Molecular signature comprising 11 platelet-genes enables accurate blood-based diagnosis of NSCLC

BACKGROUND: Early diagnosis is crucial for effective medical management of cancer patients. Tissue biopsy has been widely used for cancer diagnosis, but its invasive nature limits its application, especially when repeated biopsies are needed. Over the past few years, genomic explorations have led to the discovery of various blood-based biomarkers. Tumor Educated Platelets (TEPs) have, of late, generated considerable interest due to their ability to infer tumor existence and subtype accurately. So far, a majority of the studies involving TEPs have offered marker-panels consisting of several hundreds of genes. Profiling large numbers of genes incur a significant cost, impeding its diagnostic adoption. As such, it is important to construct minimalistic molecular signatures comprising a small number of genes. RESULTS: To address the aforesaid challenges, we analyzed publicly available TEP expression profiles and identified a panel of 11 platelet-genes that reliably discriminates between cancer and healthy samples. To validate its efficacy, we chose non-small cell lung cancer (NSCLC), the most prevalent type of lung malignancy. When applied to platelet-gene expression data from a published study, our machine learning model could accurately discriminate between non-metastatic NSCLC cases and healthy samples. We further experimentally validated the panel on an in-house cohort of metastatic NSCLC patients and healthy controls via real-time quantitative Polymerase Chain Reaction (RT-qPCR) (AUC = 0.97). Model performance was boosted significantly after artificial data-augmentation using the EigenSample method (AUC = 0.99). Lastly, we demonstrated the cancer-specificity of the proposed gene-panel by benchmarking it on platelet transcriptomes from patients with Myocardial Infarction (MI). CONCLUSION: We demonstrated an end-to-end bioinformatic plus experimental workflow for identifying a minimal set of TEP associated marker-genes that are predictive of the existence of cancers. We also discussed a strategy for boosting the predictive model performance by artificial augmentation of gene expression data.