henhaobio
Sign in

medRxivGenetic & Genomic MedicinePreprint

S-MiXcan: Inferring Cell-Type-Level Transcriptome-Wide Associations from Bulk Transcriptomics Using GWAS Summary Statistics

SINAN ZHU, Qiao FaniD, Xiaoyu SongiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Cell-type-specific regulation of gene expression plays a central role in complex disease etiology, yet most transcriptome-wide association studies (TWAS) rely on bulk tissue models. Recently, a couple methods leverage single-cell transcriptomics to perform TWAS at the cell-type resolution, but they are limited by scarce matched genotype- single-cell cohorts and restricted to peripheral blood, with minimal coverage of less accessible, disease-relevant tissues. In this study, we developed S-MiXcan, a summary-statistics-based TWAS framework that enables cell-type-aware association analysis using bulk transcriptomic data across K [≥] 2 cell types without requiring individual-level data. As a major advancement over our prior tool MiXcan, S-MiXcan jointly models genetically regulated expression (GReX) across K cell types, accounts for cross-cell-type correlations, identifies disease-associated genes, and provides probabilistic interpretations for distinguishing cell-type-specific from shared associations. In real data analyses, compared with using individual-level genotype-based implementation, S-MiXcan achieved highly concordant results (Pearsons r {approx} 1) in cell-type-aware TWAS. Applied to large-scale multi-cohort Genome-Wide Association Study (GWAS) meta-analyses from the Breast Cancer Association Consortium, S-MiXcan maintained well-controlled type I error (genomic inflation{lambda} = 1.057), identified key breast cancer risk associated genes that function in a cell-type specific manner, and revealed relevant cell types through probabilistic inference. These results demonstrate that S-MiXcan, publicly available at https://github.com/songxiaoyu/SMiXcan, provides a scalable and interpretable frame-work for cell-type-aware TWAS.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (24 Mar 2026) · CC BY 4.0

Cell-type-aware TWAS from bulk-trained expression models and GWAS summary statistics reproduces individual-level MiXcan results (r > 0.99) and finds 32 genome-wide significant breast cancer genes.

Standard transcriptome-wide association studies (TWAS; PrediXcan, FUSION) treat tissues as homogeneous, while single-cell TWAS needs large matched genotype and single-cell cohorts, mostly from blood; prior MiXcan handled only two cell types and required individual-level genotypes. S-MiXcan extends to K ≥ 2 cell types with GWAS summary statistics, adjusts for cross-cell-type correlation, and assigns probabilities to cell-type-specific versus shared associations.

Key results

  1. ShownSummary-statistics inference nearly reproduces individual-level MiXcan association results at the tissue level and within both the epithelial and stromal compartments.

    DRIVE cohort (OncoArray-30K; 58,648 European-ancestry women, 31,716 cases and 26,932 controls), with GWAS statistics computed from its genotypes and in-cohort LD: Pearson correlation of -log10(p) between S-MiXcan and MiXcan > 0.99 (abstract reports r ≈ 1). · Comparison with individual-level genotypes; Figure legends; Abstract

  2. ShownWhen applied to the Breast Cancer Association Consortium (BCAC) GWAS meta-analysis, the two-compartment (epithelial/stromal) model stays well calibrated and detects transcriptome-wide significant breast cancer risk genes.

    228,951 European-ancestry participants (122,977 cases, 105,974 controls); 6,405 tested genes; BACON genomic inflation 1.058 (abstract states λ = 1.057). Family-wise error rate (FWER) control at 5% (p < 7.8 × 10^−6) yielded 32 hits; a 10% false discovery rate (FDR) cutoff (p < 8.3 × 10^−4) flagged 76 suggestive genes. · Cell-type-aware TWAS for breast risk using GWAS…; Figure legends; Abstract

  3. SuggestedPattern probabilities suggest most suggestive genes act differently in the two compartments. A few are confidently assigned to stroma, including FES (risk-increasing) and CTSW and EP300 (risk-decreasing).

    Of 76 suggestive genes, 71 had non-identical effects across compartments: SRGAP1 mainly epithelial; seven stromal (FES, CTSW, EP300, TMEM69, CENPS, FKBP2, IQCH-AS1); 63 active in both with differing effects. FES, CTSW and EP300 had >95% probability of acting only in stroma. · Cell-type-aware TWAS for breast risk using GWAS…; Inferring cell-type–specific association…; Figure legends

  4. ShownSplitting the tissue into three cell types keeps the summary-statistics results concordant and calibrated but greatly reduces the number of discoveries, which shows the power cost of finer decomposition with small training sets.

