Pan-cancer Variance Decomposition Nominates Translationally Actionable Therapeutic Antigen Candidates Across 33 Cancer Types

Computational prioritization of therapeutic antigen candidates has largely relied on mean-level differential expression, a framework insensitive to inter-patient transcriptomic heterogeneity. Here we implement a genome-wide variance decomposition engine across the full TCGA transcriptome (60,656 genes, 33 cancer types), constructing a pan-cancer antigen variance matrix from which candidates are nominated using a pre-defined four-layer filter: pan-cancer variance recurrence, normal tissue safety, functional dependency (DepMap), and immune microenvironment classification. After exclusion of sex-chromosome genes, immunoglobulin loci, and nuclear transcription factors, 17 candidates satisfied both DepMap dependency and immune-cold classification. Three representative candidates were selected based on pre-defined druggability and pipeline criteria. CRIPTO/TDGF1, nominated across 7 cancer types, exhibited profound functional dependency across four cancer lineages (Ovarian mean Chronos = − 0.728, 98% of cell lines below threshold) and immune-cold phenotype, with no current clinical pipeline. Multi-layer computational validation confirmed a favorable therapeutic window: GTEx normal tissue expression was near-absent across 54 adult tissue sites (normal ovary median 0.080 TPM), TCGA-OV variance rank was top 13.3% (consistent with heterogeneous rather than absent expression; 95% of tumors expressing), an independent GEO cohort (GSE26193, n = 107) replicated variance rank at top 4.3%, Human Protein Atlas IHC confirmed medium-level protein expression in ovarian cancer with GPI-anchored surface accessibility, and direct comparison with mean-based DEG ranking demonstrated 2.7-fold greater sensitivity for TDGF1 discovery by variance decomposition (DEG rank top 35.9% vs. TANK rank top 13.3%). LGALS7B/Galectin-7, nominated across 8 cancer types, showed consistent functional dependency and immune-cold classification mechanistically consistent with PD-1 glycosylation-mediated T-cell suppression, representing a complete mechanism-gap opportunity. SLC34A2/NaPi2b, the highest pan-cancer recurrence candidate (20 cancer types), was nominated prior to literature consultation; subsequent search identified TUB-040 (NCT06303505), a next-generation NaPi2b-directed ADC currently in Phase 1/2 development, representing prospective convergence analogous to the CLDN6/BNT211 case in a companion study. GO enrichment analysis of high-recurrence candidates revealed significant enrichment of epithelial morphogenesis and BMP signaling pathways, consistent with reactivation of developmental gene programs. These results suggest that variance decomposition enables systematic, prospective nomination of translationally actionable antigen candidates, and reveals a natural bifurcation between immune-cold therapeutic targets and immune-hot microenvironmental regulators within the high-variance gene pool.