Disease-first public-data integration with local virtual knockout prioritizes shared proteins linking osteoarthritis and osteoporosis

Osteoarthritis and osteoporosis frequently coexist in older adults, but shared molecular programs remain unclear. We developed a disease-first multilayer public-data integration framework to identify shared protein candidates while preserving disease-specific structure before cross-disease comparison. Public bulk transcriptomic datasets from osteoarthritis synovium and cartilage and osteoporosis-related monocytes, femoral bone, and osteogenic stromal compartments were analyzed independently within each disease. Disease-level signatures were generated using differential expression analysis, robust rank aggregation, functional enrichment, and weighted gene co-expression network analysis. A shared axis was defined by concordant dysregulation, shared biological processes, and criteria-positive coexpression-module correspondences, with module matching subsequently calibrated against a size-preserving null model. Candidate genes were mapped to proteins and prioritized by integrating interaction topology, signaling priors, protein annotation, tissue support, disease-association and genetic evidence, local CARNIVAL virtual knockout, and single-cell contextual perturbation analysis. Both diseases showed reproducible signatures enriched in antigen presentation, cytokine regulation, extracellular matrix organization, osteoclast differentiation, ossification, and bone remodeling. Cross-disease comparison yielded 234 criteria-positive module pairs; 53 retained pair-specific support at empirical FDR < 0.05, whereas the total number of criteria-positive pairs did not exceed the global null expectation. Protein-level integration prioritized HLA-DRB1, HSP90AA1, CTSK, RPL7, PRG4, HLA-DRA, CLEC3B, TIMP1, SPP1, and APOE. Local virtual knockout assigned the highest model-derived CARNIVAL scores to HLA-DRB1 and HSP90AA1. Across the evaluated network configurations, HLA-DRB1 exceeded the prespecified score threshold in all four evaluable configurations, whereas HSP90AA1 exceeded it in four of six, indicating greater configuration stability for HLA-DRB1. Primary matched-null calibration of 39 evaluable candidate-compartment pairs retained nine pairs at global FDR < 0.05. External quality-control sensitivity analysis retained 125,090 of 161,470 cells and identified 10 globally FDR-supported pairs. Although only two of the nine primary pairs remained supported in the same compartment, five of the six primary supported genes retained evidence in at least one compartment. Recalculation using the quality-controlled cell-context evidence retained all primary top 10 proteins, with HLA-DRB1 and HSP90AA1 remaining the two highest-ranked candidates. Composition-aware bulk sensitivity analysis retained the direction of 36 of 40 candidate–cohort effects, while matched transcriptional-program adjustment retained 24 of 27 effects and showed greater attenuation of MHC-II candidates than of ribosomal or protein-folding candidates. These findings support a shared osteoimmune–matrix-remodeling protein axis linking osteoarthritis and osteoporosis and provide candidates for future experimental validation.