A transcriptome-wide association study (TWAS) asks which genes link genetic variants to a disease. It learns, from people who have both genotypes and expression measurements, how nearby variants predict each gene's expression. It then tests whether that genetically predicted expression tracks disease risk in large genome-wide association study (GWAS) summary statistics. Single-cell RNA sequencing now makes it possible to train these predictors separately in each cell type, which removes the averaging that hides cell-type-specific effects in bulk tissue. Current single-cell TWAS methods still reduce each gene in each cell type to one number: an average expression level. Single-cell data carry a second signal that averaging throws away. Beyond changing how much of a gene a cell makes, a variant can change the fraction of cells in which the gene is switched on at all.
We propose a TWAS that tests two separate, signed genetic channels for every gene and cell type. The first is a level channel: the genetically predicted mean count. The second is an excess-silence channel: the genetically predicted fraction of cells showing zero counts beyond what the mean level alone would produce under standard count noise. Both channels use linear weights and can be combined with GWAS summary statistics and linkage disequilibrium (LD), the correlation between nearby variants. Each gene–trait hit therefore comes with a direction ("risk rises as the gene is silenced in more CD8 T cells") and with the type of regulation behind it. Neither is available from existing cell-type TWAS.
Core Claim
In droplet single-cell data, the observed fraction of zeros is mostly a deterministic function of the mean. Under a negative binomial model with mean $\mu$ and dispersion $\phi$, the expected zero fraction is $(1+\mu/\phi)^{-\phi}$. A naive "detection rate" phenotype is therefore largely the level phenotype again, confounded by each donor's sequencing depth. The informative quantity is the residual. For donor $i$, gene $g$ and cell type $c$, we define
$$ s_{igc} = \frac{1}{n_{ic}}\sum_{k} \mathbf{1}[y_{kg}=0] \;-\; \frac{1}{n_{ic}}\sum_{k}\left(1+\frac{\ell_k \hat\mu_{igc}}{\hat\phi_{gc}}\right)^{-\hat\phi_{gc}}. $$
Here $k$ runs over the $n_{ic}$ cells of type $c$ from donor $i$, $y_{kg}$ is the count of gene $g$ in cell $k$, $\ell_k$ is that cell's library-size factor, $\hat\mu_{igc}$ is the donor's depth-normalized mean, and $\hat\phi_{gc}$ is the gene's dispersion within the cell type. Positive $s_{igc}$ means more cells are silent than the donor's average level predicts. That can come from genotype-driven changes in how often transcription switches on, or from a genotype-driven shift toward a hidden substate of the annotated cell type that does not express the gene.
We claim three things. First, for a measurable set of genes, $s$ is heritable and partly independent of $\mu$ in cis. Second, testing both channels jointly finds gene–trait associations that level-only TWAS misses. Third, the sign of each channel's association is stable and interpretable. The second possible source of $s$, hidden substates, is exactly the cellular-heterogeneity masking problem operating inside an annotated cell type. We treat telling the two sources apart as part of the research, not as noise.
Where This Sits Relative to Prior Work
The closest work is scPrediXcan. It trains a deep sequence-based predictor of cell-type expression on single-cell data, linearizes it into elastic-net SNP weights, and runs summary-statistic TWAS. On SLE it reports 129 candidate causal genes across 24 LD blocks, against 11 genes across 8 blocks for pseudobulk TWAS and 54 across 14 for bulk. Its limitations state that it does not report direction of effect. Because its final weights are linear, a sign could in principle be read off. The deeper limitation we target is that its prediction target is a single continuous expression value. It models neither the discrete count structure nor a separate on/off component.
MTWAS splits cross-tissue from tissue-specific eQTLs and applies this to pseudobulked OneK1K. Again, there is one continuous channel per gene.
SCeQTL supplies the biological premise. Its zero-inflated negative binomial regression lets genotype move the probability of being off and the active mean separately, and it outperforms Matrix eQTL in simulation when both move. However, it tests single SNP–gene pairs, treats cells as independent observations, and never aggregates to gene–trait tests. We reuse its two-part idea, move the unit of analysis to donors, and connect it to GWAS.
