Diagnostic tests that measure several blood or urine proteins usually beat single-protein tests when the proteins report on different biological processes, for example tumor growth plus clotting, or low oxygen plus inflammation. When these panels are mined from label-free quantitative proteomics, the usual feature-selection tools tend to return several proteins from the same process. Label-free quantitative proteomics (LFQ) means mass spectrometry that measures thousands of protein abundances per sample, one sample per instrument run, without chemical tags. We propose a panel-selection method that enforces "at most about one protein per biological module," with the modules taken from curated pathway databases. It also down-weights proteins that measure poorly across labs, using an existing multi-lab reference dataset.
The specific bet is about why the pathway knowledge should help in LFQ. Protein-protein correlations in label-free data are partly technical. Injection order and lab-to-lab drift make unrelated proteins rise and fall together, and low-abundance proteins drop out together. Diversity constraints built from the data's own correlations therefore inherit the batch structure. Pathway modules are defined outside the data and do not. Our prediction is that pathway-based diversity matters little within one batch but matters clearly when a panel is moved to a new lab or instrument. If that prediction fails, the pathway ingredient is not earning its place.
Core Claim
At a fixed panel size, a logistic classifier will transfer better to unseen labs and instruments if it is penalized with two things. The first is an exclusive penalty over pathway-derived protein modules, which makes proteins inside a module compete so that only one or a few survive. The second is a per-protein weight reflecting cross-lab measurement stability. It should transfer better than the same classifier with modules built from data correlations, and better than Lasso, network-smoothing penalties, or Random Forest subset search.
Concretely, for patient labels $y_i$, protein log-intensities $x_i$ and coefficients $\beta$, we minimize
$$ \sum_i \ell(y_i, x_i^\top\beta) \;+\; \lambda_1 \sum_{g} \Big(\sum_{j \in g} s_j |\beta_j|\Big)^2 \;+\; \lambda_2 \sum_j s_j|\beta_j| . $$
Here $\ell$ is the logistic loss, $g$ ranges over pathway modules, and $s_j \ge 1$ grows with protein $j$'s cross-lab coefficient of variation as estimated from reference material. The squared inner sum is the exclusive lasso of Exclusive Lasso for Multi-task Feature Selection. That paper introduced the $l_{1,2}$ penalty, which forces competition within groups, and tested it on web-page text categories. Because a protein can sit in several pathways, we handle overlapping groups following Correlated Feature Selection with Extended Exclusive Group Lasso, which extended the penalty to correlated, structured features but evaluated it on synthetic and immune-receptor data. Neither paper used biological pathways or proteomics. We supply both the groups and the setting.
Relation to Prior Work
Three groups of work motivate the gap. The first is LFQ panel studies. Proteomic landscape of primary and metastatic brain tumors for heterogeneity discovery used PLS-DA (partial least squares discriminant analysis) to select five proteins, KRT8, KRT19, KRT7, TACSTD2 and CDH1. All five are epithelial markers standing in for one fact, non-brain origin. Quantitative proteomics identified circulating biomarkers in lung adenocarcinoma diagnosis built a six-protein logistic panel that includes several glycolysis enzymes and used a STRING interaction network only to describe the result. In both papers the pathway enrichment comes after selection and plays no role in it. The same holds for Delineation of renal protein profiles in aristolochic acid I-induced nephrotoxicity in mice by label-free quantitative proteomics, which ran GO and KEGG enrichment only after calling differential proteins. The second is the measurement-noise evidence. Quartet protein reference materials and datasets for multi-platform assessment of label-free proteomics showed that randomized injection order can drop the signal-to-noise ratio to 0.59, against more than 23 for continuous injection. This is the evidence that data-derived protein correlations are technically contaminated, and it is the reason a knowledge-based grouping might help.
The second group, pathway-aware predictors, supplies methods and also constrains the design. Pathway-based subnetworks enable cross-disease biomarker discovery (SIMMS) cuts pathways into subnetworks, scores each patient's subnetwork dysregulation, and selects subnetworks with Lasso-Cox regression. On mRNA it reported an AUC of 0.75 against 0.67 for the clinical IHC4 test. We reuse its subnetwork decomposition to define modules. We differ in two ways: we select individual proteins, which is what an assay needs, and we work on LFQ data with its missing values and drift. Molecular pathway identification using biological network-regularized logistic models (Logit-Lapnet) adds a penalty that pushes connected genes toward similar coefficients. That rewards dense within-pathway clusters, which is the opposite of what a small diagnostic panel needs, so we include it as a baseline that tests the opposite hypothesis. Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression (pareg) models how much pathways overlap. We use its similarity idea only to merge near-duplicate GO, KEGG and Reactome terms before grouping. pareg itself produces cohort-level enrichment scores, not patient predictions.
The third group is the purely data-driven alternative. Quantitative proteomics and phosphoproteomics of urinary extracellular vesicles define putative diagnostic biosignatures for Parkinson's disease selected panels by Random Forest backward elimination followed by exhaustive search. It reported a 6-marker disease panel with AUC 94.3% and a 3-marker risk panel with AUC 99.8%. This is the strongest reported result on an LFQ-scale cohort in the file. These figures come from a single cohort of 141 samples, so they say nothing yet about transfer across labs.
Research Program
We first build the module library. We merge GO Biological Process, KEGG and Reactome. We collapse terms whose gene overlap is high, using pareg-style similarity, and split large pathways into SIMMS-style subnetworks. We then restrict to proteins detectable in the target sample type, so that groups reflect the proteome actually measured rather than the full genome. That correction matters because, as the brain-tumor study notes, mRNA and protein correlate as poorly as 0.54. Next we estimate per-protein stability weights $s_j$ from the Quartet data, which cover 15 labs, 9 platforms and 288 single-shot runs. These weights are computed once and are never fit to any clinical outcome.
The solver is proximal gradient descent with a known proximal operator for the exclusive penalty. We tune $\lambda_1$ and $\lambda_2$ along a path to hit target panel sizes of $k = 3, 5, 8$, because a panel's cost is set by how many proteins it measures. Missing values are imputed inside each training fold. We also run a missingness-aware variant that treats "not detected" as an additional binary feature, so that dropout-driven signal is not mistaken for biology.
Baselines
The naive reference point is the best single protein, chosen by training-fold AUC. A second reference is each study's own reported panel, refit under our protocol: the PLS-DA brain-tumor panel, the LUAD logistic panel, the ccRCC four-protein panel from Excavation of diagnostic biomarkers and construction of prognostic model for clear cell renal cell carcinoma based on urine proteomics (reported AUC 0.835), and the Parkinson's Random Forest panels. The reported numbers, AUC 0.973 for the brain panel, 92% detection for LUAD and AUC 94.3% for the Parkinson's disease panel, were obtained on small single-site cohorts. We quote them for context but compare only against our own reruns.
We treat two methods as the strongest competitors, because each represents a credible "no pathway knowledge needed" position. One is Lasso and elastic-net logistic regression. The other is Random Forest with backward elimination plus exhaustive search, reimplemented from the Parkinson's study. Logit-Lapnet represents the opposite, within-pathway-smoothing view. SIMMS-style module scores followed by picking the top-loading protein per selected module represent the two-stage view.
Every method sees identical folds, the same imputation, the same panel-size targets and the same hyperparameter search budget, set at 50 configurations per method by nested cross-validation. Every method also receives the same Quartet-filtered protein universe, so that the stability information is not an unfair advantage.
Three ablations isolate the ingredients. The first keeps the exclusive penalty but builds groups by hierarchical clustering of the training data's protein correlations, matched in group count and size. The second uses pathway groups with protein memberships randomly permuted, preserving group sizes. The third sets all $s_j = 1$. We count the method as beating a baseline only if three conditions hold together: a leave-lab-out or leave-cohort-out AUC gain whose paired-bootstrap 95% interval excludes zero, at equal $k$; no loss within the same lab; and coverage of more distinct modules per panel.
Why not just…?
Why not just run Lasso or elastic net on the intensity matrix? Under correlated features, Lasso tends to keep an arbitrary member of each correlated block. In Logit-Lapnet's breast-cancer comparison, both Lasso and elastic net missed PGR. In LFQ, part of that correlation is batch, so the arbitrary pick may be the protein that happens to track drift. That argument is plausible but untested on proteomics. The leave-lab-out experiment measures it directly, and if Lasso matches us under shift we will report that.
Why not enforce diversity from the data itself, with correlation clusters or a minimum-redundancy filter, and skip curated pathways? This is the most serious threat, because it captures most of the "orthogonal panel" idea without any database. Our only argument against it is that data correlations in LFQ are partly technical, which follows from the Quartet signal-to-noise collapse. The correlation-cluster ablation decides the question. We expect it to tie within a batch and lose across labs. If it ties across labs as well, the proposal reduces to "exclusive lasso on correlation clusters," and we would say so.
Why not run enrichment and then pick the top protein from each enriched pathway? This two-stage heuristic is what many groups do informally. It ignores whether the chosen proteins add information to one another, and enrichment computed on small cohorts is noisy. The SIMMS-plus-top-protein baseline is a stronger version of this heuristic, and we run it head-to-head.
Proposed Experiments
The decisive benchmark is built from Quartet. The four family cell lines (father F7, mother M8, twin daughters D5 and D6) differ genetically. We set up classification tasks, for example father versus mother and each parent versus the daughters, and train panels on a subset of labs and platforms. We then test on held-out labs and held-out instrument types, and separately on runs with randomized versus continuous injection order. The stability weights $s_j$ are estimated only from training labs in this experiment, to avoid leakage. This task is not clinical. It does supply the one thing the clinical cohorts lack, genuine biological differences measured repeatedly across many labs, so it can test the central prediction cleanly: the pathway-exclusive panel should gain over correlation-exclusive and Lasso panels under lab shift, and gain little within a lab.
Clinical relevance is checked on the Parkinson's urinary vesicle cohort (141 samples), the brain-tumor cohort (29), the LUAD plasma cohort (40) and the ccRCC urine cohort (23 discovery samples), wherever raw or processed matrices are publicly deposited. Where acquisition-batch or injection-order metadata exist, we use leave-batch-out splits. Otherwise we use repeated nested cross-validation and report confidence intervals honestly wide for cohorts this small. The outcomes are AUC and sensitivity at 90% specificity at fixed $k$. Alongside them we report the number of distinct modules covered, bootstrap selection stability (Jaccard overlap of selected proteins), and calibration.
Two results would falsify the idea. One is the correlation-cluster ablation matching the pathway version under lab shift. The other is the permuted-membership control performing as well as true pathways, which would mean that grouping alone, not biology, does the work. A positive result needs both a gain under shift and no loss within a lab. That combination would give a concrete, mechanistic reason to put enrichment knowledge into proteomic panel selection rather than only into its interpretation.
References
- Exclusive Lasso for Multi-task Feature Selection
- Correlated Feature Selection with Extended Exclusive Group Lasso
- Proteomic landscape of primary and metastatic brain tumors for heterogeneity discovery
- Quantitative proteomics identified circulating biomarkers in lung adenocarcinoma diagnosis
- Delineation of renal protein profiles in aristolochic acid I-induced nephrotoxicity in mice by label-free quantitative proteomics
- Quartet protein reference materials and datasets for multi-platform assessment of label-free proteomics
- Pathway-based subnetworks enable cross-disease biomarker discovery
- Molecular pathway identification using biological network-regularized logistic models
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- Quantitative proteomics and phosphoproteomics of urinary extracellular vesicles define putative diagnostic biosignatures for Parkinson’s disease
- Excavation of diagnostic biomarkers and construction of prognostic model for clear cell renal cell carcinoma based on urine proteomics