Skip to content

scPrediXcan

Summary

scPrediXcan brings TWAS down to the cell-type level. Canonical TWAS models require gene-expression reference panels of >100 individuals per tissue, which do not exist at single-cell-type resolution; scPrediXcan instead trains a deep learning model, ctPred, that predicts cell-type-specific expression by fine-tuning a pretrained sequence-to-epigenomics model (Enformer) on pseudo-bulk single-cell RNA-seq data, sidestepping the sparsity of raw scRNA-seq. Applied to type 2 diabetes (T2D) and systemic lupus erythematosus (SLE), scPrediXcan identified substantially more candidate causal genes and explained more GWAS loci than both bulk-tissue and pseudobulk TWAS baselines.[1]

Method

  • ctPred: rather than training a sequence-to-expression model from scratch on sparse single-cell data, ctPred fine-tunes Enformer (a pretrained sequence-to-epigenomics deep learning model) using pseudo-bulk scRNA-seq expression aggregated per cell type, capturing regulatory grammar that linear (elastic-net) expression models cannot.
  • scPrediXcan: applies the standard S-PrediXcan association test using ctPred-derived, cell-type-specific expression predictions in place of tissue-level linear predictors, producing a TWAS at cell-type resolution rather than tissue resolution.
  • A companion STAR Protocols paper details the full pipeline: training cell-type-specific expression models, predicting personalized expression, and testing GWAS summary-statistic associations, designed to run with modest computational resources.

Performance

  • Type 2 diabetes: scPrediXcan identified 222 candidate causal genes across 108 independent LD blocks (of 1,703 total pre-defined blocks), versus 12 genes/11 blocks for pseudobulk TWAS and 111 genes/64 blocks for bulk-tissue TWAS. Against UTMOST (a canonical bulk-tissue TWAS method) applied to pancreatic tissue, scPrediXcan (in β cells) found a similar gene count (222 vs. 221) but higher precision (0.109 vs. 0.082) and much higher recall (0.439 vs. 0.153) against a curated "silver standard" T2D gene set.
  • Systemic lupus erythematosus: scPrediXcan again identified more candidate causal genes and explained more GWAS loci than both bulk and pseudobulk baselines, including biologically plausible hits such as CFB (complement pathway, implicated in autoimmune clearance defects) and CXCR5 in plasmablasts (previously reported as differentially expressed in SLE and involved in disease progression).
  • Sensitivity in nominating candidate causal T2D genes also exceeded other gene-prioritization methods tested, including COLOC, polygenic priority score, and the effector index.

Availability

Source code: github.com/hakyimlab/scPrediXcan (archived on Zenodo). Prediction models: predictdb.org (Zenodo: 15318900, 14346661). A detailed STAR Protocols walkthrough of the full training-to-association pipeline is published alongside the primary paper.

See Also

Citations

[1] Zhou, Y., Adeluwa, T., Zhu, L., Salazar-Magaña, S., Sumner, S., Kim, H., Gona, S., Nyasimi, F., Kulkarni, R., Powell, J.E., Madduri, R., Liu, B., Chen, M., & Im, H.K. (2025). scPrediXcan integrates deep learning methods and single-cell data into a cell-type-specific transcriptome-wide association study framework. Cell Genomics, 5, 100875. [2] Zhou, Y., Sumner, S., Adeluwa, T., et al., Chen, M., & Im, H.K. (2025). Protocol to perform cell-type-specific transcriptome-wide association study using scPrediXcan framework. STAR Protocols. j.xpro.2025.104306