Integrative Serum Metabolic Fingerprints Based Multi-Modal Platforms for Lung Adenocarcinoma Early Detection and Pulmonary Nodule Classification

Abstract

Identification of novel non-invasive biomarkers is critical for the early diagnosis of lung adenocarcinoma (LUAD), especially for the accurate classification of pulmonary nodule. Here, a multiplexed assay is developed on an optimized nanoparticle-based laser desorption/ionization mass spectrometry platform for the sensitive and selective detection of serum metabolic fingerprints (SMFs). Integrative SMFs based multi-modal platforms are constructed for the early detection of LUAD and the classification of pulmonary nodule. The dual modal model, metabolic fingerprints with protein tumor marker neural network (MP-NN), integrating SMFs with protein tumor marker carcinoembryonic antigen (CEA) via deep learning, shows superior performance compared with the single modal model Met-NN (p < 0.001). Based on MP-NN, the tri modal model MPI-RF integrating SMFs, tumor marker CEA, and image features via random forest demonstrates significantly higher performance than the clinical models (Mayo Clinic and Veterans Affairs) and the image artificial intelligence in pulmonary nodule classification (p < 0.001). The developed platforms would be promising tools for LUAD screening and pulmonary nodule management, paving the conceptual and practical foundation for the clinical application of omics tools.

Keywords: deep learning; lung adenocarcinoma; metabolomics; multi-modal; pulmonary nodule.

Figure 1

Overall experimental schematic of nanoparticle based laser desorption/ionization mass spectrometry (NPLDI MS). 1 µL of native serum was incubated with ferric nanoparticles without pretreatment and directly analyzed by laser desorption/ionization mass spectrometry (LDI MS) to record Na+ and K+ adducted signals. The deep learning method was used to construct diagnostic models based on the serum metabolic fingerprints. Blind test was conducted with the same protocol to evaluate the performance of the diagnostic models.

Figure 2

Development and blind test of serum metabolic fingerprints (SMFs) based lung adenocarcinoma (LUAD) diagnostic model. A) Schematic overview of the deep learning approach used to develop and validate the SMFs based integrated LUAD diagnostic model. B) Score of the MP‐NN identified in the training cohort. p‐values were calculated using a Wilcoxon test. Error bars refer to interquartile. C) Area‐under‐curve (AUC) for individual parameters in the training cohort. p‐values were calculated using a DeLong test. Error bars refer to 95% confidence intervals (CIs). D) Receiver operating characteristic curve (ROC) of the individual parameters in the training cohort. E) Confusion tables of binary results of the MP‐NN model in the training cohort. F) ROC of the individual parameters in the test cohort. G) Confusion tables of binary results of the MP‐NN model in the test cohort. H) Detection rates of MP‐NN and carcinoembryonic antigen (CEA) at different specificity in the test cohort. I) MP‐NN score levels summarized by stage in the whole LUAD cohort. p‐values were calculated using a Chi‐square test. Error bars indicate interquartile. J) MP‐NN detection rates summarized by stage in the whole LUAD cohort. p‐values were calculated using a Chi‐square test. *p < 0.05, and *** p < 0.001.

Figure 3

Performance of MP‐NN in pulmonary nodule classification. A) Schematic depicting the approach for evaluation of the performance of MP‐NN in pulmonary nodule classification. B) MP‐NN score levels in patients with benign and malignant nodules. p‐values were calculated using a Wilcoxon test. Error bars indicate interquartile. C) Representative data of patients with pulmonary nodules. Left, data of two patients with low MP‐NN scores who were diagnosed with pulmonary infection (# 1486; # 2255). Right, data from two patients with high MP‐NN scores who were diagnosed with lung adenocarcinoma (LUAD) (# 550; # 515). D) Receiver operating characteristic curve (ROC) of MP‐NN, Image‐AI, Mayo Clinic, and Veterans Affairs model for pulmonary nodule classification in the whole cohort. E) ROC of MP‐NN in different nodule radiological subtypes. F) ROC of MP‐NN in different nodule sizes. *** p < 0.001.

Figure 4

Development and blind test of serum metabolic fingerprints (SMFs) based tri modal pulmonary nodule classification model. A) Schematic overview of the random forest approach used to develop and validate the SMFs based tri modal pulmonary nodule classification model. B) Receiver operating characteristic curve (ROC) of different machine learning algorithms using ten fold cross‐validation in the training set. C) Scatter plot for the graphical comparison of Image‐AI and MPI‐RF in the whole cohort. D) ROC of pulmonary nodule classification models in the training and test set.