Multi-Omics and Machine Learning Identify Immune-Linked Gene Signatures for LUAD Stratification

Abstract

Background: Lung adenocarcinoma (LUAD) is the most common subtype of non-small-cell lung cancer and is one of the leading causes of cancer-related deaths globally. Despite current developments, reliable biomarkers for effective diagnosis, prognosis, and patient stratification are still lacking. Methods: We analyzed publicly available TCGA-LUAD and GEO datasets using integrative bioinformatics approaches, including differential gene expression, weighted gene co-expression network analysis (WGCNA), survival modeling, mutation profiling, immune cell infiltration scores, machine learning, and bulk-RNA and single-cell RNA sequencing. Results: A total of 5581 deregulated genes were identified, with the turquoise module (298 genes) showing strong correlation with LUAD (Corr = −0.79, p < 2.2 × 10−308). The integration of two analyses yielded 281 overlapping genes, out of which nine candidates (ANO2, CHIAP2, CPED1, DNASE1L3, GSTM5, HTR3C, PRKCE, SLC14A1, and WNT3A) were selected via LASSO Cox regression to build a prognostic risk model. High-risk patients have significantly worse survival (log-rank p = 0.0027). CPED1 exhibited the highest mutation frequency, with 41% of TCGA-LUAD samples harboring mutations. Among all CPED1 mutation events, missense mutations were the most common (47%). GSEA and KEGG analysis revealed significant enrichment of pathways such as nucleocytoplasmic transport, oxidative phosphorylation, protein processing in the endoplasmic reticulum, ribosome, and ribosome biogenesis in high-risk patients. Immune infiltration analysis indicated differences in immune cell infiltration scores between high- and low-immune-score groups, with M1 macrophages showing strong statistical correlation with aDC, monocytes, and CD4+ naïve T cells. Machine learning confirmed that the combined Enet+PLS model predicted CPED1 as a core predictor, and CPED1 was successfully validated in independent GEO datasets (GSE43458 and GSE31210), showing strong diagnostic accuracy (AUCs up to 0.98). Finally, single-cell RNA sequencing revealed that CPED1 was mostly expressed in fibroblasts and myeloid cells, with CPED1 significantly downregulated in LUAD compared with normal samples. Conclusions: This study integrates multi-omics and machine learning to highlight CPED1 as a promising candidate biomarker, with potential diagnostic and prognostic relevance in LUAD. The nine-gene risk signature stratified patients by survival outcomes in the TCGA cohort. Genomic and immune analyses revealed features associated with the high-immune-score group. As the study is entirely computational and the prognostic model lacks external survival validation, these findings should be regarded as preliminary and hypothesis-generating, requiring future independent validation and functional studies to confirm the biological significance and clinical utility of CPED1 and related genes.