評估多重磁珠抗體免疫檢測方法應用在登革熱患者的血清分型

Abstract

背景 登革熱是由登革病毒（DENV）感染所引起的疾病，登革病毒分為四種不同的血清型（DENV-1, DENV-2, DENV-3, DENV-4），不同登革病毒血清型可流行於同一區域內。因此，可能發生不同血清型的重覆感染，透過血清型分析幫助我們了解當次或是過去感染的病毒血清型別。其中，利用血清抗體進行血清型分型的方法可以增進我們對登革病毒感染史與疾病嚴重程度或免疫保護力之相關性的理解。目前，辨認野生型類病毒顆粒（Wild-type Virus-Like Particle, WT-VLPs）、突變型類病毒顆粒（Cross-reactive reduced fusion-peptide mutated VLPs, CRR-VLPs）和非結構蛋白1（Nonstructural protein 1, NS1）的免疫球蛋白M（IgM）和免疫球蛋白G（IgG）已被運用於區分登革病毒感染和其餘黃病毒感染，或是登革病毒的血清分型。然而，目前未知何種抗原（WT-VLPs, CRR-VLPs, NS1）的IgM或IgG抗體較能有效的區分現在及過去感染的登革病毒血清型。 研究目標 本研究的目標是使用WT-VLPs、CRR-VLPs和NS1抗原的多重磁珠抗體免疫檢測方法（MIA），偵測急性期和恢復期間IgM和IgG的抗體反應，評估其不同抗體反應區分感染血清型的能力。此外，同時評估各種建模方法推測感染登革病毒血清型的能力。 材料與方法 我們收集234例分別感染四種血清型之登革熱患者在急性期和恢復期的血清樣本，並使用MIA偵測血清中針對四種血清型的WT-VLPs、CRR-VLPs和NS1之IgM和IgG抗體反應。MIA以中位數螢光強度（Median Fluorescence Intensity, MFI）值呈現抗體反應的強度，並進一步轉換成常態化MFI（Normalized MFI, nMFI）等其他MFI轉換型式。利用nMFI的抗體反應數值建立多元羅吉斯回歸（Multinomial Logistic Regression, MLR）以推斷感染血清型。我們將MLR模型的血清型分型能力與支持向量機器（Support Vector Machine, SVM）和不平衡資料分類統合集成分析（Meta Imbalanced Classification Ensemble, MICE）進行比較，並使用8比2測試驗證方法評估模型表現。 結果 當考量利用急性期和恢復期對NS1之IgM和IgG抗體反應進行血清型分型時，MLR-RUS的平均血清型分型精確度為0.591至0.673間，其表現低於SVM-RUS模型（0.641至0.754）和MICE（0.740至0.765）的精確度。這個差異可以歸因於SVM-RUS和MICE相對於MLR-RUS，在推斷DENV-1感染擁有更高的精確度，但推斷DENV-2, -3, -4感染的精確度較低。相較之下，利用恢復期對WT-VLPs的IgM抗體反應的MLR-RUS模型最有效區分登革病毒感染血清型，對推測四種血清型感染的精確度較平衡，分別為對DENV-1的精確度為0.690、對DENV-2的精確度為0.417、對DENV-3的精確度為0.700、對DENV-4的精確度為0.579。此外利用IgG抗體反應的MLR-RUS模型則以辨認NS1抗體較好，在急性期和恢復期的平均精確度分別為0.663和0.673。 總結 MLR-RUS模型適合利用nMFI資料型態的抗體反應進行血清分型，且與其他機器學習演算法相比，MLR-RUS在分別推斷四種血清型感染時提供相對平衡的精確度。因此我們建議針對急性期的登革熱患者，能採用NS1之MIA偵測血清型專一性之IgG抗體反應；針對恢復期的登革熱患者血清，則採用WT-VLPs之MIA區分血清型專一性之IgM抗體反應。此外，若在IgM抗體消退之恢復期後期，NS1之MIA可有效偵測並區別血清型專一性之IgG抗體反應。

Background Dengue fever is caused by infection with any four serotypes of the dengue virus (DENV), which can co-circulate in areas where the disease is endemic. Repeat infection by a different serotype can occur. The use of serological serotyping methods can improve our understanding of the association between DENV infection history and disease severity or protective immunity. IgM and IgG antibody responses to wild-type (WT) virus-like particles (VLPs), cross-reactive reduced fusion-peptide mutated VLP (CRR-VLPs), and non-structural protein 1 (NS1) have been used to differentiate DENV infection from other flavivirus infections or differentiate one serotype of DENV from another serotype. However, it remains uncertain which antibody (IgM and IgG) responses against which antigens (WT-VLPs, CRR-VLPs, or NS1) are more effective in determining the status of current and past serotype infection. Objective Our objective was to assess the performance of multiplex microsphere immunoassay (MIA) using WT-VLPs, CRR-VLPs, and NS1 to distinguish serotypes in serotype-specific IgM or IgG antibody responses during the acute and convalescent phases. Additionally, we aimed to evaluate various modeling methods for inferring the infected serotypes. Materials and Methods The MIA was used to simultaneously measure IgM and IgG antibody reactivity against WT-VLPs, CRR-VLPs, and NS1 of all four serotypes in the acute and convalescent serum panels. These panels were collected from a total of 234 dengue patients. The antibody reactivities measured by MIA were obtained as a median fluorescence intensity (MFI), which was further transformed into a normalized MFI (nMFI) dataset and others. We first used the multinomial logistic regression model (MLR) utilized the nMFI dataset of antibody reactivity to infer the infected serotype. The performance of the MLR model was compared to that of support vector machine (SVM) and meta imbalanced classification ensemble (MICE). The model performance was validated by the 8:2 validation approach. Results The MLR-RUS models demonstrated an average accuracy range of 0.591 to 0.673, which was lower compared to the SVM-RUS models (0.641 to 0.754) and the MICE (0.740 to 0.765) when utilizing IgM and IgG antibody reactivity to NS1 for serotype inference. This difference in performance can be attributed to the SVM-RUS and the MICE achieving more accurate predictions for DENV-1 infection but displaying lower accuracy for DENV-2, DENV-3, and DENV-4 infections compared to the MLR-RUS models. The most effective MLR-RUS model for DENV serotyping involved utilizing convalescent IgM antibody reactivity to WT-VLPs, yielding a balanced accuracy of 0.690 for DENV-1, 0.417 for DENV-2, 0.700 for DENV-3, and 0.579 for DENV-4. Additionally, the MLR-RUS models utilizing IgG antibody reactivity to NS1 demonstrated higher accuracies of 0.663 and 0.673 during the acute and convalescent phases, respectively. Conclusions The nMFI dataset of antibody reactivity is well-suited for MLR-RUS models as it allows for a better balance when inferring serotype-specific DENV infection compared to other machine learning models. We propose utilizing NS1-based MIA and WT-VLPs-based MIA to distinguish serotype-specific IgG and IgM antibody responses in sera collected from dengue patients during the acute and convalescent phases, respectively. NS1-based MIA can effectively detect serotype-specific IgG antibodies after the decline of IgM antibodies.