Centre for Rheumatology, University College London, London, United Kingdom.
Centre for Cardiometabolic and Vascular Medicine, University College London, London, United Kingdom.
Front Immunol. 2020 Jul 17;11:1527. doi: 10.3389/fimmu.2020.01527. eCollection 2020.
Neutralizing anti-drug antibodies (ADA) can greatly reduce the efficacy of biopharmaceuticals used to treat patients with multiple sclerosis (MS). However, the biological factors pre-disposing an individual to develop ADA are poorly characterized. Thus, there is an unmet clinical need for biomarkers to predict the development of immunogenicity, and subsequent treatment failure. Up to 35% of MS patients treated with beta interferons (IFNβ) develop ADA. Here we use machine learning to predict immunogenicity against IFNβ utilizing serum metabolomics data. Serum samples were collected from 89 MS patients as part of the ABIRISK consortium-a multi-center prospective study of ADA development. Metabolites and ADA were quantified prior to and after IFNβ treatment. Thirty patients became ADA positive during the first year of treatment (ADA+). We tested the efficacy of six binary classification models using 10-fold cross validation; k-nearest neighbors, decision tree, random forest, support vector machine and lasso (Least Absolute Shrinkage and Selection Operator) logistic regression with and without interactions. We were able to predict future immunogenicity from baseline metabolomics data. Lasso logistic regression with/without interactions and support vector machines were the most successful at identifying ADA+ or ADA- cases, respectively. Furthermore, patients who become ADA+ had a distinct metabolic response to IFNβ in the first 3 months, with 29 differentially regulated metabolites. Machine learning algorithms could also predict ADA status based on metabolite concentrations at 3 months. Lasso logistic regressions had the greatest proportion of correct classifications [F1 score (accuracy measure) = 0.808, specificity = 0.913]. Finally, we hypothesized that serum lipids could contribute to ADA development by altering immune-cell lipid rafts. This was supported by experimental evidence demonstrating that, prior to IFNβ exposure, lipid raft-associated lipids were differentially expressed between MS patients who became ADA+ or remained ADA-. Serum metabolites are a promising biomarker for prediction of ADA development in MS patients treated with IFNβ, and could provide novel insight into mechanisms of immunogenicity.
中和性抗药物抗体（ADA）会大大降低用于治疗多发性硬化症（MS）患者的生物制药的疗效。然而，导致个体产生 ADA 的生物学因素还没有得到很好的描述。因此，人们迫切需要生物标志物来预测免疫原性的发展，以及随后的治疗失败。多达 35%的接受β干扰素（IFNβ）治疗的 MS 患者会产生 ADA。在这里，我们使用机器学习来预测针对 IFNβ 的免疫原性，利用血清代谢组学数据。 作为 ABIRISK 联盟的一部分，从 89 名 MS 患者中采集了血清样本-一项针对 ADA 发展的多中心前瞻性研究。在 IFNβ 治疗前后对代谢物和 ADA 进行了定量。在第一年的治疗过程中，有 30 名患者 ADA 呈阳性（ADA+）。我们使用 10 倍交叉验证测试了六种二分类模型的有效性；k-最近邻、决策树、随机森林、支持向量机和带有/不带有交互作用的最小绝对收缩和选择算子（Lasso）逻辑回归。 我们能够从基线代谢组学数据中预测未来的免疫原性。带有/不带有交互作用的 Lasso 逻辑回归和支持向量机分别是识别 ADA+或 ADA-病例最成功的方法。此外，ADA+患者在最初的 3 个月内对 IFNβ 有明显的代谢反应，有 29 种差异调节的代谢物。基于 3 个月时的代谢物浓度，机器学习算法也可以预测 ADA 状态。Lasso 逻辑回归的正确分类比例最高[F1 评分（准确性度量）= 0.808，特异性= 0.913]。最后，我们假设血清脂质可以通过改变免疫细胞脂质筏来促进 ADA 的发展。这一假设得到了实验证据的支持，实验表明，在接触 IFNβ 之前，ADA+或 ADA-的 MS 患者之间，脂质筏相关脂质的表达存在差异。 血清代谢物是预测接受 IFNβ 治疗的 MS 患者 ADA 发展的有前途的生物标志物，并且可以为免疫原性的机制提供新的见解。
