Dual-Branch Semi-Supervised Deep Learning for Improved Lung Cancer Diagnosis With Metabolomics

Metabolomics, a comprehensive omics approach for profiling small-molecule metabolites within biological systems, has become an indispensable tool for disease diagnosis, prognostic assessment, and prediction of therapeutic response. In lung cancer research, metabolomics shows particular promise in early detection, subtype classification, treatment monitoring, and outcome prediction. Lung cancer remains one of the leading causes of cancer-related mortality worldwide, characterized by high treatment costs and persistently low five-year survival rates, underscoring the urgent need for improved diagnostic strategies. The Lung Cancer Metabolome Database (LCMD) is a pioneering, freely accessible online resource that aggregates lung cancer–associated metabolites identified from mass spectrometry–based studies, providing a valuable foundation for computational modeling. However, metabolomics data are often challenged by limited labeled samples and high levels of missingness. To address these challenges, we propose a novel Dual-Branch Semi-Supervised Deep Learning (DBSL) framework for lung cancer metabolomics analysis. The model incorporates a Multi-Features Consistency Attention (MFCA) mechanism to jointly capture global temporal patterns and local feature interactions. In addition, a semi-supervised optimization strategy integrating pseudo-labeling, stochastic data augmentation, and consistency regularization is designed to effectively exploit unlabeled data. Under extreme low-label regimes (5%, 10%, and 15% labeled data), DBSL achieves state-of-the-art performance on the LCMD dataset. These results demonstrate the strong potential of DBSL as a scalable and cost-effective approach for lung cancer diagnosis, contributing to the urgent clinical demand for early detection and intervention.