Bayesian model selection reveals biological origins of zero inflation both motivates and constrains the design. It supports excess zeros having biological causes in some genes. It also implies that a two-part model is unnecessary for many genes. For that reason, our status channel is defined as a residual that is near zero when no excess exists, rather than as a mandatory zero-inflation parameter.
Count models such as scNBMF and Count Bridges show the value of treating counts as integers, but they have no association component. We borrow the negative binomial baseline only. Iterative bridge sampling across 982 donors, every gene and every cell type is not affordable in a TWAS loop.
On the association side, two papers constrain what counts as success. The ten-blood-cell-phenotype TWAS found that 93% of significant associations vanish when conditioning on sentinel GWAS variants. Gene score regression showed LD-driven co-regulation inflating TWAS statistics, with a gene-connectivity correlation of 0.76 in schizophrenia. Raw gene counts are therefore not evidence of a better method, and every comparison below includes conditional and calibration checks.
Research Program
The construction has three stages.
Per-donor phenotypes. For each OneK1K cell type with enough cells, we fit a gene-wise negative binomial with donor random intercepts. This yields each donor's level $\log\hat\mu_{igc}$ and excess-silence residual $s_{igc}$, the latter with a delta-method variance that shrinks as $n_{ic}$ grows. A precision weight is attached to each donor–cell-type phenotype, so donors with few cells of a rare type are down-weighted rather than dropped. This is a cheap, donor-level treatment of sparsity that avoids per-cell likelihoods in the association step.
Two weight vectors per gene and cell type. We train elastic-net cis weights $w^{L}_{gc}$ and $w^{S}_{gc}$ on these phenotypes, with genotype principal components and depth covariates regressed out. We report cross-validated prediction accuracy and heritability for each channel. We also report the genetic correlation between the channels, which tells us how often the status channel is genuinely distinct.
Joint summary-statistic test. For GWAS z-scores and a reference LD matrix, S-PrediXcan gives two signed statistics per gene–cell-type pair, $z^{L}$ and $z^{S}$. Their correlation $\rho$ follows from the two weight vectors and the LD matrix. We use the two-degree-of-freedom statistic $T = \mathbf{z}^\top R^{-1}\mathbf{z}$ with $R=\begin{pmatrix}1&\rho\\ \rho&1\end{pmatrix}$, which is $\chi^2_2$ under the null because both z-scores are linear in the same Gaussian GWAS statistics. We then report which channel carries the signal and its sign.
A diagnostic stage addresses the substate interpretation. If $s$ reflects a genotype-driven shift in a hidden subpopulation, the $s$ values of that subpopulation's marker genes should share the same cis signal. We test this by subclustering each cell type and asking whether a status-channel eQTL also predicts the donor's subcluster proportions, in the spirit of the cell-fraction GWAS in SPLITR. Genes whose status signal is explained by subcluster proportion are labeled "composition-mediated". The remainder are labeled "kinetic".
Baselines
The naive reference point is pseudobulk TWAS: sum counts per donor and cell type, log-normalize, fit elastic-net weights, and run S-PrediXcan. This is the TWAS-pseudobulk arm of scPrediXcan, which reports 11 genes across 8 LD blocks for SLE. We will re-implement it ourselves, with proper depth normalization and the same donors, genotype covariates and cross-validation folds as our method. The published figure may understate a well-tuned pseudobulk. A fair version is the threat our method must beat, not a strawman.
The strongest existing method is scPrediXcan's l-ctPred weights with S-PrediXcan. The best reported number is SLE: 129 candidate causal genes across 24 LD blocks, with ctPred Pearson correlations of 0.753 to 0.892 across cell types scPrediXcan. We will use released weights if available, or retrain on the same OneK1K split. We will run it against the same GWAS files and LD reference as our method. We also run MTWAS on OneK1K. There is no reported result on Lacuna for MTWAS on OneK1K gene–trait discovery counts; its reported figure is a 47.4% average gain in prediction $R^2$ over PrediXcan on GTEx MTWAS.
Beating these baselines does not mean producing a higher raw gene count. It means three things. The method must recover more LD blocks that survive conditioning on sentinel GWAS variants. It must remain calibrated under permuted-phenotype nulls. Its status-only hits must replicate in an independent GWAS of the same trait, or in UK Biobank phenotypes, at a rate comparable to its level hits.
Three ablations isolate the key ingredient. The first uses the level channel only, with the same count-aware phenotype. This measures what the status channel adds. The second replaces $s$ with the raw detection fraction. This tests the claim that the residual, not detection per se, carries independent signal. We expect raw detection to correlate strongly with level and add little. The third replaces our donor-level moment estimates with a per-cell zero-inflated negative binomial fit in the style of SCeQTL. This asks whether full count likelihoods buy anything over the cheap residual. All arms get the same compute budget and the same hyperparameter grid.
Why not just…?
Why not just use bulk TWAS on GTEx or whole blood? Bulk profiles have no cells, so they cannot define a within-cell-type silent fraction at all. The file documents that bulk averaging hides signals. SPLITR recovered 1,182 eGenes invisible in bulk brain data, and the blood-trait TWAS could not model CSF3 because its expression is too low in mature blood. Bulk whole-blood TWAS remains a reported comparator to show where our hits overlap it.
Why not just pseudobulk and add a detection-rate phenotype? This is the most serious alternative, and our design incorporates it rather than dismissing it. Our method is a pseudobulk method: phenotypes are computed per donor, not per cell. The disagreement is only over which second phenotype to compute. The second ablation runs exactly this alternative. If raw detection fraction performs as well as the residual $s$, the simpler version wins and the negative binomial correction is unnecessary. We predict it will not win, because raw detection is mostly the mean again plus a donor-depth confound, but the experiment decides.
Why not just run SCeQTL per SNP and aggregate with a gene-set tool like MAGMA or PASCAL? SCeQTL treats cells from the same donor as independent. With hundreds of cells per donor, this overstates the effective sample size, which is 982 donors, not a million cells. Its per-cell likelihood across all SNPs, genes and cell types is also expensive. Separately, SNP-to-gene aggregation tools suffer from the LD problem: GSR reports PASCAL at an area under the precision–recall curve of 0.260 versus 0.925 for gene score regression in simulation. The third ablation measures whether per-cell modeling adds power once donor clustering is respected.
Why not just read signs off scPrediXcan's linear weights? For the level channel this is legitimate, and we will do it as a comparator. It still gives one signed channel. It cannot tell a variant that silences a gene in a subset of cells from one that lowers its output uniformly. Those two mechanisms imply different interventions.
Proposed Experiments and Decisive Checks
Simulation. We use real OneK1K genotypes and simulate counts with known cis effects on level only, on excess silence only, or on both, including a scenario where silence arises from a genotype-driven hidden subcluster. We measure type I error of $T$ and of each single-channel statistic under nulls. We measure power by channel and the accuracy of the recovered sign. We also check whether the substate diagnostic separates composition-mediated from kinetic silence.
Real data. In OneK1K, the first decisive check is descriptive: how many gene–cell-type pairs have a cross-validated status-channel prediction $R^2 > 0.01$ with low genetic correlation to level. If this set is tiny, the proposal's premise fails, and we will report that as a negative result.
We then run both channels against SLE summary statistics from the GWAS Catalog and against UK Biobank immune-related phenotypes, comparing with the baselines above. We report sentinel-conditioned survival, following the 93% dropout finding in the blood-cell TWAS, and apply gene score regression to check that status-channel hits are not driven by LD with neighboring genes.
The T2D islet data (29 donors) are too small to train reliable weights and will not be used for training. For direction, we check sign agreement between level and status channels for the same gene. Opposite signs in a gene that is both silenced and raised are unlikely from genuine biology and would flag a modeling artifact. We also check agreement with the direction of known SLE interferon-pathway biology, where it is established.
References
- scPrediXcan integrates advances in deep learning and single-cell data into a powerful cell-type–specific transcriptome-wide association study framework
- Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations
- SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data
- Bayesian model selection reveals biological origins of zero inflation in single-cell transcriptomics
- A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A large-scale transcriptome-wide association study (TWAS) of ten blood cell phenotypes reveals complexities of TWAS fine-mapping
- Partitioning gene-based variance of complex traits by gene score regression
- Single-cell deconvolution of 3,000 post-mortem brain samples for eQTL and GWAS dissection in mental disorders