Protein Expression Net: an integrated graph-guided computational framework and virtual perturbation approach for prioritizing ageing-associated target hypotheses from plasma proteomics

Ageing-associated plasma proteomic signatures can predict chronological age, but translating predictive biomarkers into mechanistically interpretable and experimentally testable target hypotheses remains challenging. We developed Protein Expression Net (PEN), a six-stage computational framework for linking plasma proteomic ageing signatures to candidate mechanistic target hypotheses. PEN integrates six sequential analytical stages: (1) plasma proteome input, (2) proteomic age prediction using a multilayer perceptron (MLP), (3) diagnostic protein identification and functional module characterization using Integrated Gradients, (4) graph-based target discovery using a protein–protein interaction network with graph convolutional propagation and diffusion, (5) discovery-oriented reranking integrating multi-component biological evidence, and (6) Geneformer-based virtual perturbation as an orthogonal in silico support stage. Applied to 44,179 UK Biobank participants profiled for approximately 2,920 plasma proteins, the MLP achieved accurate age prediction (test R² = 0.8723, Pearson r = 0.934, MAE = 2.30 years). Graph-based prioritization identified candidate mechanistic target hypotheses including FBN1, FBLN5, EDA, RLN3, LHCGR, COL14A1, HLA-E, and TSPAN4. Geneformer-based virtual perturbation in an independent ageing human skin single-cell transcriptomic dataset (GSE130973) provided orthogonal in silico support for selected candidate hypotheses. After recalibration using an expanded null distribution of 500 random token-valid genes, FBLN5 and HLA-E showed significant positive OLD-to-YOUNG transcriptional state shifts in specific cell populations, whereas other candidates displayed weaker or context-dependent effects. PEN provides a reproducible six-stage computational framework for translating plasma proteomic ageing signatures into interpretable candidate mechanistic target hypotheses. The final Geneformer perturbation stage provides orthogonal in silico functional support, strengthening biological plausibility without constituting experimental validation. The framework is generalizable to other biomarker-driven target discovery problems.