ctPred

Transcriptome-wide association studies link GWAS signals to genes by predicting expression from genotype and testing that prediction against a trait. Their weak point is the reference panel: expression QTL data is overwhelmingly bulk tissue, and where cell-type-resolved expression exists at all, cohorts are small enough that linear SNP-based models converge for only a few percent of genes. TWAS therefore names genes but rarely the cell type in which they act. ctPred removes that bottleneck by predicting cell-type-specific expression from DNA sequence instead of from an eQTL cohort.

ctPred is a four-layer multilayer perceptron that maps sequence-derived epigenomic features onto pseudobulk gene expression for a given cell type. It is the reusable model inside scPrediXcan, the surrounding TWAS framework built by the Im lab at the University of Chicago with collaborators at Harvard Medical School, the Garvan Institute of Medical Research and UNSW Sydney, Argonne National Laboratory, and the National University of Singapore. The framework runs in three stages: train ctPred on single-cell pseudobulk expression, linearize it into a SNP-based elastic-net model called l-ctPred, then run association testing with S-PrediXcan against GWAS summary statistics. A user with a new GWAS skips the first two stages and applies released l-ctPred weights directly.

The design leans on transfer rather than scale. Instead of learning regulatory grammar from raw DNA, ctPred consumes the epigenomic track predictions of Enformer and learns only the much smaller mapping from those tracks to expression in one cell type, which is what makes it trainable on cohorts of a few dozen donors. It was published in Cell Genomics in May 2025.

#Key Features

• No matched genotypes required: Because the input features come from sequence, adding a cell type needs only expression measurements, not a genotyped eQTL cohort.
• Whole protein-coding transcriptome coverage: ctPred predicts roughly 19,000 genes per cell type, where a pseudobulk elastic net converges for about 3.5% of them.
• Accuracy without parameter count: On CD4+ T cells it reaches a Pearson correlation of 0.787 against observed expression, above the 0.666 reported for a comparison model with more than 60 million parameters.
• Linearization for summary-statistic TWAS: l-ctPred distils the network into SNP weights, with median 10-fold cross-validated Spearman correlation above 0.83 to ctPred across all 40 cell types, and ships as per-cell-type PredictDB databases with matching SNP covariance matrices.

#Technical Details

Enformer emits 5,313 genomic tracks across 896 sequence bins; ctPred averages the central four bins around the transcription start site into one 5,313-dimensional vector per gene. That vector passes through four hidden layers of 64 units with ReLU, dropout of 0.05, and weight decay of 5×10⁻⁴, ending in one output: the gene's rank-based expression percentile between 0 and 1. The model holds about 0.4 million parameters, with genes split by chromosome across train, validation, and test.

Training used OneK1K (982 individuals, 29 immune cell types), a type 2 diabetes pancreatic islet dataset (29 individuals, 11 cell types), and Tabula Sapiens (15 individuals, over 150 cell types across 14 organs), requiring at least 125 cells per cell type. Pearson correlations against observed pseudobulk expression run 0.836–0.892 across OneK1K cell types, 0.753–0.815 on islet types, and 0.823–0.885 on the Tabula Sapiens types examined. Against GTEx genetically regulated expression, ctPred recovers 15,339–16,277 correlated genes per cell type versus 277–1,646 for the pseudobulk elastic-net baseline. Linearization applies ctPred in silico to 462 European individuals from the 1000 Genomes panel, then fits an elastic net on SNP dosages to those predictions.

#Applications

The framework suits statistical geneticists who already hold GWAS summary statistics and want cell-type resolution without collecting new data. On type 2 diabetes, scPrediXcan nominated 222 candidate causal genes across 108 linkage disequilibrium blocks, against 12 genes in 11 blocks for a pseudobulk TWAS and 111 in 64 blocks for bulk pancreas, and recovered 24 silver-standard T2D genes at Bonferroni significance versus 4 and 13. On systemic lupus erythematosus it nominated 129 genes across 24 blocks, against 11 and 54. The cell-type attribution is the point: CASR surfaced in gamma cells and LPL in stellate cells, PYCARD in CD4+ T cells and ITGAM in monocytes.

#Impact

ctPred demonstrates that a sequence-based epigenomic predictor can stand in for an eQTL reference panel at cell-type resolution, reframing what limits a TWAS from cohort size to the quality of the upstream regulatory model. Released weights cover 40 cell types from two studies — 11 pancreatic islet types and 29 immune types — so another tissue means training a new ctPred model. The published limitations follow from that upstream dependency: some predictions correlate negatively with observed expression, so the direction of a disease effect cannot be read off reliably; Enformer captures promoter-proximal regulation better than the distal enhancers that carry cell-type specificity; prediction uncertainty is not propagated into the association test; and both the reference panel and the GWAS used are of European ancestry.