Machine learning-based plasma proteomic analysis identifies a novel disease-defining biomarkers

Recent advances in serum proteomic analysis tools have expanded our understanding of various diseases. To identify a novel disease-specific proteomic markers, we used the high-throughput proximity extension assay (PEA) platform to profile plasma proteins from 62 patients with atopic dermatitis (AD), or ulcerative colitis (UC) and 29 healthy subjects. Differentially expressed protein (DEP) analysis yielded 85 (20 unique) and 99 (14 unique) upregulated proteins in AD and UC, respectively, compared to healthy subjects. We integrated a machine-learning (ML) model and selected 24 proteins to distinguish between the diseases, which accurately predicted disease-defining biomarkers based on distinctive proteomic signatures and disease severity. Correlation analysis with disease severity identified upregulated ML-selected proteins such as CCL13, CCL26, CD70, CDON, LY6D and MMP1 in AD, and ITM2A and REG4 in UC. Among these identified proteins, we suggest CDON, CD70 and LY6D, which had highly correlated expression levels, as AD-specific biomarkers. Thus, our results provide insight into the application of ML algorithms for disease diagnosis, and our model may be expanded to other disease contexts.

血清蛋白质组学分析工具的最新进展，拓展了我们对多种疾病的认知。为筛选新型疾病特异性蛋白质组标志物，本研究采用高通量邻近延伸测定（proximity extension assay, PEA）平台，对62例特应性皮炎（atopic dermatitis, AD）、溃疡性结肠炎（ulcerative colitis, UC）患者及29例健康受试者的血浆蛋白质组进行了谱分析。差异表达蛋白质（differentially expressed protein, DEP）分析结果显示，与健康受试者相比，AD组与UC组分别上调表达85种（含20种特有蛋白）与99种（含14种特有蛋白）蛋白质。本研究整合机器学习（machine learning, ML）模型，筛选出24种可区分两类疾病的蛋白质；该模型基于独特的蛋白质组特征与疾病严重程度，可精准预测疾病特异性生物标志物。与疾病严重程度的相关性分析显示，经ML筛选的上调蛋白质中，AD组包括CCL13、CCL26、CD70、CDON、LY6D及MMP1，UC组包括ITM2A与REG4。在上述筛选得到的蛋白质中，本研究提出表达水平高度相关的CDON、CD70及LY6D可作为AD特异性生物标志物。综上，本研究结果为机器学习算法在疾病诊断中的应用提供了新思路，且本模型可推广至其他疾病场景。