Architectural Fragmentation and Immune Reprogramming Define Early Malignant Transition in Lung Adenocarcinoma

The earliest steps of lung adenocarcinoma (LUAD) evolution remain poorly resolved because existing spatial-omics technologies cannot scale to the large archival cohorts required to capture rare transitional states. Here, we reconstructed a stage-resolved spatial atlas of LUAD progression using ROSIE, a deep-learning framework inferring 50-plex protein expression from routine H&E slides (114 specimens; 3.8 million cells). Using a pathology-defined precursor continuum (Normal[->]AAH[->]AIS[->]MIA[->]IAC) as a pseudotemporal axis, this scalable virtual multiplexing approach reveals a previously unrecognized two-phase immune reprogramming cascade tightly coupled to a measurable structural transition we term architectural fragmentation. During the AAH[->]AIS transition, transient adaptive immune engagement collapses, marked by a 48.4% loss of CD8 + T cells and the emergence of microfragmented epithelial-stromal interfaces. Subsequent AIS[->]MIA[->]IAC transitions are dominated by myeloid-driven immunosuppression, expansion of CD163 + macrophage niches, and the establishment of an immune-excluded architecture. Fragmentation severity in AIS strongly predicts downstream invasive potential, positioning it as an early, quantifiable hallmark of malignant transition. These findings define a mechanistic sequence linking structural disintegration, immune collapse, and invasive outgrowth and demonstrate that H&E-based virtual multiplexing enables cohort-scale dissection of early cancer evolution and identification of actionable windows for immunoprevention. The ROSIE predictions were validated against imaging mass cytometry (IMC) across 25 precursor samples, achieving concordance up to r [≥] 0.90 (93% of marker-cell-type pairs at r [≥] 0.80). A complementary comparison against MIPHEI-ViT, an independent virtual multiplexing framework, corroborated the structural and pan-immune components of this signature across all stages, while highlighting fine-grained immune-lineage resolution as a shared open challenge for H&E-based virtual multiplexing.