Serum Protein Fishing for Machine Learning-Boosted Diagnostic Classification of Small Nodules of Lung

Diagnosis
of benign and malignant small nodules of the lung remains
an unmet clinical problem which is leading to serious false positive
diagnosis and overtreatment. Here, we developed a serum protein fishing-based
spectral library (ProteoFish) for data independent acquisition analysis
and a machine learning-boosted protein panel for diagnosis of early
Non-Small Cell Lung Cancer (NSCLC) and classification of benign and
malignant small nodules. We established an extensive NSCLC protein
bank consisting of 297 clinical subjects. After testing 5 feature
extraction algorithms and six machine learning models, the Lasso algorithm
for a 15-key protein panel selection and Random Forest was chosen
for diagnostic classification. Our random forest classifier achieved
91.38% accuracy in benign and malignant small nodule diagnosis, which
is superior to the existing clinical assays. By integrating with machine
learning, the 15-key protein panel may provide insights to multiplexed
protein biomarker fishing from serum for facile cancer screening and
tackling the current clinical challenge in prospective diagnostic
classification of small nodules of the lung.