When a cancer biologist compares invasive and non-invasive cells by RNA sequencing (RNA-seq, which counts every messenger RNA in a sample), the result is a list of a thousand or more genes that changed. The list is then summarized by functional enrichment: checking which known pathways, such as "cell adhesion" or "MAPK signaling," contain more changed genes than chance would predict. The resulting pathway story is what gets written up as the driver of cancer spread. Two facts undercut this. First, messenger RNA levels track protein levels only 40–60% of the time (A high-stringency blueprint of the human proteome). Second, the confirming experiment, usually a Western blot (an antibody test for one protein at a time), is slow, so labs check only a few genes. Those genes are typically the largest fold-changes, which says little about whether the pathway-level claim holds at the protein level.
We propose to treat protein validation as a budgeted decision problem. Each pathway gets a probability that it would still be enriched if protein levels were measured. The method then picks the few genes whose protein measurement would most change those probabilities, and updates after each assay. The claim is that with five to twenty targeted assays, chosen this way, pathway calls agree with full-proteome pathway calls much better than they do under unweighted enrichment or under the usual practice of blotting the top hits. The question is not "which gene is most changed" but "which measurement most changes what we would conclude."
What exists and what is missing
Several studies already join RNA-seq, enrichment, and metastasis, and they show the gap directly. The zebrafish serial-selection study (Identifying drivers of breast cancer metastasis in progressively invasive subpopulations of zebrafish-xenografted MDA-MB-231) enriched increasingly invasive MDA-MB-231 cells and nominated CTSD, SERPINE1, DDIT4 and MT1X from RNA-seq. It did no protein validation. The SETD8 hepatocellular carcinoma study (Impact of SETD8/KMT5A overexpression on hepatocellular carcinoma progression and prognosis) reported Gene Ontology enrichment for adhesion and MAPK after knockdown, but it validated only SETD8 itself at the protein level. The RBM47 work (Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer) validated a single downstream protein, DKK1. The Na<sub>v</sub>1.5 study (A novel Nav1.5-dependent feedback mechanism driving glycolytic acidification in breast cancer metastasis) released knockdown RNA-seq (GSE228621) without testing whether the invasion programs it found appear in protein. These four papers motivate the proposal: the field already validates proteins one at a time, but the gene choice is ad hoc and is never connected to the pathway claim.
A cautionary case is the ABC-transporter study (Expression Profiling of ABC Transporters in a Drug-Resistant Breast Cancer Cell Line Using AmpArray). There, ABCA4 rose about 132-fold at the RNA level, yet its protein could not be detected. Picking validation targets by fold-change can therefore spend the budget on exactly the wrong genes.
Our method borrows from two earlier tools. ActivePathways (Integrative pathway enrichment analysis of multivariate omics data) combines evidence from several data types into one enrichment test using Brown's method, a way of merging P-values that accounts for correlation between them. It has been run on DNA and RNA, not protein, and it takes the data as given rather than deciding what to measure next. MoPC (Integrated microRNA and proteome analysis of cancer datasets with MoPC) measures how much microRNAs uncouple RNA from protein across matched patient samples. We reuse its outputs as one predictor of which genes' transcripts will fail to show up as protein changes. The network approach of Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling finds hidden hub proteins on an interaction network. It answers a different question (which unmeasured proteins connect the hits), and we do not compete with it.
Core Claim
Pathway-level conclusions from metastasis RNA-seq can be made protein-robust at small cost, with the gain coming from choosing assays by expected information about pathway calls. Neither better concordance weighting alone nor more assays chosen by fold-change should deliver the same gain. If concordance weighting alone matches the active selection, the proposal reduces to a simple weighting tool, and we will report that.
Research Program
For each changed gene $g$, we model a hidden indicator $z_g$ that is 1 if the protein changes in the same direction as the RNA. Its prior probability $\pi_g$ is predicted from features available in an RNA-only experiment:
- the gene's typical across-tumor mRNA–protein correlation in matched CPTAC data;
- MoPC-style microRNA uncoupling scores;
- evidence of RNA-binding-protein targeting, using the roughly 2,500 RBM47 binding sites as a first source;
- splicing changes;
- protein class, such as membrane, secreted, or channel.
A pathway $P$ counts as protein-enriched if the hypergeometric overrepresentation test, applied to genes with $z_g=1$, passes the chosen false-discovery threshold. Sampling the $z_g$ values gives each pathway a probability $q_P$ of being protein-enriched.
Given these probabilities, the next assay goes to the gene that maximizes the expected reduction in total uncertainty over pathway calls:
$$g^\ast = \arg\max_g \Big[ \sum_P H(q_P) - \mathbb{E}_{z_g}\sum_P H(q_P \mid z_g) \Big],$$
where $H(q)=-q\log q-(1-q)\log(1-q)$ is the binary entropy. In plain terms, this picks the measurement that, on average, most sharpens our yes/no answers about the pathways. Overlap between Gene Ontology terms matters here. One gene shared by several redundant terms can settle many calls at once, and the selection naturally exploits this.
The deliverable is a short ranked assay list per experiment, plus a per-pathway report of how much of its enrichment rests on genes predicted not to translate into protein.
One limitation is built into the design. Our ground truth is total protein abundance. Channels such as ANO1 drive invasion only when trafficked to the cell surface (Suppression of 14-3-3γ-mediated surface expression of ANO1 inhibits cancer progression of glioblastoma cells), so for membrane proteins we flag abundance as necessary but not sufficient evidence.
