TWAS sign errors reach 33% in complex-trait analyses

A large genomic study found that transcriptome-wide association studies, or TWAS, frequently assign the wrong direction to a gene’s effect on a trait. In analyses reported in Nature Genetics, the estimated false sign rate was 23% for plasma proteins and 33% for complex traits, based on comparisons with genetic burden tests that provide an independent expectation for the direction of gene effects. The results show that a statistically significant TWAS association can point toward the opposite of a gene’s expected biological influence.

TWAS is widely used to connect genetic variants with genes and to infer whether increased expression of a gene is likely to raise or lower a trait or disease risk. It does so by combining genome-wide association study summary statistics with models that predict gene expression. The resulting associations can be valuable for nominating causal genes, but their signs also carry a mechanistic interpretation: a positive coefficient might be read as evidence that more gene expression increases a phenotype, while a negative coefficient might suggest the opposite. The new study specifically examined the reliability of those directional interpretations.

The analysis evaluated TWAS signs against loss-of-function and gene-duplication burden tests. Such tests aggregate disruptive variants that reduce or increase gene dosage, providing what the authors described as a “gold standard” for the expected direction of a gene’s effect. The approach was tested first with 2,923 plasma proteins measured in the UK Biobank Pharma Proteomics Project. Loss-of-function variants reduced the abundance of 98.6% of proteins detected as significant in those tests, whereas duplications increased the abundance of 100% of proteins detected as significant. These observations supported the expected positive relationship between gene expression and abundance of the encoded protein, even though RNA and protein levels are imperfectly correlated.

Using FUSION TWAS models trained on 47 GTEx tissues, the researchers then estimated genetically predicted expression for genes encoding those plasma proteins. Sixteen proteins were excluded because of quality problems or because more than one gene encoded the same protein, leaving 2,907. Of these, 1,543 had at least one statistically significant TWAS association after Bonferroni correction and could be used to estimate directional accuracy. Among them, 23% had a TWAS direction opposite to the burden-test expectation, with a 95% confidence interval of 20% to 25%. Requiring stronger agreement among tissue categories reduced the error only modestly, indicating that ambiguous voting did not account for most of the problem. In practical terms, roughly one in four genes with an evaluable association had a sign that contradicted the expected expression-to-protein relationship.

One example was DDT, which encodes D-dopachrome tautomerase. Variants predicted to increase DDT expression were associated with lower plasma levels of the encoded protein, the reverse of the expected relationship. By contrast, variants linked to higher PON1 expression in the liver also corresponded to higher plasma paraoxonase 1 levels, as expected. The analysis also showed that many nearby genes were labeled significant: before colocalization filtering, 85.3% of protein-encoding genes had at least one significant TWAS gene within two megabases, consistent with numerous signals tagging a causal locus rather than acting through the named gene. After filtering, that proportion fell to 43.3%, although many nearby associations remained.

Colocalization can help by asking whether an expression association and a trait association are likely to share a causal variant. The team used the posterior probability for a shared causal variant, PPH4, from both coloc and a fine-mapping-aware version called coloc-SuSiE. Stricter filters lowered the false sign rate, but also discarded many associations, and the gene-level error remained above 10% even at the strictest threshold tested, PPH4 = 0.99. Simulations indicated that high linkage disequilibrium between distinct causal variants was the largest contributor to persistent errors. In such cases, a TWAS model can correctly implicate the right genomic haplotype while attributing the trait effect to the wrong regulatory variant, so colocalization may establish a shared signal without establishing the correct tissue or direction of effect.

The researchers also tested newer methods intended to address local genetic confounding. cTWAS produced a 24% gene-level false sign rate for proteins, compared with 23% for FUSION, while multigroup cTWAS, which jointly analyzes multiple tissue datasets, yielded similar performance and recall. Filtering by posterior inclusion probability reduced errors somewhat but caused a substantial loss of associations. The protein results therefore appeared to reflect the underlying architecture of expression and trait signals rather than a problem confined to one TWAS implementation. The analyses also distinguished false signs from the broader question of whether a TWAS association identifies the correct gene, because the study’s primary measurement was the agreement between inferred direction and the burden-test benchmark.

For complex traits, the researchers assembled 730 gene–trait pairs with significant loss-of-function burden-test results from 85 UK Biobank traits. TWAS detected at least one significant association for 484 pairs. Summarizing signs across tissue categories revealed a 33% discordance rate, with a 95% confidence interval of 29% to 38%. In 12% of all evaluated gene–trait pairs, TWAS pointed in the wrong direction in more than 90% of tissues with available expression models. Colocalization again reduced the error at the cost of recall, while a model of gene density, recombination, gene constraint, cis-heritability and trait monotonicity did not identify any single predictor that convincingly explained the differences among loci. The study used burden-test pairs as a reference, and the authors cautioned that genes selected by this method may not represent all genes involved in a trait.

The GCK locus illustrated how statistical confidence can coexist with biological misdirection. Loss-of-function variants in GCK are associated with increased blood glucose, and increased GCK activity in pancreatic beta cells is expected to increase insulin secretion and lower glucose. Yet the analysis found little expression-QTL signal in pancreas, preventing a significant TWAS association there. Strong positive TWAS signals instead appeared in 17 other tissue categories, including tibial nerve and thyroid. The thyroid association reached a Z-score of 33.2, and its expression and glucose signals appeared to colocalize, yet both pointed in the direction opposite to established GCK biology. The authors proposed that such results can arise when regulatory variants in non-trait-relevant tissues are in very high linkage disequilibrium with the true causal variant and carry a different sign.

Simulations involving 1,000 genes supported this mechanism. False signs were most common when expression effects in non-causal tissues were strongly anticorrelated with those in the blood, the designated causal tissue, and became less frequent as cross-tissue co-regulation increased. Even after coloc filtering, strongly anticorrelated tissues remained enriched among false-sign associations, showing that apparent sharing of a genetic signal cannot guarantee the correct tissue or direction. The authors also identified other possible contributors, including opposite expression effects across tissues, co-regulation among neighboring genes, non-monotone dosage relationships, mismatches between linkage-disequilibrium reference panels and study cohorts, and uncertainty in expression-imputation models. These mechanisms were presented as possible contributors rather than individually proven explanations for every discordant result.

Restricting analyses to a protein’s tissue of origin modestly improved protein results: the false sign rate fell from 23% to 20% for 228 tissue-specific proteins. Excluding brain-specific proteins lowered it further to 14%. Combining tissue-of-origin information with coloc filtering at PPH4 of at least 0.5 produced no discordant cases among the 40 retained genes, although the sample was small; coloc-SuSiE yielded a 7% rate, with a 95% confidence interval of 2% to 17%, while retaining 58 genes. The findings support evaluating TWAS signs alongside false discovery rates and treating tissue relevance and colocalization as complementary safeguards. The results do not, by themselves, settle TWAS’s value for gene discovery, but they show that its inferred directions require biological and genetic context rather than reliance on association strength alone.

Subject of Research: Accuracy of gene-effect direction estimates from transcriptome-wide association studies

Article Title: High false sign rates in transcriptome-wide association studies

Article References: Gerlach, P. A., Milind, N., Spence, J. P., & Pritchard, J. K. (2026). High false sign rates in transcriptome-wide association studies. Nature Genetics. https://doi.org/10.1038/s41588-026-02796-4

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02796-4

Keywords: transcriptome-wide association studies, TWAS, false sign rate, gene regulation, linkage disequilibrium, colocalization, eQTL, GWAS, plasma proteins, complex traits, High, false