A spatially aware deep learning framework for multiscale cellular ecology profiling to predict 5-year recurrence in invasive lung adenocarcinoma

Recurrence remains a major clinical challenge in surgically resected invasive lung adenocarcinoma (ILA), where existing staging and grading systems fail to capture the cellular complexity that underlies tumor aggressiveness. We present PathRosetta, an AI framework that models cellular ecology of tumor microenvironment (TME) from hematoxylin-and-eosin slides. Rather than treating slides as set of independent patches, PathRosetta represents individual cells within their local spatial neighborhoods and integrates this cellular context with tissue-level architecture to predict five-year recurrence. Across 430 patients (876 slides), PathRosetta reached an AUC of 0.78 ± 0.04 on the internal cohort, outperforming IASLC grading (0.71), AJCC staging (0.64), and competing AI models (0.62–0.72), and generalized to external TCGA (0.75), CPTAC (0.76), and WFBCCC (0.75) cohorts. It achieved a hazard ratio of 9.54 and a concordance index of 0.70 and remained independently prognostic after adjusting for other variables (adjusted HR:4.05, p < 0.001 ). Most notably, within clinically homogeneous AJCC Stage I patients, PathRosetta isolated a high-risk subgroup with markedly poorer outcomes (5-year RFS 46.7% vs. 84.7%; p < 0.001 ). Exploratory analysis further suggested morpho-spatial subgroups within individual cell types, motivating future molecular validation. By characterizing the spatial cellular ecology of the TME, PathRosetta delivers interpretable, generalizable recurrence prediction that complements existing clinical systems.