Most measurements of gene activity in tumor biopsies count RNA reads that land on a gene's coordinates and treat the total as "how much of this gene's message is present." DNA has two strands, though, and many genes overlap an antisense transcript: an RNA copied from the opposite strand at the same location. Sequencing protocols that ignore read direction ("unstranded") fold those antisense reads into the gene's count. This proposal asks how much of the well-known mismatch between transcript and protein levels, for metabolic enzymes in particular, comes from that folding. Part of the mismatch may be a measurement artifact, where antisense reads inflate the apparent message. Part may be real regulation, where the antisense RNA changes how a message is processed. A third part may come from mixing many cell types in one biopsy, so that a shift in antisense level reflects a shift in which cells are present rather than any change within cells.
This matters because enzyme transcript levels are routinely used as stand-ins for metabolic activity. Tools that infer metabolic flux (the rates of biochemical reactions) from RNA, such as GEFMAP, concede that expression is an imperfect proxy for reaction rate. The HUPO blueprint in A high-stringency blueprint of the human proteome is cited for the claim that mRNA tracks protein only 40–60% of the time. If a measurable share of that gap is caused by strand handling and cell mixing, it can be corrected with data that already exist. The method also produces a list of enzyme genes whose transcript counts should not be trusted as protein proxies.
Core Claim
We claim that the transcript-to-protein residual for metabolic enzymes in bulk tumors has three separable parts. The first is a strand-leakage artifact: antisense reads counted as sense in unstranded quantification. The second is an antisense regulatory effect: a contribution of antisense level to protein that remains after sense message is measured correctly. The third is a composition effect: variation in antisense and sense levels that is explained by cell-type proportions. These parts can be estimated within a single stranded dataset by computing, for the same samples, the unstranded count a strand-blind pipeline would have produced.
Here is what is new. Culbertson et al. supply the motivating mechanism. They show that NQO1-AS binds the NQO1 3' untranslated region, changes which RNA-binding proteins attach, favors a long, roughly four-fold more stable isoform, and raises NQO1 enzyme activity. That study is one locus in two bulk cell lines, and it has no protein-versus-transcript analysis across samples and no cell-mixture modeling. In the other direction, Tan et al. show in ovarian cancer that proteomic subtypes predict survival where transcriptomic classifications do not, but they work with unstranded single-cell data and do not ask why the transcript and protein layers disagree. No work in the file measures how much of transcript-protein discordance is attributable to strand handling, with or without accounting for composition.
Research Program
The core object is a per-gene, across-sample regression of protein abundance on RNA features, fitted in cohorts that have both mass-spectrometry proteomics and RNA-seq. The gene set is restricted to metabolic enzymes in the Recon3D network that GEFMAP uses, plus the specific enzymes highlighted in the file: NQO1, HK2, the pyruvate dehydrogenase kinases, TYMP, ABCG1 and ACOT8.
For each gene $g$ and sample $i$ we compute four quantities from the same aligned reads:
- $s_{gi}$, the sense-strand exonic count;
- $a_{gi}$, the antisense-strand count over the overlapping region;
- $n_{gi}$, the sense intronic count, which serves as a rough proxy for transcription rate;
- $u_{gi} = s_{gi} + a_{gi}$, the count an unstranded pipeline would report.
We also estimate cell-type proportions $\pi_i$ from sense counts of marker genes only. Antisense features never enter this step, which matters because single-cell reference atlases are strand-blind, as POPPY and the 3CA atlas it uses illustrate. The decomposition then compares nested models for log protein $p_{gi}$. The leakage artifact is the difference in explained variance between a model built on $\log u$ and one built on $\log s$. The composition effect is what $\pi_i$ adds to $\log s$. The antisense regulatory effect is what $\log a$ adds after $\log s$, $\log n$ and $\pi_i$ are all in the model.
Including $n$ is deliberate. An antisense RNA may simply mark an active, open locus. If so, its apparent effect should vanish once the transcription proxy is in the model. A stability-type mechanism like NQO1-AS should survive that control.
Strandedness of public libraries cannot be assumed, so the first step is empirical. For each cohort we infer library orientation from read direction at genes with no antisense overlap, which is the same principle as the strand-based artifact filtering in ShaNTi. Only cohorts confirmed as stranded enter the decomposition.
The intended candidates are the CPTAC 2016 and 2023 ovarian cohorts and the Xiangya cohort used by Tan et al. Tan et al. explicitly report proteomics for all three. RNA availability is confirmed only for the Xiangya single-cell data, which is unstranded. For the CPTAC cohorts, whether matched bulk RNA-seq exists and whether it is stranded must be established as the first step. The MDA-Par and MDA-LM2 lines from Culbertson et al. serve as the positive control, because NQO1-AS there must show a regulatory effect.
