Machine-learning-guided single-cell transcriptomic interrogation identifies biomarkers and candidate therapeutics for immune-cold lung adenocarcinoma

Background The heterogeneity of the tumor immune microenvironment (TME) significantly influences the efficacy of immunotherapy in lung adenocarcinoma (LUAD). However, a systematic classification framework to distinguish immune-inflamed (hot) and immune-desert (cold) tumors in LUAD remains lacking. Methods We performed consensus clustering based on immune gene expression profiles to stratify LUAD samples into hot and cold tumor subtypes. Multi-omics integration including transcriptomic, genomic, and single-cell RNA sequencing (scRNA-seq) data was used to dissect the immune landscape, identify subtype-specific gene modules, and construct a Consensus Machine Learning–based Signature (CMLS) for prognostic evaluation. The predictive features were further interpreted using SHAP (Shapley Additive Explanations) analysis, and key regulators such as SHCBP1 were validated through functional assays. Results The hot tumor subtype was characterized by enhanced infiltration of CD8⁺ T cells, M1 macrophages, and NK cells, alongside increased immunogenicity and favorable survival outcomes. In contrast, cold tumors exhibited immunosuppressive features, with enrichment of M2 macrophages, Tregs, and TGF-β signaling, correlating with poor prognosis and resistance to immunotherapy. The CMLS risk model, derived from 113 machine learning algorithm combinations, showed robust performance in stratifying patients across multiple cohorts. SHAP analysis identified SHCBP1 as a top predictive gene; functional experiments confirmed its oncogenic role in promoting LUAD cell proliferation and migration. Furthermore, drug sensitivity analysis revealed distinct therapeutic vulnerabilities between subtypes. Single-cell and cell–cell communication analyses further highlighted the immune exhaustion and exclusion phenotypes of cold tumors. Conclusions This study proposes a robust immune-based classification and risk scoring framework for LUAD, uncovering mechanistic differences between hot and cold tumors. Our findings provide novel biomarkers and therapeutic targets, paving the way for precision immunotherapy in LUAD patients.