Hospital archives hold enormous numbers of tumor blocks preserved by formalin fixation and paraffin embedding (FFPE), the standard chemical preservation used in pathology. These blocks are linked to years of patient follow-up. When researchers sequence RNA from them and ask "which biological pathways change between groups?", they usually run overrepresentation analysis. This test checks whether a pathway's genes appear among the differentially expressed genes more often than chance would predict. We argue that two systematic effects break this test on archival tissue, and that both can be repaired at the level of the test rather than the data.
The first effect is degradation. Fixation makes some genes far easier to call differentially expressed than others: abundant, fragment-robust transcripts pass the threshold, while faint ones do not. The second effect is cell composition. A pathway can look "activated" simply because the cell type that expresses it grew in number. Our proposal is an enrichment test that asks whether pathway membership predicts differential-expression status after conditioning on each gene's FFPE detectability and its predicted composition-only fold change. We would validate it against matched fresh-frozen tissue as ground truth.
Core Claim
The textbook enrichment null assumes every gene is equally likely to be called differentially expressed, apart from its biology. In an all-FFPE cohort this assumption is false in a specific, measurable way. Whole-genome DASL profiling, a probe-based assay built for degraded RNA, shows the smallest detectable fold change rising from 1.3–1.5 in intact RNA to 1.5–2.0 in FFPE. The same work shows agreement with an independent PCR measurement for low-abundance transcripts falling from $R^2 \sim 0.87$ to $R^2 \sim 0.55$ (Whole-Genome Gene Expression Profiling of Formalin-Fixed, Paraffin-Embedded Tissue Samples). Paired fresh-frozen and FFPE sequencing finds 1,494 genes that diverge recurrently between preservation methods across tissue types (Next-Generation Sequencing of RNA and DNA Isolated from Paired Fresh-Frozen and Formalin-Fixed Paraffin-Embedded Samples of Human Cancer and Normal Tissue).
Two consequences follow. When every sample is FFPE, fixation is not a sample-level covariate that a regression can remove. It is a gene-level difference in statistical power, and power differences bias which pathways look enriched. Composition shifts behave the same way. Given a single-cell reference, each gene has a predictable fold change produced purely by the change in cell mixture between groups.
Both effects are therefore gene-level nuisance covariates. They can be built into the enrichment null in the same way gene-length bias is handled. The test we propose fits
$$\operatorname{logit} P(D_g = 1) = f(a_g, c_g) + \beta \, \mathbb{1}[g \in S],$$
where $D_g$ indicates that gene $g$ is called differentially expressed, $a_g$ is its FFPE detectability score, $c_g$ is its composition-only expected log fold change, $f$ is a flexible smooth function fit genome-wide, $S$ is the pathway, and $\beta$ is the enrichment effect we test.
The composition term is $c_g = \log \sum_k \bar\pi^{B}_k x_{gk} - \log \sum_k \bar\pi^{A}_k x_{gk}$. Here $x_{gk}$ is the reference expression of gene $g$ in cell type $k$, and $\bar\pi^{A}_k, \bar\pi^{B}_k$ are the mean estimated fractions of cell type $k$ in groups A and B.
Using group-mean fractions rather than per-sample fractions is a deliberate choice. Deconvolution of fixed tissue is noisy and biased, and averaging over a group damps that noise. Per-sample covariates instead push the noise straight into every gene's test.
Relation to Prior Work
Three deconvolution lines motivate the composition term and supply the fractions. TAPE is an autoencoder whose decoder weights act as a tissue-adapted cell-type signature matrix. POPPY aligns bulk profiles to a single-cell atlas using the Cell Ontology. Count Bridges deconvolves integer counts with a Poisson birth-death bridge. All three assume intact RNA, and none feeds its output into a pathway test.
We do not propose another deconvolver. We reuse their fractions, and we measure how far errors in those fractions propagate into enrichment calls. To do this, we run all three plus a reference-free linear unmixing and treat the disagreement among them as uncertainty, propagated by resampling.
ActivePathways supplies the closest existing enrichment method. It merges gene P-values across omic layers with Brown's method and then runs a ranked hypergeometric test. It does not model detectability or composition, so we use it as a baseline, not a component.
The FFPE papers act as constraints. They tell us the size and shape of the detectability effect, and the paired cohort provides the ground truth. The iST benchmark (Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues) compares imaging methods that read RNA in place on FFPE sections. Its matched FFPE sections plus single-cell data give a second ground truth in which within-cell-type changes can be read off directly.
Transcript-protein discordance plays a deliberately limited role here. Up to 40–60% of protein variance is not explained by mRNA (The Discordance Between Transcript and Protein Expression). A calibrated transcript-level enrichment is therefore still not a claim about protein-level signaling. We report this as a separate annotation rather than folding it into the test. For each significant pathway, we flag what fraction of its supporting genes have documented low mRNA-protein correspondence. The reference for this is the CITE-seq panels used by scChord, which pair RNA with surface-protein measurements in the same cells; scChord itself learns to predict protein from RNA. These panels cover only surface proteins in blood, so the annotation is coarse and we say so.
We also do not attempt network propagation in the style of prize-collecting Steiner trees, which link measured hits through unmeasured hub proteins in an interaction network (Huang et al.; Tunçbağ et al.). Those methods depend on fresh-lysate proteomics and lie outside our scope.
Research Program
The work has three parts.
