Ma Qiong, Huang Chun-Xia, He Jia-Wei, Zeng Xiao, Qu Yu-Li, Xiang Hong-Xia, Zhong Yang, Lei Mao, Zheng Ru-Yi, Xiao Jun-Jie, Jiang Yu-Ling, Tan Shi-Yan, Xiao Ping, Zhuang Xiang, You Li-Ting, Fu Xi, Ren Yi-Feng, Zheng Chuan, You Feng-Ming
Hospital of Chengdu University of Traditional Chinese Medicine, Chengdu, Sichuan Province, China.
College of Artificial Intelligence, Xi'an Jiaotong University, Xian, Shanxi Province, China.
Int J Surg. 2025 Feb 1;111(2):2055-2071. doi: 10.1097/JS9.0000000000002152.
Determining the benign or malignant status of indeterminate pulmonary nodules (IPN) with intermediate malignancy risk is a significant clinical challenge. Oral microbiota-lung cancer (LC) interactions have qualified oral microbiota as a promising non-invasive predictive biomarker in IPN.
Prospectively collected saliva, throat swabs, and tongue coating samples from 1040 IPN patients and 70 healthy controls across three hospitals. Following up, the IPNs were diagnosed as benign (BPN) or malignant pulmonary nodules (MPN). Through 16S rRNA sequencing, bioinformatics analysis, fluorescence in situ hybridization (FISH), and seven machine learning algorithms (support vector machine, logistic regression, naïve Bayes, multi-layer perceptron, random forest, gradient-boosting decision tree, and LightGBM), we revealed the oral microbiota characteristics at different stages of HC-BPN-MPN, identified the sample types with the highest predictive potential, constructed and evaluated the optimal MPN prediction model for predictive efficacy, and determined microbial biomarkers. Additionally, based on the SHAP algorithm interpretation of the ML model's output, we have developed a visualized IPN risk prediction system on the web.
Saliva, tongue coating, and throat swab microbiotas exhibit site-specific characteristics, with saliva microbiota being the optimal sample type for disease prediction. The saliva-LightGBM model demonstrated the best predictive performance (AUC = 0.887, 95%CI: 0.865-0.918), and identified Actinomyces, Rothia, Streptococcus, Prevotella, Porphyromonas , and Veillonella as biomarkers for predicting MPN. FISH was used to confirm the presence of a microbiota within tumors, and external data from a LC cohort, along with three non-IPN disease cohorts, were employed to validate the specificity of the microbial biomarkers. Notably, coabundance analysis of the ecological network revealed that microbial biomarkers exhibit richer interspecies connections within the MPN, which may contribute to the pathogenesis of MPN.
This study presents a new predictive strategy for the clinic to determine MPNs from BPNs, which aids in the surgical decision-making for IPN.
确定具有中等恶性风险的肺结节（IPN）的良恶性状态是一项重大的临床挑战。口腔微生物群与肺癌（LC）的相互作用使口腔微生物群成为IPN中有前景的非侵入性预测生物标志物。
前瞻性收集来自三家医院的1040例IPN患者和70例健康对照的唾液、咽拭子和舌苔样本。随访后，将IPN诊断为良性肺结节（BPN）或恶性肺结节（MPN）。通过16S rRNA测序、生物信息学分析、荧光原位杂交（FISH）和七种机器学习算法（支持向量机、逻辑回归、朴素贝叶斯、多层感知器、随机森林、梯度提升决策树和LightGBM），我们揭示了HC-BPN-MPN不同阶段的口腔微生物群特征，确定了具有最高预测潜力的样本类型，构建并评估了用于预测效能的最佳MPN预测模型，并确定了微生物生物标志物。此外，基于对机器学习模型输出的SHAP算法解释，我们在网络上开发了一个可视化的IPN风险预测系统。
唾液、舌苔和咽拭子微生物群具有部位特异性特征，唾液微生物群是疾病预测的最佳样本类型。唾液-LightGBM模型表现出最佳的预测性能（AUC = 0.887，95%CI：0.865-0.918），并确定放线菌属、罗氏菌属、链球菌属、普雷沃菌属、卟啉单胞菌属和韦荣球菌属为预测MPN的生物标志物。FISH用于确认肿瘤内微生物群的存在，并使用来自LC队列的外部数据以及三个非IPN疾病队列来验证微生物生物标志物的特异性。值得注意的是，生态网络的共丰度分析表明，微生物生物标志物在MPN内表现出更丰富的种间连接，这可能有助于MPN的发病机制。
本研究为临床从BPN中确定MPN提供了一种新的预测策略，有助于IPN的手术决策。