    Epithelial, adipocyte/endothelial and fibroblast models from the same 125 samples: 6,403 predictable genes; -log10(p) r > 0.99 vs individual-level DRIVE analyses; inflation 1.064 in BCAC; only 17 FWER-significant hits and 33 suggestive ones (FDR threshold p < 4 × 10^−4), compared with 32 and 76 in the two-cell-type analysis. · Cell-type-aware TWAS performance under K > 2…

  5. ShownCell-type genetically regulated expression (GReX) models could be trained from bulk GTEx breast tissue for nearly all candidate genes, and most showed compartment-specific genetic regulation.

    125 GTEx v8 mammary samples from European-ancestry women, BayesDeBulk proportions and 10-fold cross-validation: models for 6,405 of 6,443 genes; 4,734 had distinct epithelial vs stromal weights and 1,671 had shared regulation. · Model training in GTEx mammary tissue samples

Methods, data and reproducibility

Bulk expression is deconvolved into K cell types (BayesDeBulk); a contrast-reparameterized elastic net learns cell-type genetically regulated expression (GReX) weights. Summary-statistic Z-scores with reference LD become ridge-stabilized joint Z-scores, combined by the aggregated Cauchy association test (ACAT) and given Primo-style pattern probabilities.

  • GTEx v8 mammary tissue bulk RNA-seq with matched genotypes from European-ancestry women, used to train two- and three-cell-type GReX models · human · n = 125 samples; 6,443 genes · dbGaP phs000424.v8.p2
  • DRIVE breast cancer case-control cohort genotyped on OncoArray-30K, used to compare S-MiXcan against individual-level MiXcan · human · n = 58,648 women (31,716 cases, 26,932 controls) · dbGaP phs001265.v1.p1
  • BCAC breast cancer risk GWAS meta-analysis summary statistics in European-ancestry women · human · n = 228,951 participants (122,977 cases, 105,974 controls)
  • 1000 Genomes Project Phase 3 European-ancestry reference panel for SNP matching and LD matrices · human (1000 Genomes reference panel)
Methods & tools
S-MiXcan, MiXcan, BayesDeBulk, Elastic-net regression, Helmert contrast coding, S-PrediXcan (summary-statistic Z-score derivation), ACAT (aggregated Cauchy association test), Primo (association-pattern probabilities), BACON (genomic inflation estimation), PEER factors (expression covariates), OncoArray-30K genotyping array
Reproducibility gaps
GTEx and DRIVE genotypes are obtained via dbGaP accessions; derivations, marker genes and gene lists are in Supplementary Materials not provided here; release of trained GReX weights is not stated; the printed 1000 Genomes link appears truncated.
Limitations (4)
  • Power falls as the number of cell types grows with limited training samples, because of more parameters, harder decomposition and stronger GReX collinearity; the three-cell-type run found 17 FWER-significant and 33 suggestive genes versus 32 and 76 with two cell types. · stated by the authors
  • Correlation adjustment cannot fully remove leakage of non-linear signals from other cell types, so cell-type-level p-values may stay inflated; the authors advise basing gene-level conclusions on the tissue-level ACAT p-value. · stated by the authors
  • Evidence is confined to one tissue and to European ancestry (125 GTEx training samples, 6,443 preselected genes, European GWAS and LD panel); compartment assignments such as stromal FES, CTSW and EP300 lack experimental or single-cell confirmation. · noted in this summary
  • The main text reports no null simulations and no head-to-head comparison with tissue-level S-PrediXcan/FUSION or single-cell TWAS. Calibration rests only on genomic inflation, which the abstract gives as 1.057 and the results give as 1.058. · noted in this summary

Read this if you run TWAS with GWAS summary statistics and want cell-type-resolved gene associations in tissues, such as breast, that lack matched genotype and single-cell cohorts.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSong '22Guo '26Sistig '25Zhou '24Liu '23Cao '26
88% semantic matchBioinformatics
Xiaoyu Song, Jiayi Ji, Joseph H. Rothstein+10 more·bioRxiv·Mar 15, 2022· 1 citations