Computed Tomography-Based Quantitative Texture Analysis and Gut Microbial Community Signatures Predict Survival in Non-Small Cell Lung Cancer

SIMPLE SUMMARY: There is a lack of understanding of the pathogenesis and mechanisms accounting for the large variability in tumor response to immune checkpoint inhibition. In this study, we investigate the role and composition of the human gut microbiome in the clinical setting by integrating shotgun metagenomics and quantitative texture analysis (QTA) of CT images in NSCLC patients treated with anti-PD-L1 immunotherapy using a novel machine learning approach. Using all available parameters, the XGB machine learning system predicted therapeutic response with an accuracy of 83% and correctly separated long-term survival patients from short-term survival patients with an accuracy of 69%. Our findings show that an integrated signature of these characteristics may predict outcomes more accurately than separate measures and may have potential therapeutic implications in the future. ABSTRACT: This study aims to combine computed tomography (CT)-based texture analysis (QTA) and a microbiome-based biomarker signature to predict the overall survival (OS) of immune checkpoint inhibitor (ICI)-treated non-small cell lung cancer (NSCLC) patients by analyzing their CT scans (n = 129) and fecal microbiome (n = 58). One hundred and five continuous CT parameters were obtained, where principal component analysis (PCA) identified seven major components that explained 80% of the data variation. Shotgun metagenomics (MG) and ITS analysis were performed to reveal the abundance of bacterial and fungal species. The relative abundance of Bacteroides dorei and Parabacteroides distasonis was associated with long OS (>6 mo), whereas the bacteria Clostridium perfringens and Enterococcus faecium and the fungal taxa Cortinarius davemallochii, Helotiales, Chaetosphaeriales, and Tremellomycetes were associated with short OS (≤6 mo). Hymenoscyphus immutabilis and Clavulinopsis fusiformis were more abundant in patients with high (≥50%) PD-L1-expressing tumors, whereas Thelephoraceae and Lachnospiraceae bacterium were enriched in patients with ICI-related toxicities. An artificial intelligence (AI) approach based on extreme gradient boosting evaluated the associations between the outcomes and various clinicopathological parameters. AI identified MG signatures for patients with a favorable ICI response and high PD-L1 expression, with 84% and 79% accuracy, respectively. The combination of QTA parameters and MG had a positive predictive value of 90% for both therapeutic response and OS. According to our hypothesis, the QTA parameters and gut microbiome signatures can predict OS, the response to therapy, the PD-L1 expression, and toxicity in NSCLC patients treated with ICI, and a machine learning approach can combine these variables to create a reliable predictive model, as we suggest in this research.