利用機器學習方式以臨床特徵和免疫細胞為基礎預測HBe抗原陰性B肝帶原者的HBs抗原清除與後續臨床表現

Abstract

研究背景 慢性B肝的感染是造成肝硬化與肝癌的危險因子，之前的研究發現慢性B肝感染且HBs抗原未清除者罹患肝癌的風險比清除者高，而且未清除HBs抗原的感染者體內高劑量的HBs抗原可能會造成免疫耗損以及更嚴重的肝臟病變，但是目前抗病毒藥物僅能清除感染者體內的HBV DNA，鮮少能清除HBs抗原，因此尋找HBs抗原清除的機轉，以及觀察抗原清除後對於免疫系統的影響和後續肝功能是否會回復，變得至關重要。

目的 我們希望能利用機器學習的方式找出和HBs抗原清除有關的預測因子，進而找出HBs抗原清除的機轉，此外我們也希望觀察HBs抗原清除後免疫系統是否有變化，以及從免疫的變化程度建立免疫分數，觀察免疫的改變是否與後續肝功能有關。

方法 本次研究分成兩個階段，第一階段我們利用先前發表有包含宿主因子與病毒因子的縱貫性研究資料，以建立一個包含920名HBe陰性感染者的回溯性世代研究，並用機械學習的方式找出HBs抗原清除的預測因子。第二階段，我們利用222名來自前述世代研究下所建立的PBMC子世代建立前瞻性研究，以流式細胞儀分析12個免疫標記。我們用逐級線性判別分析區分清除與未清除者的免疫差異，並用羅吉斯迴歸係數建立免疫分數。 結果 第一階段我們發現病毒量、病毒基因型、ALT與空腹胰島素為穩定的HBs抗原清除預測因子，並用不同演算法預測10年~20年的HBs清除機率，所得AUC介於0.71~0.83；第二階段我們利用線性判別分析發現清除與非清除者有9個標記在HBs抗原清除後會發生改變，包含NK bright cells、NK dim cells、CD4 T cells, 細胞激素製造(IFN-γ+ NK cells、IFN-γ+ CD8 T cells)、免疫耗損標記(PD1+ CD8 T cells、PD1+ CD4 T cells)、以及早期活化標記(CD69+ CD4 T cells、CD69+ CD8 T cells). 並觀察到除了NK bright cells外(p=0.01)，其他細胞皆不受抗原預測因子的影響。最後我們建立抗原清除相關的細胞建立免疫分數，觀察到ALT下降會受到抗原清除相關的細胞影響(p=0.04)。 結論 我們的資料指出HBs抗原清除後，免疫圖譜也會隨之改變，而且此改變也會影響後續肝臟發炎的程度。

Background Chronic Hepatitis B (CHB) is strong risk factor for both liver cirrhosis and hepatocellular carcinoma. Patients with CHB rarely achieve a functional cure after successful anti-HBV therapy. It remains uncertain whether there is a further HCC reduction in patients who achieve functional cure, indicated by hepatitis B surface antigen (HBsAg) seroclearance, after viral suppression. More-over, immune exhaustion is related to high titter of HBsAg, which may lead to severe clinical outcome. Little is known about how immune functions are restored after functional cure in CHB patients. Specific Aims First, we wanted to find out predictors for HBs seroclearance with machine learning approached, then we wanted to understand whether loss of HBsAg may result in change in immune cellular profiles after taking account the identified predictors of HBsAg seroclearance. Finally, we wanted to construct an immune cell score as a prognostic factor and to investigate its’ impact on subsequent liver function. Methods This is a two-stage study. In the first stage, we searched HBsAg seroclearance predictors using a retrospective cohort study by machine learning algorithms with data from a published longitudinal viral load study, which included various host factors and viral factors. Nine hundred and twenty HBeAg negative CHB individuals were included. In the second stage, we did a prospective analysis 222 of the 920 CHB, who were recruited from a PBMC cohort to examine their immune cell phenotypes with 12 markers by multidimensional flow cytometry. Stepwise linear discriminant analysis was used for feature extraction, and logistic regression coefficients for risk scoring. Result The most predictive variables for HBsAg seroclearance from different machine learning approaches were HBV DNA, HBV genotype, ALT and fasting plasma levels of insulin, and the area under the receiver operating characteristic curve varied between 0.71 and 0.83 for 10~20 years prediction of HBsAg seroclearance, The phase 2 study revealed that subjects with HBsAg seroclearance and those without displayed differences in nine lymphocyte subsets, based on linear discriminate analysis. Among these, there are lymphocytes pertaining to innate and adaptive immunity, including NK bright cells, NK dim cells, CD4 T cells, cytokine production function (IFN-γ+ NK cells, IFN-γ+ CD8 T cells), immune exhaustion (PD1+ CD8 T cells, PD1+ CD4 T cells), and early activation markers (CD69+ CD4 T cells, CD69+ CD8 T cells). No association were found between cellular markers and determinants of HBsAg except for NK bright cells (p=0.01). We also established a prognosis immune risk score that was found to be inversely associated subsequent ALT levels (p=0.04). Conclusion The data suggest seroclearance of HBs antigen maybe associated with distinct changes in peripheral immune cell subsets, resulting in improvement of liver dysfunction.