The Human Interactome Vulnerability Atlas: Exact Genome-Scale Node-Loss Mapping Reveals Degree-Confounded Cancer Dependency Signals

Background: Network-medicine studies often use centrality to prioritize biologically important genes, yetprotein-interaction networks are strongly shaped by degree heterogeneity and ascertainment bias. Counterfactualeffective-resistance centralities offer a global measure of structural change after node loss, but theirgenome-scale biological value has not been established under nonlinear degree control.Methods: We constructed the Human Interactome Vulnerability Atlas from BioGRID 4.4.230, comprising19,835 proteins and 824,511 unique undirected interactions. For 19,150 non-articulation nodes, we computedexact normalized node-deletion dissimilarity from pre- and post-deletion Kirchhoff traces using a grounded-Laplacian low-rank update solver; 684 articulation vertices were recorded as structural sentinels and RPA2(Entrez 6118) served as the excluded ground. We then evaluated 16,397 continuously scored genes againstDepMap 26Q1 Chronos profiles across up to 1,208 models. The prespecified model compared nonlinear controlsfor degree, expression and coverage with and without KVD percentile. A separate 17,068-gene analysis testedwhether articulation status predicted common essentiality.Results: The full Atlas completed in 3,221.6 s with peak resident memory of 2.90 GiB. On independentlyrerun synthetic grids, exact updates agreed with explicit deletion-and-pseudoinverse oracles to maximumabsolute KVD error 1.55 × 10−14 and achieved 2.1–35.6-fold amortized speedups for N = 100–1,000. AtlasKVD was strongly coupled to node degree (Spearman ρ = 0.9048), although ranking overlap remained partial.Adding KVD percentile increased in-sample R2 by 0.003263 and repeated-cross-validation R2 by 0.003188(95% CI 0.002766–0.003610). However, the adjusted coefficient was negative (β = −0.1920, HC3 95% CI−0.2335 to −0.1506), contradicting the prespecified positive vulnerability hypothesis. Articulation verticesshowed strong crude enrichment for common essentiality (22.35% vs 10.67%), but this disappeared afternonlinear adjustment (OR 0.8618, 95% CI 0.6791–1.0937; p = 0.2213) and reversed under strict degreematching (22.37% vs 27.19%).Conclusions: Exact genome-scale node-loss mapping is computationally feasible and yields a reusablehuman interactome Atlas. In this static, unweighted interactome, however, global KVD does not validate thehypothesized positive cancer-dependency signal, and crude articulation enrichment is explained by covariatestructure. The results support counterfactual spectral mapping as a measurement resource while requiringnonlinear degree-aware null models before biological interpretation.