Most gene-expression pipelines quietly assume that every sample contains about the same total amount of RNA. Under that assumption, which the literature calls "biological stasis," dividing by sequencing depth recovers meaningful per-gene changes. Tumors break the assumption routinely. If a cancer doubles its total messenger RNA, a gene whose copy number never moved shows up at half its former share and gets called "down-regulated." In simulations with varying transcriptome size, edgeR's false discovery rate reaches 54.89% (Cai et al.). Methods built on within-sample gene-pair orderings (is gene A above gene B in this sample?) avoid this, because a global rescaling of one sample cannot reorder its genes. They have two weaknesses. They see only full order flips, so subtle shifts are missed. And they treat orderings as immune to lab-to-lab technical differences, which they are not: about 10% of pairs reorder purely from technical variation (Liu et al.). Standard batch correction such as ComBat has the opposite problem. It handles lab differences but works on depth-normalized values. When tumors and normals come from different labs, it can absorb a real global RNA expansion into the "batch effect" and subtract it.
We propose a single test that addresses all three problems. For each gene, we compute its continuous log-ratio against a fixed panel of well-measured anchor genes inside each sample. That step cancels any per-sample global scale, whether it comes from sequencing depth or from true RNA amplification. We then estimate gene-specific lab offsets on those ratios, using only samples of the same phenotype that appear in more than one lab. Finally, we compare phenotypes with a stratified rank test on the corrected ratios. The claim is that this matches the near-zero false discovery rate of order-flip methods while recovering much of the power of count-based tools, including in multi-lab designs where phenotype and lab are partly confounded. It also states explicitly when the design gives no information at all.
Core Claim
Write the measured log-expression of gene $g$ in sample $s$ as
$$\log y_{gs} = c_s + \theta_{g,k(s)} + \gamma_{g,b(s)} + \varepsilon_{gs},$$
where $c_s$ is a per-sample constant absorbing library size and total-RNA content, $\theta_{g,k}$ is the biological level of gene $g$ in phenotype $k$, $\gamma_{g,b}$ is a gene-specific offset for lab or batch $b$, and $\varepsilon_{gs}$ is noise with no assumed distribution.
ComBat's model (pyComBat) has no $c_s$ term separate from what depth normalization removes. As a result, any phenotype-linked global expansion leaks into $\gamma$ or $\theta$. Taking within-sample differences against an anchor gene $a$ removes $c_s$ exactly:
$$d_{gas} = (\theta_{g,k} - \theta_{a,k}) + (\gamma_{g,b} - \gamma_{a,b}) + (\varepsilon_{gs} - \varepsilon_{as}).$$
The batch term is constant within a lab. It can therefore be estimated robustly, for example as a median of between-lab differences of $d_{gas}$, from any phenotype observed in two or more labs. After subtracting it, the phenotype contrast $(\theta_{g,k}-\theta_{a,k})$ is tested with a stratified Wilcoxon-type statistic, with null calibration by permuting phenotype labels within labs.
The pair-level estimand needs no stasis assumption. Calling a single gene "changed" does require one assumption: that the median anchor is stable in absolute copies per cell. This is the same kind of anchor logic RankCompV2 uses. We state it as an assumption and stress-test it rather than hiding it. When phenotype and lab are fully confounded with no bridging labs, the offsets are not identifiable. In that case the method reports the design as uninformative instead of producing calls.
Relation to Prior Work
Three papers motivate the gap. Cai et al. introduce RankCompV2 and prove that order reversals against stable partners imply absolute change. They also acknowledge low power for shifts that do not flip an order, and they model no batches. The review by Zheng et al. argues that rank-based pairwise analysis aids cross-platform integration, but it offers no explicit technical-covariate model for unbalanced designs. Cao et al. show across 6,590 tumors that total tumor mRNA varies widely and tracks progression. Global amplification is therefore a real biological signal that a batch method must not erase. Their deconvolution, however, needs matched DNA data.
Two papers constrain the design. Liu et al. find that pair orderings degrade below 250 pg input. For that reason we restrict to bulk data and choose anchors by high expression and low dropout. reComBat shows that ComBat breaks when batch and biology are entangled across many public batches. This tells us that the offset-estimation step needs regularization and an explicit check of which labs are connected through shared phenotypes.
Ramdas et al. supply the method. They show that rank-based two-sample tests are distribution-free while optimal-transport tests degrade in high dimension. That result is why we run many one-dimensional rank tests on log-ratios rather than one multivariate transport test. We deliberately do not use the permutation-space consensus trees of Clémençon and Irurozki: they are demonstrated at around 150 items, far below a transcriptome.
Research Program
We will build the estimator and test in four steps.
The first step selects the anchors: roughly 200 to 500 genes, chosen by expression level and cross-lab stability in a held-out reference set. Restricting to anchors reduces the number of pairs from about $G^2$ to about $500G$.
The second step estimates lab offsets on anchor-ratios. We fit them by median polish over the graph of labs that share a phenotype, with shrinkage across genes in the spirit of empirical Bayes. A connectivity check flags labs that no bridge reaches.
The third step computes a stratified Hodges-Lehmann shift and a van Elteren statistic per gene-anchor pair. These are aggregated to a gene-level score as the median over anchors, with false discovery control by within-lab permutation.
