Saharan Seema Singh, Nagar Pankaj, Creasy Kate Townsend, Stock Eveline O, Feng James, Malloy Mary J, Kane John P
Department of Statistics, University of Rajasthan, Jaipur, India.
UCSF Kane Lab, San Francisco, USA.
Proc (Int Conf Comput Sci Comput Intell). 2021 Dec;2021:312-317. doi: 10.1109/csci54926.2021.00125. Epub 2022 Jun 22.
Presently, the role of cytokines in severe illness like COPD, cancer, cardiac disease associated with smoking is being explored to enable preemptive diagnosis and delivery of treatment interventions. We are investigating the connection between the elevation of inflammatory plasma cytokine in smokers versus nonsmokers. Disease indicator cytokines can be used to monitor the progression of disease which can help in the crucial task of prognosis and definitive diagnosis. Powerful and versatile Machine Learning algorithms can be leveraged to extract insights that cannot be obtained manually. We have applied Support Vector Machine (SVM) on 65 plasma cytokines and other traditional biomarkers to differentiate smokers and nonsmokers. To optimize the classification separability, we have used the following techniques: Principal component analysis (PCA), 10-fold cross validation and variable importance. The primary metric of evaluation is Area Under Receiver Operating Curve (AUROC), though we have additionally recorded and compared prediction accuracy across classifiers. The results are very promising. The AUROC classification accuracy achieved by SVM using the selected predictor feature variables is 89.2% with a 95%CI (85.4%,93.1%). The most prominent cytokines, contributing to the classification, in the order of importance are: I-TAC, Age, TG, G-CSF-CSF-3, MDC-CCL22, Eotaxin-3, LIF, IL-2, Eotaxin-2, MIP-3alpha. The AUROC classification accuracy improved to 93% with a 95% CI (90.1%,99.5%) upon choosing the five most prominent cytokines. The versatile prowess of Machine Learning algorithms such as Support Vector Machine can translate pioneering molecular discoveries into actionable insights that can be applied in the field of translational and precision medicine to save life.
目前，正在探索细胞因子在慢性阻塞性肺疾病（COPD）、癌症、与吸烟相关的心脏病等严重疾病中的作用，以实现早期诊断和提供治疗干预措施。我们正在研究吸烟者与非吸烟者中炎症性血浆细胞因子升高之间的联系。疾病指标细胞因子可用于监测疾病进展，这有助于进行预后和明确诊断这一关键任务。可以利用强大且通用的机器学习算法来提取无法通过手动获得的见解。我们已将支持向量机（SVM）应用于65种血浆细胞因子和其他传统生物标志物，以区分吸烟者和非吸烟者。为了优化分类可分离性，我们使用了以下技术：主成分分析（PCA）、10折交叉验证和变量重要性。评估的主要指标是受试者工作特征曲线下面积（AUROC），不过我们还额外记录并比较了各分类器的预测准确性。结果非常有前景。使用选定的预测特征变量，支持向量机实现的AUROC分类准确率为89.2%，95%置信区间为（85.4%，93.1%）。对分类贡献最大的最突出细胞因子，按重要性排序为：γ-干扰素诱导的单核因子（I-TAC）、年龄、甘油三酯（TG）、粒细胞集落刺激因子-集落刺激因子3（G-CSF-CSF-3）、巨噬细胞来源的趋化因子-CCL22（MDC-CCL22）、嗜酸性粒细胞趋化因子3、白血病抑制因子（LIF）、白细胞介素2（IL-2）、嗜酸性粒细胞趋化因子-2、巨噬细胞炎性蛋白-3α（MIP-3alpha）。选择五种最突出的细胞因子后，AUROC分类准确率提高到93%，95%置信区间为（90.1%，99.5%）。支持向量机等机器学习算法的通用能力可以将开创性的分子发现转化为可操作的见解，应用于转化医学和精准医学领域以挽救生命。
