Additional file 1 of Cyclic multiplex fluorescent immunohistochemistry and machine learning reveal distinct states of astrocytes and microglia in normal aging and Alzheimer’s disease

Additional File 1: Table S1. Demographic and neuropathological characteristics of study subjects. Description: Abbreviations: ADNC = AD neuropathological changes; APOE = Apolipoprotein E genotype; CAA = cerebral amyloid angiopathy; CVD = cerebrovascular disease; F = female; LBD = Lewy body disease; M = male; NA = Not available/applicable; NOS = not otherwise specified; NP Dx = neuropathological diagnosis. Table S2. Primary and secondary antibodies used in this study and sequence of immunohistochemistry cycles. Description: Note: GFAP and DAPI detection are needed in all the cycles to guarantee an adequate alignment of the images. Abbreviations: AF488 = AlexaFluor 488; Cy = cyanine; Dk = donkey; Gt = goat; Ms = mouse; Rb = rabbit. All secondary antibodies were purchased from Jackson ImmunoResearch Labs, West Grove, PA. Table S3. Results of mixed effects regression models. Description: Results of mixed effects regression models with diagnosis (CTRL vs. AD) or state (homeostatic vs. intermediate vs. reactive) as a fixed effect, respectively, and subject ID as random effect in both cases, are reported. Table S4. Model performance statistics for CTRL vs. AD binary classifiers. Description: Model performance statistics for the binary classification task of CTRL vs. AD for both the gradient boosting machine (GBM) and the convolutional neural network (CNN) machine learning models are reported. For all heuristics except for AUC and AUCPR (which are not threshold-dependent), the threshold was chosen by maximizing the accuracy. 95% confidence intervals were estimated by bootstrapping the hold-out test set across 500 iterations. Table S5. Results of Bayesian hyperparameter optimization. Description: The final hyperparameters determined by the Optuna hyperparameter tuning framework are reported. The Optuna optimizer maximized the out-of-sample area under the receiver operating characteristic (ROC) curve (AUC), which in turn was determined by 3-fold cross-validation for each trial.

附加文件1：表S1。研究对象的人口统计学与神经病理学特征。
说明：缩写说明：AD神经病理改变（AD neuropathological changes, ADNC）、载脂蛋白E基因型（Apolipoprotein E genotype, APOE）、脑淀粉样血管病（cerebral amyloid angiopathy, CAA）、脑血管疾病（cerebrovascular disease, CVD）、女性（female, F）、路易体病（Lewy body disease, LBD）、男性（male, M）、无可用/不适用（Not available/applicable, NA）、未另行说明（not otherwise specified, NOS）、神经病理学诊断（neuropathological diagnosis, NP Dx）。

表S2。本研究所用一抗、二抗及免疫组织化学循环序列。
说明：注：所有循环中均需开展胶质纤维酸性蛋白（glial fibrillary acidic protein, GFAP）与4',6-二脒基-2-苯基吲哚（4',6-diamidino-2-phenylindole, DAPI）检测，以确保图像可实现充分对齐。
缩写说明：AF488 = AlexaFluor 488；Cy = 花青素（cyanine）；Dk = 驴源；Gt = 山羊源；Ms = 小鼠源；Rb = 兔源。所有二抗均购自宾夕法尼亚州西格罗夫市的杰克逊免疫研究实验室（Jackson ImmunoResearch Labs）。

表S3。混合效应回归模型结果。
说明：本文报告了两类混合效应回归模型的结果：其一以诊断状态（对照组[CTRL] vs. 阿尔茨海默病组[AD]）为固定效应，其二以细胞状态（稳态[homeostatic] vs. 中间态[intermediate] vs. 反应态[reactive]）为固定效应，两类模型均以受试者ID作为随机效应。

表S4。对照组与阿尔茨海默病二分类分类器的模型性能统计量。
说明：本文报告了梯度提升机（gradient boosting machine, GBM）与卷积神经网络（convolutional neural network, CNN）两种机器学习模型在对照组（CTRL）与阿尔茨海默病（AD）二分类任务中的性能统计量。除受试者工作特征曲线（receiver operating characteristic, ROC）下面积（area under the curve, AUC）与精确召回曲线下面积（area under the precision-recall curve, AUCPR）这两类与阈值无关的指标外，其余所有启发式指标均通过最大化准确率选取最优阈值。通过对留存测试集开展500次迭代自助抽样，估算得到95%置信区间。

表S5。贝叶斯超参数优化结果。
说明：本文报告了通过Optuna超参数调优框架确定的最终超参数。Optuna优化器以留样本外受试者工作特征曲线（receiver operating characteristic, ROC）下面积（area under the curve, AUC）最大化为优化目标，该指标通过每次试验的3折交叉验证计算得到。