Artificial intelligence-assisted proteomic signatures for discriminating malignant from benign pulmonary nodules

Lung cancer remains a major global health burden. Although low-dose CT (LDCT) is effective for early detection, its clinical application is limited by a high false-positive rate and the need for subsequent invasive confirmation, such as biopsy. Meanwhile, the extensive protein-level alterations observed in lung cancer provide a rationale for developing minimally invasive plasma proteomics-based screening a panel of biomarkers. In this study, untargeted proteomic profiling was performed on a tissue cohort (n = 61, paired tumor and normal adjacent tissue samples spanning pre-invasive (AAH/AIS), MIA and IAC), a discovery plasma cohort (n = 221), and an independent validation plasma cohort (n = 144). Integrative tissue-plasma analysis was used to identify concordant proteins. Classifier training was conducted on the plasma proteomics data through a LightGBM-based pipeline with a two-stage feature selection strategy to derive plasma protein panels and construct binary classifiers that distinguish malignant from benign nodules. The diagnostic performance of the trained models was then evaluated in an independent validation cohort. Furthermore, binary classifiers were constructed to discriminate the invasiveness of lesions among LUADs. A subset of proteins was identified whose expression in both tissue and plasma was significantly associated with tumor invasiveness. The compact pipeline identified reproducible plasma protein panels and trained LightGBM classifiers that robustly discriminated malignant from benign pulmonary nodules. The classifiers maintained strong performance in the independent validation cohort (AUC-ROC: 0.949-0.986). We developed separate classification models to stratify lung adenocarcinoma subtypes by distinct patterns of invasiveness (AUC-ROC: 0.762-0.803). High-depth proteomics combined with LightGBM enables the identification of robust plasma protein panels for discriminating pulmonary nodules and stratifying invasiveness. This generalizable strategy, along with reproducible biomarker panels, supports non-invasive early diagnosis of lung adenocarcinoma.