The first part estimates $a_g$. From the 38 paired fresh-frozen/FFPE specimens, we estimate for each gene the probability that a fold change present in fresh-frozen tissue is recovered at the same direction and significance in FFPE. We model this as a smooth function of abundance and of membership in the 1,494-gene artifact set. We then check that this function transfers to the DASL characteristics and to held-out tissue types. If $a_g$ does not transfer across tissues, the method is cohort-specific, and we will say so.
The second part estimates $c_g$ from atlas references (3CA) and the deconvolution ensemble. We propagate fraction uncertainty by refitting the test across resampled fraction estimates.
The third part builds a semi-synthetic benchmark. Pseudo-bulks are drawn from 3CA and Kassandra cell profiles, with known pathway perturbations planted either in a single cell type (cell-autonomous) or produced only by a composition shift. These pseudo-bulks are then passed through an FFPE corruption model fit to the paired data: compression, a raised detection threshold, and artifact-gene distortion.
Baselines
The naive reference point is standard hypergeometric enrichment, with GO and Reactome gene sets, on bulk FFPE differentially expressed genes. Every method receives the same gene sets, the same gene-level calls, and the same multiple-testing correction.
The first targeted baseline removes the 1,494 recurrent artifact genes before running that same test. It is the obvious FFPE fix, and it costs nothing.
The second baseline is the common practice of adding per-sample deconvolved fractions as covariates in the differential-expression model, followed by enrichment. We run it with TAPE, Count Bridges and POPPY fractions under identical tuning budgets, and we report the best of the three. This makes it the strongest competitor on the composition question.
The third baseline runs TAPE's cell-type-specific profiles through per-cell-type enrichment.
The fourth is ActivePathways' ranked test on the same P-values. This tests whether an existing integrative enrichment method already absorbs these biases.
For the strongest baseline we found no reported pathway-level accuracy, FFPE-versus-fresh-frozen agreement, or false-positive rate on Lacuna. The papers report gene-level and discovery-count figures instead. ActivePathways found enriched pathways in 89% of PCAWG cohorts. DASL reports 74% overlap of differentially expressed genes between preservation methods. We will use the 74% gene-level overlap as a reference for how much pathway-level agreement could plausibly be lost.
Two ablations isolate our ingredients: dropping $a_g$ and dropping $c_g$. A third replaces both with permuted covariates, which controls for the extra model flexibility alone.
Beating the baselines means two things. In semi-synthetic null settings, our false-positive rate must stay near nominal where the baselines' does not. On real paired data, agreement with fresh-frozen pathway calls must exceed the best baseline by more than the bootstrap interval.
Why not just…?
Why not just drop the 1,494 artifact genes and run ordinary enrichment? That set captures genes whose level shifts with fixation. It does not capture the power gradient: a gene can carry no systematic shift yet still be hard to detect at 1.5–2.0-fold thresholds. Dropping genes also removes pathway members outright, which itself biases the test. The ablation that keeps only artifact-set membership in $a_g$ directly measures how much of the gain comes from this simpler idea. If it matches the full detectability model, we will adopt it.
Why not just put deconvolved fractions into the differential-expression model? This is the real threat. With accurate fractions it may suffice. Our argument is that fractions estimated from fixed tissue are biased, since no current deconvolver models degradation, and per-sample covariates inject that error into every gene's statistic. The semi-synthetic benchmark settles the question by varying deconvolution error. We expect the covariate approach to match ours at low error and to lose calibration as error grows. If it does not lose calibration, the composition half of our contribution reduces to a cheaper equivalent, and we will report it as such.
Why not just deconvolve into cell-type profiles with TAPE and run enrichment per cell type? TAPE struggles with rare and closely related cell types (TAPE), and its profiles inherit uncorrected FFPE compression. Per-cell-type enrichment also multiplies the number of tests. We include TAPE as a baseline. The spatial benchmark, where within-cell-type truth is observable, shows whether its per-cell-type pathways are more accurate than ours.
Proposed Experiments
The decisive real-data check uses the paired cohort's tumor-versus-normal contrasts within tissue, for example colon carcinoma versus normal colon. We run every method on the FFPE side and score precision, recall, and rank agreement of pathway calls against the same analysis on fresh-frozen.
The second real-data check uses the iST tumor and normal FFPE microarray sections. Pseudo-bulking them gives a bulk input, while segmented single-cell counts give true within-cell-type change. We ask how many bulk-enriched pathways each method attributes correctly to composition versus cell-autonomous change. Here we must account for segmentation errors in stromal and dense regions, and we will restrict truth to well-segmented regions.
The semi-synthetic benchmark measures calibration, using QQ plots under composition-only and degradation-only nulls, and power under planted signals.
A fourth experiment tests transfer: we fit $a_g$ on some tissue types and evaluate on the others.
The proposal fails cleanly if neither covariate improves fresh-frozen agreement beyond the artifact-removal and fraction-covariate baselines.
References
- Whole-Genome Gene Expression Profiling of Formalin-Fixed, Paraffin-Embedded Tissue Samples
- Next-Generation Sequencing of RNA and DNA Isolated from Paired Fresh-Frozen and Formalin-Fixed Paraffin-Embedded Samples of Human Cancer and Normal Tissue
- Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Integrative pathway enrichment analysis of multivariate omics data
- Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- Network Modeling Identifies Patient-specific Pathways in Glioblastoma