TL;DR
This paper introduces scPrediXcan, a novel framework that leverages deep learning and single-cell RNA sequencing (scRNA-seq) data to perform transcriptome-wide association studies (TWAS) at the cell-type level. By accurately predicting cell-type-specific gene expression, scPrediXcan significantly outperforms traditional TWAS methods in identifying candidate causal genes for complex diseases like type 2 diabetes and systemic lupus erythematosus, offering deeper insights into disease mechanisms.
What Problem Does This Paper Address?
Traditional transcriptome-wide association studies (TWAS) aim to identify genes associated with diseases by linking genetic variants to gene expression. However, these methods often rely on tissue-level gene expression data, which can mask crucial cell-type-specific regulatory mechanisms that drive disease. The sparsity and limited scale of available cell-type-specific expression data have historically hindered the application of TWAS at this finer resolution. This paper addresses this gap by developing a method that can accurately predict cell-type-specific gene expression, even with limited scRNA-seq data, and integrate it into a TWAS framework.
Methods
The core of scPrediXcan lies in its ctPred model, a deep learning approach that predicts cell-type-specific gene expression. This model leverages transfer learning from Enformer, a state-of-the-art deep learning model that predicts epigenetic features from DNA sequences. By using Enformer's output as input, ctPred can learn complex gene regulatory patterns that linear models miss.
To make this computationally intensive model practical for TWAS, it's linearized into l-ctPred, a SNP-based elastic net model. This allows for TWAS using readily available GWAS summary statistics, bypassing the need for individual-level genotype data. The overall scPrediXcan framework involves three steps: 1) training ctPred for cell-type-specific gene expression prediction, 2) linearizing ctPred into l-ctPred, and 3) performing association tests using S-PrediXcan with GWAS summary statistics and l-ctPred.
Datasets
The authors utilized several publicly available scRNA-seq datasets for training and evaluating their models:
- OneK1K dataset: A large dataset comprising 29 cell types from 982 individuals.
- T2D islet dataset: Contains 11 cell types from 29 individuals, specifically focused on pancreatic islets.
- Tabula Sapiens dataset: A comprehensive atlas of single-cell transcriptomics from various organs. A subset of 6 cell types from 15 individuals was used for computational efficiency.
For GWAS summary statistics, they used:
- DIAGRAM consortium data: For type 2 diabetes (T2D).
- GWAS Catalog: For systemic lupus erythematosus (SLE).
Key Insights and Ideas
- Deep learning for gene expression prediction: The paper demonstrates that deep learning models, particularly when leveraging pre-trained epigenetic models like Enformer, can achieve high accuracy in predicting cell-type-specific gene expression from DNA sequences.
- Bridging the gap between scRNA-seq and TWAS: scPrediXcan effectively overcomes the limitations of sparse scRNA-seq data by using deep learning for robust expression prediction and then linearizing the model for efficient TWAS with GWAS summary statistics.
- Cell-type specificity is crucial: The results highlight that focusing TWAS at the cell-type level can uncover candidate causal genes and pathways that are missed by tissue-level analyses, providing a more nuanced understanding of disease etiology.
- Parameter efficiency matters: The ctPred model, despite its deep learning architecture, is significantly more parameter-efficient than other contemporary single-cell expression prediction models like emb2cell.
Key Results
- High prediction accuracy: ctPred achieved Pearson correlations ranging from 0.753 to 0.892 across various cell types and datasets, demonstrating its ability to accurately predict gene expression.
- Outperforms canonical TWAS: In both T2D and SLE analyses, scPrediXcan identified a significantly larger number of candidate causal genes and explained more GWAS loci compared to traditional TWAS-pseudobulk and TWAS-bulk methods.
- For T2D, scPrediXcan identified 222 candidate causal genes across 108 LD blocks, compared to 12 for TWAS-pseudobulk and 111 for TWAS-bulk.
- For SLE, scPrediXcan identified 129 candidate causal genes from 24 LD blocks, significantly outperforming TWAS-pseudobulk (11 genes, 8 LD blocks) and TWAS-bulk (54 genes, 14 LD blocks).
- Enhanced sensitivity: scPrediXcan showed higher sensitivity in identifying T2D-associated genes, particularly when compared to other gene prioritization methods.
- Cell-type-enriched associations: The framework revealed a substantial number of genes with cell-type-enriched associations, underscoring the importance of cell-type-specific analysis.
Limitations and Future Work
- Direction of correlation: The current analysis primarily focuses on the p-values of correlations and associations, not the direction of correlation. This limits the ability to discern whether disease risk is associated with up- or down-regulation of a nominated gene, which is crucial for drug target selection. The authors are working on improving this aspect with enhanced pre-trained models.
- Cis-regulatory focus: Like other TWAS frameworks, scPrediXcan primarily focuses on cis-regulatory mechanisms and does not account for trans-effects or other regulatory mechanisms.
- Linkage Disequilibrium (LD) errors: The method is susceptible to LD-induced errors, a common challenge in genetic association studies.
- Future Work: The authors plan to refine the directionality of gene expression changes and explore trans-effects. They also make scPrediXcan and its predictors publicly available to facilitate broader adoption and research.
Summary updated