Source-linked AI summary
Differential meta-analysis of RNA-seq data from multiple studies
Andrea Rau, Guillemette Marot, Florence Jaffrézic
TL;DR
Limited biological replication reduces RNA-seq power, motivating methods that integrate multiple related studies. The paper adapts inverse-normal and Fisher p-value combinations and compares them with a fixed-study-effect negative binomial GLM. The GLM performs well with low inter-study variation and few studies, whereas meta-analysis methods perform better with moderate to large variation and more studies.
Problem
Few biological replicates limit RNA-seq differential-expression power, motivating integration of additional related studies despite RNA-seq’s discrete, heterogeneous data structure.
Method
The paper combines per-study RNA-seq p-values using inverse-normal and Fisher methods and compares them with a negative binomial GLM including a fixed study effect.
Results
Meta-analysis methods outperform the fixed-study-effect GLM under moderate to large inter-study variability and larger numbers of studies, while the GLM performs well under low variability and few studies.
Takeaways & Limitations
P-value combination methods provide a valuable approach for differential meta-analysis of RNA-seq data across related studies.
Takeaways & Limitations
The methods require every study to include all conditions and must address differences in objectives, designs, populations, technologies, library preparation, and laboratory effects.
Abstract
from arXiv · showhide
High-throughput sequencing is now regularly used for studies of the transcriptome (RNA-seq), particularly for comparisons among experimental conditions. For the time being, a limited number of biological replicates are typically considered in such experiments, leading to low detection power for differential expression. As their cost continues to decrease, it is likely that additional follow-up studies will be conducted to re-address the same biological question. We demonstrate how p-value combination techniques previously used for microarray meta-analyses can be used for the differential analysis of RNA-seq data from multiple related studies. These techniques are compared to a negative binomial generalized linear model (GLM) including a fixed study effect on simulated data and real data on human melanoma cell lines. The GLM with fixed study effect performed well for low inter-study variation and small numbers of studies, but was outperformed by the meta-analysis methods for moderate to large inter-study variability and larger numbers of studies. To conclude, the p-value combination techniques illustrated here are a valuable tool to perform differential meta-analyses of RNA-seq data by appropriately accounting for biological and technical variability within studies as well as additional study-specific effects. An R package metaRNASeq is available on the R Forge.
1 Background
RNA-seq studies often have few biological replicates, limiting differential-expression detection power, while follow-up studies create a need to integrate related datasets. This paper adapts meta-analysis methods to RNA-seq and compares them with existing integrated-analysis strategies.
- Few biological replicates in typical RNA-seq experiments can limit power to detect differential expression.
- Microarray meta-analysis methods increase effective sample size by integrating related datasets, but many are not directly applicable to discrete RNA-seq counts.
- The paper adapts inverse normal and Fisher p-value combinations and compares them with per-study analyses and a negative binomial GLM with a fixed study effect.
- Integrated analyses assume that every study includes all experimental conditions, because otherwise condition and study effects can become confounded.
2 Methods
The methods obtain per-study RNA-seq p-values under negative binomial models, combine them across studies, and compare them with a global negative binomial GLM. The combination procedures require calibrated null p-values and attention to conflicting expression directions across studies.
- P-value combination from independent analyses: Per-study differential analysis models gene counts with a negative binomial distribution and produces raw per-gene p-values.
- Inverse normal method: The inverse normal method transforms each study’s p-value with the standard normal quantile function and combines the transformed values using study-specific weights.
- Inverse normal method: Studies receive larger inverse-normal weights when they contain more biological replicates, although users may define alternative weights based on data quality.
- Inverse normal method: The inverse-normal statistic follows N(0, 1) under the null, enabling a right-tailed test followed by false-discovery-rate correction.
- Fisher combination method: Fisher’s method combines the per-study p-values into a statistic that follows a χ2 distribution with 2S degrees of freedom under the null.
- Additional considerations: P-value combination requires approximately uniform null p-values, so low-count genes are filtered when discretization creates a peak near 1.
- Additional considerations: Genes with conflicting expression directions across studies are identified and removed from the final differential-expression list.
- Global differential analysis: The global analysis fits negative binomial GLMs with condition and study in the full model and study alone in the reduced model, comparing them with a χ2 likelihood-ratio test.
3 Results and Discussion
In real melanoma RNA-seq data, filtering corrected nonuniform per-study p-values before combination, while Fisher and inverse-normal meta-analysis increased detection relative to intersecting studies. Simulations showed that fixed-study GLMs work well under low variability, whereas meta-analysis methods gain advantage as inter-study variability and study count increase.
- Real-data analysis: Filtering weakly expressed genes made the per-study p-values appear to satisfy the null-uniformity assumption required for combination methods.The unfiltered data, especially Study B, showed a peak near 1 caused by p-value discretization.
- Real-data analysis: 3,578 genes were shared by the meta-analysis methods and global analysis, compared with 1,583 genes in the intersection of individual studies.The intersection approach was described as conservative, while uniquely identified and shared genes included potential cancer- or melanoma-related candidates.
- Simulation study: For σ = 0 or 0.15, methods showed no practical ROC difference, but differences became apparent at σ = 0.3 or 0.5 across study counts.The simulations evaluated 2, 3, and 5 studies using settings based on real RNA-seq experiments.
- Simulation study: At large inter-study variability, Fisher and inverse-normal meta-analysis produced larger AUCs than the fixed-study global analysis, while naive global analysis had the lowest AUC.Including a study effect improved performance over the naive global analysis without one.
- Simulation study: Meta-analysis improved sensitivity over global analysis for moderate-to-large variability with three or more studies, while all methods controlled FDR below 5%.With two studies and σ = 0.50, meta-analysis and the fixed-study global model had similar detection power.
- Simulation study: For two studies with large inter-study variability, roughly 80% of true positives were uniquely detected by either Fisher meta-analysis or the fixed-study global model.With more than two studies, the proportion uniquely detected by Fisher increased relative to the global analysis as variability increased.
4 Conclusions
The proposed p-value meta-analysis methods perform similarly and increasingly outperform a fixed-study-effect global GLM as inter-study variability and the number of studies increase. Their practical scope includes RNA-seq studies with shared conditions, while differences across designs, populations, technologies, and laboratories remain important constraints.
- Conclusions: Meta-analysis methods gain significantly over the global analysis as inter-study variability increases, particularly with more than two studies.The gains concern AUC, sensitivity, and the proportion of true positives among uniquely identified genes.
- Conclusions: Figure 4 compares sensitivity across 2, 3, and 5 studies at σ = 0.15 and σ = 0.50 for five analysis strategies.The strategies are individual analyses, inverse normal, Fisher, DESeq without a study effect, and DESeq with a fixed study effect.
- Conclusions: Figure 5 compares the proportion of true positives among unique discoveries for Fisher and DESeq with a fixed study effect across study counts and inter-study variability.Error bars show one standard deviation, while parenthetical values give mean total unique discoveries for the two methods.
- Conclusions: Both p-value combination methods are implemented in the R package metaRNASeq, available on R-Forge.
- Limitations: The methods are intended for analyses in which every study includes all experimental conditions, avoiding condition–study confounding.The authors also note differences in objectives, designs, populations, sequencing technology, library preparation, and laboratory effects.
Authors’ contributions
The authors describe their respective roles in study design, simulations, data analysis, package development, manuscript drafting, and study conception.
- Authors’ contributions: AR conducted simulations and data analyses, GM wrote the associated R package, and FJ conceived the study.All authors contributed to study design or manuscript drafting, and approved the final manuscript.