As a downstream consequence check, we replace enzyme weights in GEFMAP with strand-corrected and composition-adjusted estimates. We then ask whether the inferred dominant metabolic objective changes in samples where leakage is large.
Baselines
The naive reference point is the standard practice: unstranded gene counts $u$ regressed on protein, gene by gene, with no composition terms. No reported result was found on Lacuna for per-gene mRNA–protein correlation on the CPTAC ovarian cohorts. The only quantitative anchor in the file is the HUPO statement that mRNA levels correlate with protein abundance only 40–60% of the time. We will therefore compute the naive baseline ourselves on identical samples.
The strongest existing competitor for the composition part is deconvolution. Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) models bulk counts as sums of integer single-cell counts and is reported to outperform CIBERSORTx and MuSiC. The file gives no numeric deconvolution accuracy for it, so no reported result was found on Lacuna for Count Bridges proportion error on these cohorts. To stay within budget we also run quanTIseq, which Mock et al. found more robust than CIBERSORT at shallow depth. All deconvolution methods get the same marker-gene inputs and the same reference.
The strongest competitor for predicting protein from RNA is a flexible learner on sense counts alone. We use gradient-boosted trees, plus a totalVI-style probabilistic model (totalVI) where its assumptions transfer. Each gets the same cross-validation folds and the same hyperparameter search budget as our antisense-augmented models. scChord reports lower RMSE and higher Pearson correlation than totalVI, sciPENN and scTranslator on the GSE100866 and GSE164378 CITE-seq benchmarks. That claim concerns single-cell surface proteins, so it is not a direct baseline here, and we say so rather than borrow its numbers.
The key ablation is a locus-permuted antisense feature. Antisense counts are reassigned to random non-overlapping genes, which keeps their distribution while breaking the locus link. Beating this baseline is the minimum evidence that antisense information at the right locus matters.
Beating the baselines means two things. First, across held-out samples, adding antisense features must raise the median per-gene protein variance explained, among enzymes with antisense partners, above the composition-adjusted sense-only model. Second, the gain must exceed the gain from locus-permuted antisense, with significance assessed by paired tests across genes.
Why not just…?
Why not just add antisense counts as one more feature in an off-the-shelf protein predictor? This is essentially what we do, on purpose. The contribution is not a new architecture. It is the controlled decomposition: the computed unstranded counterfactual, the transcription proxy, the composition terms and the locus permutation, which together say why the feature helps. A flexible model given $u$, $s$ and $a$ will happily use antisense as a marker of a cell type or of locus activity. Without the nested controls that gain would be misattributed to regulation.
Why not just deconvolve and stop, since composition may explain most of the discordance? This is a real threat. Tan et al.'s finding that transcript-derived cell fractions miss clinical behavior suggests composition is not the whole story, but it does not rule out that composition explains most of the enzyme residual. The decomposition settles this directly. If $\pi$ absorbs nearly all variance that $a$ would add, we report that antisense effects in tumors are mostly composition. That would be a useful negative result for anyone tempted to read bulk antisense changes as regulation.
Why not just use proteomics and skip RNA altogether? For enzymes whose protein is measured, that is the right answer in the moment. However, proteomic cohorts are few and costly, and roughly 1,900 human proteins are still not confidently detected according to HUPO. Most metabolic inference, GEFMAP included, therefore runs on RNA. Our output is a gene-level map of where RNA proxies are trustworthy, which only a paired RNA–protein analysis can produce.
Proposed Experiments
The decisive check is the leakage counterfactual. Within each stranded cohort we compare protein variance explained by $\log u$ versus $\log s$ across enzymes, ranked by antisense-to-sense ratio. If leakage is real, the drop from $s$ to $u$ should grow with that ratio. If no such trend appears, the artifact story fails.
The second check is the regulatory test under full controls. For enzymes that do have antisense partners, does $\log a$ retain a significant held-out effect beyond $\log s$, $\log n$ and $\pi$, and does it beat the locus-permuted ablation?
The positive control must pass in both settings. In MDA-Par versus MDA-LM2, the NQO1 locus should show an antisense association with the long isoform and with protein. In the cohorts, the direction of the NQO1 effect should match Culbertson et al. If the method cannot recover the one mechanism established in the file, its tumor-wide estimates are not credible.
The final check is downstream. In GEFMAP runs, we measure how often the inferred top metabolic objective changes when enzyme weights are switched from unstranded counts to corrected values, stratified by estimated leakage. Large changes concentrated in high-leakage samples would show that strand handling materially distorts metabolic inference in heterogeneous tumors.
References
- Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning
- A high-stringency blueprint of the human proteome
- A sense-antisense RNA interaction promotes breast cancer metastasis via regulation of NQO1 expression
- Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Transcriptome profiling for precision cancer medicine using shallow nanopore cDNA sequencing
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation