize the findings. Because public summary statistics do not provide individual-level genotypes, the pipeline cannot perform formal genotype-by-sex interaction testing or stratum-specific conditional regressions, nor can it account for individual-level demographic covariates or exclusion criteria. Furthermore, the artificial intelligence evaluation was restricted to a single model architecture tested across five task families, with primary prompts executed once, meaning findings describe specific task performance rather than universal artificial intelligence reliability. Ultimately, the author positions MultiGWAS-Explorer as an exploratory, locally reproducible bioinformatics framework rather than an autonomous discovery system. By systematically connecting allele harmonization, linkage-disequilibrium-aware locus prioritization, and multi-scale visualization, the platform provides geneticists with a transparent tool for exploring stratified genomic data and generating robust hypotheses for downstream experimental validation.
Subject of Research: Bioinformatics workflow for comparative analysis and visualization of stratified GWAS summary statistics
Article Title: MultiGWAS-Explorer: an AI-assisted workflow for comparative analysis and visualization of stratified GWAS summary statistics
Article References: Cheng, Z.-S. (2026). MultiGWAS-Explorer: an AI-assisted workflow for comparative analysis and visualization of stratified GWAS summary statistics. Frontiers in Genetics, 17, Article 1922359. https://doi.org/10.3389/fgene.2026.1922359
Image Credits: AI Generated
DOI: 10.3389/fgene.2026.1922359
Keywords: MultiGWAS-Explorer, Genome-Wide Association Studies, Stratified Analysis, Schizophrenia Genetics, Psychiatric Genomics Consortium, Summary Statistics Harmonization, Linkage Disequilibrium Clumping, Sex-Stratified Comparison, Ancestry Stratification, Bioinformatics Pipeline, PLINK2, GWAMA Benchmark, EasyStrata Benchmark, Codex Evaluation, Genomic Data Visualization, CYP26B1 Locus, Genotype-by-Sex Interaction
News Source: Juliet Wilcox. (November 19, 2026). MultiGWAS-Explorer Standardizes Comparative Analysis Across Stratified Genome Wide Association Datasets. Scienmag.