Baselines
The do-nothing reference is plain overrepresentation analysis on the RNA list with zero assays. This is what the SETD8 and zebrafish studies report. Every other method must beat its agreement with protein-derived pathway calls.
The practice baseline spends the same budget of $k$ assays on the genes with the largest absolute fold-change, the way the ABC-transporter and RBM47 studies chose targets. It then reruns the enrichment with validated genes fixed. It is the strongest competitor on realism, because it is what labs actually do.
The strongest computational baseline uses no assays. It weights each gene by its CPTAC across-tumor mRNA–protein correlation and runs a weighted enrichment. A second version combines the RNA evidence with the $k$ validated proteins through ActivePathways' Brown fusion, with $k$ chosen by fold-change. We found no reported result on Lacuna for pathway-level agreement between RNA-based and protein-based enrichment, for any of these methods. The nearest figures concern a different quantity. MoPC cut candidate microRNA–target pairs from 30–60% to 17–18%. ActivePathways found 192 prognostic pathways in METABRIC, 33 of them only through combining data types. Neither number is directly comparable to ours.
Two ablations isolate what drives any gain:
- Flat-prior ablation: the information-based selection with a flat prior, which tests whether the concordance features matter.
- Fold-change-selection ablation: the concordance prior with fold-change selection, which tests whether the selection rule matters.
Random selection serves as a sanity floor. Full proteomics serves as the ceiling.
All methods share the same changed-gene list, gene-set database (Gene Ontology and KEGG, the Kyoto Encyclopedia of Genes and Genomes), false-discovery threshold, and budget $k$. Prior features are fit on cohorts disjoint from the evaluation cohort. Beating the baselines means a higher area under the agreement-versus-$k$ curve than both the fold-change practice baseline and concordance weighting, across tumor types, with non-overlapping bootstrap intervals.
Why not just…?
Why not just weight genes by their CPTAC mRNA–protein correlation and skip the assays? This is the real threat. Across-tumor correlation is a reasonable proxy for whether a change between conditions will reach the protein. The concordance-weighted baseline is designed to test this directly. If it matches active selection at $k\le 10$, the honest conclusion is that a lookup table suffices, and we will publish it as such. We expect it to fall short in two places. Some genes have low variance across tumors but large perturbation effects, so their correlation is uninformative. And a weight cannot resolve which redundant pathway calls hinge on a single shared gene.
Why not just run full proteomics on every experiment? Mass spectrometry that measures proteins broadly, such as data-independent acquisition, covers thousands of proteins. However, it misses low-abundance proteins and is noisy from run to run (Data-independent acquisition proteomics reveals circulating biomarkers of coronary chronic total occlusion in humans). It is also unavailable for archived RNA-seq datasets like GSE228621. We use it as the ground truth and the ceiling. The method is aimed at the common case where a lab can afford a handful of antibodies.
Why not just validate the top fold-change genes, as everyone does? The ABCA4 case shows that huge RNA changes can have no protein behind them. More basically, top hits are chosen for gene-level salience, not pathway-level information. The fold-change baseline measures this cost directly.
Why not predict protein from RNA with a learned translator like scChord? scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation maps single-cell RNA to surface protein. It does so only for immune-cell panels of 13–228 antibodies, and its accuracy on joint protein structure, measured by correlation-matrix distance, degraded as the panel grew. It does not cover the intracellular proteins that make up adhesion or MAPK pathways in tumor cells.
Proposed Experiments
Retrospective study. We use matched CPTAC/TCGA breast, glioblastoma, and lung data (77, 106, and 201 samples, as used by MoPC). We define metastasis-relevant contrasts, such as basal versus luminal breast tumors, and run differential expression on RNA. We then simulate budgeted validation by revealing proteomics for the selected genes only. Protein-derived pathway calls, from the same enrichment run on full proteomics, serve as truth. The metric is the agreement curve of pathway calls (F1) as $k$ goes from 0 to 30, together with calibration of the probabilities $q_P$. Priors are trained leave-one-tumor-type-out.
Prospective study. We repeat the zebrafish serial selection on MDA-MB-231 parental, F1, and F2 cells. We profile both RNA-seq and data-independent acquisition proteomics on the same samples, then compare each method's $k$-assay pathway calls to the proteome-derived calls. We also run Western blots for CTSD, SERPINE1, DDIT4 and MT1X, which the original study never did. The same protocol applies to SETD8-knockdown SNU475 cells, including phospho-ERK blots to test the MAPK claim.
Decisive checks. We need to see three things. Active selection must beat both fold-change selection and concordance weighting at matched $k$ in the retrospective cohorts. The flat-prior ablation must lose most of that gain. And in the zebrafish model, at least one published pathway-level conclusion must flip or survive in the direction the method predicted before the proteomics were unblinded.
References
- A high-stringency blueprint of the human proteome
- Identifying drivers of breast cancer metastasis in progressively invasive subpopulations of zebrafish-xenografted MDA-MB-231
- Impact of SETD8/KMT5A overexpression on hepatocellular carcinoma progression and prognosis
- Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer
- A novel Nav1.5-dependent feedback mechanism driving glycolytic acidification in breast cancer metastasis
- Expression Profiling of ABC Transporters in a Drug-Resistant Breast Cancer Cell Line Using AmpArray
- Integrative pathway enrichment analysis of multivariate omics data
- Integrated microRNA and proteome analysis of cancer datasets with MoPC
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- Suppression of 14-3-3γ-mediated surface expression of ANO1 inhibits cancer progression of glioblastoma cells
- Data-independent acquisition proteomics reveals circulating biomarkers of coronary chronic total occlusion in humans
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation