henhaobio
Sign in

bioRxivCancer BiologyPreprint

Transparent Weighting of Heterogeneous Evidence for Auditable Candidate Prioritization

Tue Nguyen MauiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Motivation: Differential-expression and network-based prioritization capture complementary molecular evidence. However, integrating these criteria poses a trade-off between score integration and methodological transparency. Existing approaches derive weights through optimization against predictive or benchmark metrics, or use manual or equal-weight settings without explicitly encoding the rationale for relative criterion importance. Results: We introduce AHP-CDS, an auditable candidate-ranking method that makes criterion importance explicit through weights derived from the Analytic Hierarchy Process from explicit pairwise judgments, which allows the resulting rankings to be evaluated through consistency, sensitivity and contribution analyses. The resulting pairwise judgments achieved acceptable consistency ($\text{Consistency Ratio} = 0.050$). The leave-one-criterion-out analysis showed that ranking sensitivity followed the assigned weight ordering: the removal of Fold-change producing most of the disruption (Spearman compare to full method $\rho = 0.796$), whereas removing betweeness causes the least ($\rho = 0.992$). Score decomposition further made individual rank changes traceable to their evidence dimension. Contact: tue1661@gmail.com or 25001662@hus.edu.vn Availability: all code is uploaded on \url{https://github.com/NguyenMauTue/PRIDE-breast-cancer-exosome-biomarker-discovery} Supplementary: Is all posted online

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 6 (15 Sept 2026) · CC BY 4.0

Analytic Hierarchy Process weights merge fold-change, FDR and protein-interaction centrality into a traceable composite ranking of 143 triple-negative breast cancer exosomal protein candidates.

Combining differential-expression and network-centrality evidence requires specifying criterion importance, which existing approaches either optimize against benchmarks or set equal or manually without a stated rationale. AHP-CDS (Analytic Hierarchy Process Composite Decision Score) encodes importance as consistency-checked pairwise judgments and adds contribution and leave-one-criterion-out analyses so each rank change can be audited.

Key results

  1. ShownExplicit Saaty-scale pairwise judgments produce an internally consistent weight vector. It ranks fold-change first, then network degree, FDR and betweenness, and each choice has a stated biological or statistical reason.

    Weights 0.518 (|log2 FC|), 0.242 (degree), 0.164 (FDR), 0.076 (betweenness); Consistency Ratio 0.050 (Table 1: 0.0503), below the 0.10 acceptability cutoff. FDR was down-weighted relative to degree because of n = 3 per group. · 4.2. Case-study parameterization; Tables; Abstract

  2. ShownWhole-pool agreement is high with the Equal-weight and FC-only rankings and lower with Centrality-only, yet AHP weighting still substantially reorders individual candidates. The divergence is strongest against the centrality-only ranking.

    Across 143 candidates, Spearman ρ/Kendall τ was 0.904/0.736 vs Equal-weight, 0.876/0.684 vs FC-only, 0.602/0.454 vs Centrality-only. Vs Equal-weight, 71.3% moved >5 ranks and 27.3% >20 (max 42); vs Centrality-only, 84.7% and 49.7% (max 86) with RBO 0.399 at depth 10. Top-10 overlap 8, 5, 2 of 10 with FC-only, Equal-weight, Centrality-only. · 5.1. Rank-level evaluation

  3. ShownLeave-one-criterion-out sensitivity follows the AHP weight order. Fold-change dominance is concentrated among the top-ranked candidates rather than spread across the pool.

    Dropping FC: Spearman ρ 0.796 vs full ranking, 80.4% of candidates shift >5 ranks, max Δr 87; dropping degree ρ 0.947 (64.3%), FDR ρ 0.981 (41.3%), betweenness ρ 0.992 (top-10/15 unchanged). Mean FC contribution share 0.531 in top-15 vs 0.416 in the rest; degree 0.216 vs 0.278. · 5.2. Criterion-level evaluation; Abstract

  4. SuggestedAHP-CDS top-k sets keep balanced representation across curated, network-linked and unannotated evidence tiers. They beat a random-draw null only at intermediate depths, unlike a random-walk diffusion (RWR) comparator and Equal-weight.

    Null band from 10,000 random draws from 143 candidates: AHP-CDS G* 0.865 at k = 5 (RWR 0.952, at the 95% bound), 0.991 at k = 10 (tying the bound), 1 at k = 15, 0.991 at k = 20. AHP-CDS gains 3, 4, 5, 6 of 18 Tier-1 genes; RWR exactly two at each depth; Equal-weight's top 5 lacks Tier-4. · 5.3. Portfolio-level evaluation; 4.4. Evaluation procedures

  5. ShownScore decomposition makes each rank move attributable to a criterion. Every single-criterion baseline gets corrected toward the criterion it leaves out, and cases such as TSG101 and VCL show both the gains and the trade-offs of integration.

    At T = 15, candidates promoted vs FC-only had mean degree share 0.415 vs 0.097, and those promoted vs Centrality-only had mean FC share 0.601 vs 0.253 for demoted ones. TSG101: AHP-CDS rank 17 vs 43 (FC-only) and 25 (Centrality-only), CDeg 0.160, CFC 0.233. VCL: rank 95 vs 22 (Centrality-only), rising to 44 without FC (51-position gain, largest in pool). · 5.2. Criterion-level evaluation; 5.4. Case-study

Methods, data and reproducibility

Exosomal label-free proteomics is processed through quality control, left-censored imputation, limma differential expression, ReactomePA GSEA and a STRING subnetwork. Four min-max-scaled criteria (|log2 FC|, -log10 FDR, log(1+x) degree and betweenness) are summed with weights from Saaty-scale pairwise comparisons via the approximate eigenvector method.

  • Label-free (MaxQuant LFQ) proteomics of exosomal fractions from MDA-MB-231 TNBC cells and non-tumorigenic MCF-10A cells · human (MDA-MB-231 and MCF-10A cell lines) · n = 3 biological replicates per condition, two technical each · PRIDE PXD056161
  • Candidate funnel: identified proteins filtered by quality control, core-enrichment membership in retained pathways, and contaminant/blacklist removal · human · n = 1,782 identified; 866 after QC; 195 core-enrichment; 143 final
  • STRING protein-protein interaction subnetwork restricted to candidates (combined score ≥ 0.7) · human (Homo sapiens STRING identifiers)
  • Comparative Toxicogenomics Database gene-disease associations used to assign candidates to evidence tiers · n = 18 Tier-1 genes in the 143-candidate pool
Methods & tools
MaxQuant (label-free quantification output), limma (lmFit, contrasts.fit, eBayes, topTable), Perseus-style left-censored Gaussian imputation, ReactomePA gsePathway (GSEA), Ensembl BioMart, Gene Ontology, STRING REST API, igraph, Comparative Toxicogenomics Database (CTD), Analytic Hierarchy Process (Saaty scale), Rank-Biased Overlap, Spearman and Kendall rank agreement, Heat-kernel random-walk diffusion ranking (Ren et al. 2019 Global Ranking analogue)
Software & environment
All analyses in R; sessionInfo and the full pipeline are in the GitHub repository; STRING, Reactome (ReactomePA), Ensembl BioMart and CTD were all queried in August 2026, with STRING and CTD reflecting their state at query time rather than a fixed release.
Reproducibility gaps
Input data are public via PRIDE accession PXD056161 (no URL printed); STRING and CTD are continuously updated with no fixed release, so results may differ on re-query; the six-gene blacklist is only in Supplementary material.
Limitations (6)
  • Weights come from one rater's fixed a priori pairwise judgments, not varied in the primary analysis nor aggregated across raters via group-AHP methods; effects of judgment magnitude (e.g., 3- vs 5- or 7-fold preferences) are untested, and CR verifies coherence, not correctness. · stated by the authors
  • Independence among the four criteria, assumed by the weighted-sum form, was not formally tested, and normalization choices (log1p for centralities, -log10 for FDR) alter each criterion's effective contribution independently of its assigned weight. · stated by the authors
  • Portfolio comparisons are exploratory: the 143-candidate pool is small and tier-imbalanced, giving a coarse null distribution at small k, no correction across depths was applied, and AHP-CDS exceeds the null band only at k = 15 and 20. · stated by the authors
  • Acknowledged general caveats include n = 3 biological replicates per condition, possible imputation artifacts, attrition of nodes in the STRING network, and no wet-lab validation of prioritized candidates. · stated by the authors
  • Only one public dataset contrasting a single TNBC cell line with a single non-tumorigenic line is analyzed; no independent cohort or patient-derived exosome data tests whether rankings, leave-one-out patterns or tier balance generalize. · noted in this summary
  • Usefulness rests on an author-defined evenness index over CTD tiers that rewards including unannotated Tier-4 candidates; without a benchmark of validated biomarkers, a higher G* does not demonstrate better prioritization accuracy. · noted in this summary

Read this if you need a transparent, auditable scheme for weighting expression and network evidence when prioritizing proteomic biomarker candidates from small-sample studies.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHeil '23Fine '19Kanduri '17Patil '16Chen '24Jain '22
43% semantic matchScientific Communication & Education
B. J. Heil, C. S. Greene·bioRxiv·Jan 5, 2023