The fourth step is analysis. We will prove invariance to $c_s$ and identify exactly which contrasts are estimable for a given lab-phenotype design. We will also characterize when the gene-level call fails, namely when more than half of the anchors change in absolute terms.
Baselines
The naive reference is TMM normalization followed by a per-gene Wilcoxon test that ignores lab. It represents what most analysts do, and in the transcriptome-size simulation its count-based cousin edgeR reaches a false discovery rate as high as 54.89% (Cai et al.). Beating it means a false discovery rate near the nominal level under amplification.
The strongest existing method for this question is RankCompV2 itself. It reports a false discovery rate below 0.15% with sensitivity above 95% and roughly 98% directional concordance across independent datasets in its simulation (Cai et al.). That result has no batch structure, and the reported sensitivity is not broken down by effect size. We will run RankCompV2 on identical simulated and real data, with the same gene filtering, and pool its calls across labs both naively and by per-lab meta-analysis. Beating it means equal false discovery control together with materially higher sensitivity on changes below the size that flips an order, for example log-fold changes under 0.5. In multi-lab designs, beating it also means lower false discovery rate where it treats technical reorderings as biology.
The strongest batch-correction competitor is ComBat via pyComBat, or reComBat when designs are singular. Phenotype is supplied as a covariate, and the output is followed by both a Wilcoxon test and RankCompV2. No reported result found on Lacuna for ComBat or reComBat under transcriptome-size variation. All competitors get the same anchor and gene filters and the same compute budget. Tuning (shrinkage strength, number of anchors) is done on simulation seeds disjoint from the evaluation seeds, and the same number of configurations is tried for every method.
Two ablations isolate the ingredients. The first drops the lab-offset step, leaving the anchor-ratio rank test alone. The second replaces continuous ratios with binary orderings while keeping the offset correction. The first ablation tests whether batch modeling matters. The second tests whether keeping magnitude is what buys power.
Why not just…?
Why not just block on lab with a stratified rank test on normalized values? A van Elteren test on TMM values handles lab offsets whenever both phenotypes appear in each lab. It still inherits the normalization error, because TMM divides by a denominator that amplification inflates. The first ablation, run on amplification-only simulations, will show whether its false discovery rate tracks edgeR's.
Why not just run ComBat first and then RankCompV2? This is a real threat in balanced designs. ComBat with a phenotype covariate preserves $\theta$ there, and pair orderings afterwards remove most of $c_s$. They do not remove all of it, because ComBat rescales each gene by its own variance factor, which turns a common sample scale into gene-specific distortions. If the experiments show this pipeline matching ours in balanced designs, we will narrow the claim to partially confounded designs and to power for subtle shifts. Those are the regimes where ComBat must either drop the phenotype covariate or face a near-singular fit.
Why not just apply RankCompV2 within each lab and combine p-values? This avoids batch modeling entirely, but it wastes every lab that holds only one phenotype, and it keeps the order-flip power ceiling. Comparing the two on the 114-batch compendium from reComBat, where many batches hold a single condition, will show how many testable contrasts each approach retains.
Why not just rescale samples by estimated total mRNA? The tumor total-mRNA score of Cao et al. needs matched DNA sequencing for purity and ploidy, which most cohorts lack. Where it exists (TCGA), we use it as an external check rather than as a competitor.
Proposed Experiments
The first experiment extends the transcriptome-size simulation of Cai et al. with lab structure. Labs add gene-specific offsets drawn from offsets estimated on real GEO batches. Designs range from balanced, to partially confounded with one or two bridging labs, to fully confounded. The amplification factor varies from 1 to 3, and true changes are stratified by size. The metrics are false discovery rate, sensitivity by effect size, and the rate of correct "uninformative design" flags.
The second experiment is semi-synthetic, so that the ground truth is real biology. We take normal-tissue samples spanning several real labs, split them into pseudo-groups, and inject known absolute changes plus a global multiplication of one group's counts. We then downsample both groups to a fixed library size, so that measurements become relative as in real sequencing.
The third experiment is a technical null. GSE50856 profiles the same cells under different amplification protocols, so protocol acts as lab with no biology to find. Any call is a false positive.
The fourth experiment uses real tumors. We split TCGA tumors from one cancer type by high versus low total-mRNA score. A method should not call a gene "down" merely because its group has more total RNA. We will also report the LIHC cancer-driver enrichment and cross-dataset directional concordance used by Cai et al.
The decisive check is the semi-synthetic partially confounded design with amplification factor 2. If our method does not hold the false discovery rate near nominal while exceeding RankCompV2's sensitivity on sub-flip changes, the proposal fails. It also fails if ComBat followed by RankCompV2 does equally well.
References
- Identification and characterization of genes with absolute mRNA abundances changes in tumor cells with varied transcriptome sizes
- Robust transcriptional signatures for low-input RNA samples based on relative expression orderings
- pyComBat, a Python tool for batch effects correction in high-throughput molecular data using empirical Bayes methods
- Less is more: relative rank is more informative than absolute abundance for compositional NGS data
- Estimation of tumor cell total mRNA expression in 15 cancer types predicts disease progression
- reComBat: batch-effect removal in large-scale multi-source gene-expression data integration
- On Wasserstein Two-Sample Testing and Related Families of Nonparametric Tests
- Beyond Kemeny Medians: Consensus Ranking Distributions Definition, Properties and Statistical Learning