Integrating computational pathology and multi-transcriptomics to characterize lung adenocarcinoma heterogeneity and prognostic modeling

Background: Lung adenocarcinoma (LUAD) is the most prevalent subtype of non-small cell lung cancer (NSCLC), characterized by high molecular and pathological heterogeneity. While traditional histopathology plays a key role in LUAD diagnosis, integrating computational pathology with multi-omics analysis provides novel insights into tumor microenvironment (TME) dynamics and molecular mechanisms. However, the relationship between pathological histological features and genomic instability in LUAD remains poorly understood. Methods: This study employed whole-slide images (WSIs) from the TCGA-LUAD dataset, which were processed into image patches for deep learning feature extraction using ResNet-50 and pathological feature selection using CellProfiler. Copy number variations (CNV) were inferred using inferCNV, and high-dimensional weighted gene co-expression network analysis (hdWGCNA) was performed to identify key regulatory modules associated with CNV-defined malignant cell populations. Additional analyses included intercellular communication using CellChat, pseudotime trajectory inference with Monocle2, and immune landscape profiling. Finally, we performed correlation analyses between the gene expression patterns of high-CNV (HCNV) cell lines and pathological image features, followed by prognostic model construction using a machine learning benchmark framework. Results: This study first identified LUAD malignant cells with high CNV scores. These cells also exhibited high stemness, and their proportion gradually increases with the progression of LUAD. CNV-driven tumor subpopulations exhibited distinct metabolic and immune signatures, with HCNV cells showing enhanced glycolysis, MYC signaling, and immune evasion. Intercellular communication analysis highlighted VEGF, MK and IGF signaling pathways as key mediators of HCNV-stroma interactions. A set of 192 imaging features significantly correlated with CNV burden in LUAD was identified, including 11 pathological features from CellProfiler and 181 deep learning features from ResNet-50. Machine learning-based prognostic modeling using deep learning and pathology features demonstrated robust survival prediction, with high-risk patients exhibiting lower immune infiltration and reduced immunotherapy responsiveness. Conclusion: This study provides a comprehensive multi-dimensional framework integrating computational pathology and single-cell multi-omics to characterize LUAD heterogeneity. By identifying CNV-associated imaging features and key molecular regulators, we propose potential biomarkers for prognosis and therapeutic targeting in LUAD. However, as this study is based primarily on retrospective bioinformatics analysis, the clinical utility of these findings requires further validation through prospective cohorts and experimental studies. These results lay the groundwork for future translational applications but should be interpreted with caution in the absence of functional validation.