8027 Background: Precise management of pulmonary ground-glass nodules (GGNs) is hindered by an incomplete understanding of the biological determinants driving the transition from indolence to aggressive progression. Methods: In a multicenter longitudinal study of 845 patients, we stratified GGNs into three trajectories via serial CT: Stable, ground glass opacity (GGO)-predominant progression, and solid-predominant progression. Integrative profiling was performed using WES, WGS, bulk RNA-seq, Visium HD, CosMx SMI, and imaging mass cytometry. We developed a deep learning model to predict solid-predominant progression using baseline radiographic features. Results: During a median follow-up of 55.8 months, our analysis indicated GGNs progression was driven predominantly by chromosomal instability rather than distinct point mutational profiles. Specifically, progressive GGNs harbored catastrophic structural variations, including chromothripsis, kataegis, and extrachromosomal DNA which were absent in stable lesions. Spatially, the immune microenvironments diverged sharply. While stable GGNs were quiescent, GGO-predominant growth lesions were significantly enriched with antigen-presenting macrophages (p = 0.00054), which facilitated the spatial aggregation of activated lymphocytes and promoted the formation of early tertiary lymphoid structure (TLS)-like niches through CCL19-CCR7 signaling. Conversely, in solid-predominant aggressive progression nodules, macrophages underwent a phenotypic shift from antigen-presenting toward SPP1+ state, accompanied by a marked reduction in CD74 and HLA-DR expression (p = 0.0019). Within the tumor core, SPP1+ macrophages suppressed T-cell effector function via SPP1-CD44 signaling. At the tumor margin, hypoxia-adapted endothelial cells exhibited significant upregulation of CD274 (PD-L1), contributing to T-cell inhibition, while activated fibroblasts deposited dense collagen matrices formed a physical barrier restricting B-cell infiltration. Together, these mechanisms established a coordinated immunosuppressive niche facilitating tumor growth. Leveraging these biological insights, we developed a deep learning model extracting radiomic features from the baseline CT nodule and an extended 2.5mm perinodular margin. This approach significantly outperformed models using the nodule alone or the nodule plus a 1mm margin (p < 0.05), achieving AUCs of 0.89 in the training cohort (n = 471), 0.85 in the internal test cohort (n = 202) and 0.83 in the external cohort (n = 117). Conclusions: Distinct radiographic progression trajectories of GGNs are underpinned by divergent chromosomal and spatial immune programs at the tumor core and margin, providing a biological foundation for margin-aware AI strategies to predict aggressive progression and guide optimal surgical intervention.
Jin et al. (Thu,) studied this question.